INFORMATION TO USERS

This manuscript has been reproduced from the microfilm master. UMi films the text directly from the original or copy submitted. Thus, some thesis and dissertation copies are in typewriter face, while others may be from any type of computer printer.

The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleedthrough, substandard margins, and improper alignment can adversely affect reproduction.

In the unlikely event that the author did not send UMI a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion.

Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning the original, beginning at the upper left-hand comer and continuing from left to right in equal sections with small overlaps. Each original is also photographed in one exposure and is included in reduced form at the back of the book.

Photographs included in the original manuscript have been reproduced xerographically in this copy. Higher quality 6” x 9” black and white photographic prints are available for any photographs or illustrations appearing in this copy for an additional charge. Contact UMI directly to order. UMI Bell & Howell Information and Learning 300 North Zeeb Road, Ann Arbor, Ml 48106-1346 USA 800-521-0600

MULTIMODALITY ELECTROPHYSIOLOGICAL MONITORING IN THE NEUROINTENSIVE GARE UNIT

DISSERTATION

Presented in Partial Fulfillment of the Requirements for

the Degree Doctor of Philosophy in the

Graduate School of The Ohio State University

By

André van der Kouwe, M.Eng.

*****

The Ohio State University

1999

Dissertation Committee: Approved by

Professor DeLiang Wang, Adviser

Professor Manjula Waldron Adviser Professor Urbashi Mitra Biomedical Engineering Richard C. Burgess, M.D., Ph.D. Center UMI Number: 9941449

UMI Microform 9941449 Copyright 1999, by UMI Company. All rights reserved.

This microform edition is protected against unauthorized copying under Title 17, United States Code.

UMI 300 North Zeeb Road Ann Arbor, MI 48103 ABSTRACT

In the neurointensive care unit (NICU) there is a need for a single bedside moni­ tor for continuously monitoring the function of the patient’s central nervous system.

This dissertation describes the development of a system of hardware and software for continuously and automatically monitoring the ongoing neurological condition of patients in the NICU. The system samples several electrophysiological waveforms at regular intervals along with routinely monitored physiological parameters. The electrophysiological data consists of brainstem auditor}", somatosensor}' and visual evoked potentials and epochs of the electroencephalogram (EEC) . The system ap­ plies peak detection and spectral analysis to extract salient parameters from the raw waveforms. These parameters are assembled into a state vector representing the patient’s neurophysiological state. The parameters include the latencies and ampli­ tudes of the significant peaks in the brainstem auditor}", somatosensory and visual evoked potentials, the spectral pole positions, peak frequencies and energy distribu­ tions of the EEC and such physiological parameters as heart rate, blood pressure and intracranial pressure. A display summarizing the history of the patient’s state is provided for visual review. The results are also made available on the network for review and further analysis on the local network. A web-based interface makes review possible anywhere within the hospital’s secure intranet during and after monitoring.

Experimental data from six intensive care patients show that certain regions of the

11 State space correspond with particular pathologies and this may have diagnostic value in particular patients. Experimental data from angioplasty and stent patients sug­ gest that the electrophysiological changes correspond with the induced physiological changes in the central nervous system as they occur in real time. If these results are extrapolated to NICU patients, it implies that further analysis of the state space djmamics may be of prognostic value in these patients. In particular, the results suggest that continuous multimodality electrophysiological monitoring of this type may potentially contribute to the quality of care of stroke patients with hemorrhagic and edematous brain injuries during the critical period during which tissue shifts and possible herniation may give rise to pressure on the brainstem and other structures, with associated morbidity or mortality.

I ll Dedicated to my parents and Evelina

IV ACKNOWLEDGMENTS

I am ver\' grateful to my research adviser, Dr. Richard C. Burgess of the Cleve­ land Clinic Foundation, for his expert knowledge and guidance in this research, for his patience and enthusiasm and his authoritative suggestions. He has been a con­ tinuous source of knowledge and inspiration. His efficient and valuable review of any material at short notice are greatly appreciated. I would also like to thank him for his optimistic and confident attitude and his concern not only for my progress in this research project, but also for the development of my personal career. 1 appreciate his provision of the hardware necessary to build up this monitoring system,together with the supporting infrastructure and environment for testing it.

Thanks are due to Professor Wang for his careful review of the final manuscript, and his assistance and authority in parallel research and coursework. My thanks also to Professors Mitra and Waldron for their helpful and wise suggestions related to the manuscript.

A special debt of thanks is due to Linda Gagnon, who is an evoked potential and intraoperative monitoring technologist with vast experience and enthusiasm. I appre­ ciate her tireless enthusiasm and assistance even at short notice, meticulous attention to the experimental setup, cheerfulness and encouragement. Her compassionate in­ teraction with her patients and colleagues has been a great source of inspiration. Thanks also to the other technologists on this qualified team for their assistance in the experimental sections of the research.

I wish to thank Dr. Jeffrey I. Frank of the Neurointensive Care Unit for his willingness to host the testing of the monitoring system in his unit, his assistance in identifying suitable patients, and his suggestions as to the most appropriate subjects to include in the protocol.

I would like to thank the intensive care fellows. Dr. Waleed El-Feky and Dr.

Harold McGrade, for their help in identifying suitable patients for the NICU study.

Thanks to Dr. Derek Krieger of the Neurolog}' Department for his helpful sugges­ tions, logistical assistance, explanations and encouragement.

Thanks to Doctors Thomas Masaryk and Perl for their willingness to assist in the evaluation of the monitoring system in the interventional neuroradiology suite, and for their suggestions and advice.

I would like to express my appreciation to the members of the Section of Neu­ rological Computing who contributed to this project by providing expert technical advice. In particular, Martin Andrews always had the answers to my UNIX-related questions, trivial or complex, tolerating interruptions at any time with advice and solutions. Thanks also to Misha Rekhson for his advice related to hardware drivers and communications on the UNIX platform. Thanks to John Turnbull for mathe­ matical contributions. Thanks also to Karl Horning for his support in issues related to purchasing the components for the system and assembling the hardware, and for his great sense of humor. It was a privilege to work with this accomplished team of scientists and engineers.

VI Finally, my thanks to the Graduate School of the Ohio State University for their investment in this research in the form of a Graduate Student Alumni Research

Award.

vu VITA

30 December, 1970 ...... Bom - Jonquière, Canada

1992 ...... B. Eng. Electronics, University of Pre­ toria 1995 ...... M. Eng. Electronics (Bioeng.), Univer- sitv of Pretoria

PUBLICATIONS

Research Publications

A..1.W. van der Kouwe and R.C. Burgess, "Peak detection in auditor}- and so­ matosensory evoked potentials by means of the zero-crossings wavelet transfor- m", Proc. 34th Rocky Mnt. Bioeng. Sv-mp., Dayton, OH, (1996).

2. .A..J.W. van der Kouwe and D.L. Wang, “Temporal alignment, spatial spread and the linear independence criterion for blind separation of voices", Proc. 19th Int. Conf. of IEEE EMBS, Chicago, IL, (1997).

3. -\..I.W. van der Kouwe and R.C. Burgess, “Examining the dynamics of a vector representing neurophysiological state during intensive care". InterJ. Complex Systems, doc. no. 257, (1998).

A.J.W. van der Kouwe and R.C. Burgess, “Long-term evoked potential and EEC monitoring system for the neurointensive care unit", J. Clin. Neurophys., 16 (2), (1999).

V lll 5. A.J.W. van der Kouwe and R.C. Burgess, "A continuous electrophysiological monitoring system with web-based review for the neurological intensive care unit”, Proc. 3rd Int. Workshop on Biosig. Interp., Chicago, IL, (1999).

6. A.J.W. van der Kouwe, D.L. Wang and G.J. Brown, "A Comparison of Auditor}' and Blind Separation Techniques for Speech Segregation", Tech. Rep. OSÜ- CISRC-6/99-TR15, Dept, of Comp, and Info. Sci., The Ohio State University, Columbus, OH, (1999).

FIELDS OF STUDY

Major Field: Biomedical Engineering

IX TABLE OF CONTENTS

Page

A b s tra c t...... ii

D edication ...... iv

Acknowledgments ...... v

V i t a ...... viii

List of T a b le s ...... xiv

List of Figures ...... xvi

Preface ...... xxvii

Chapters;

1. Background to multimodality neurological intensive care unit monitoring 1

1.1 Introduction ...... 1 1.2 Stroke monitoring in the neurointensive care u n it...... 4 1.3 Evoked poten tials ...... 6 1.3.1 Brainstem auditory evoked potentials...... 8 1.3.2 Somatosensor}^ evoked potentials ...... 12 1.3.3 Visual evoked potentials...... 18 1.4 The electroencephalogram ...... 22 1.5 Endovascular techniques ...... 25 1.6 O ther physiological pa ra m e te rs ...... 27 1.7 Physiological state representations ...... 31 1.8 Continuous multimodality intensive care m onitoring ...... 35 1.9 S u m m a ry ...... 37

X 2. The NICU monitoring system: Design and implementation ...... 39

2.1 Design overview ...... 39 2.2 Data acquisition hardware design ...... 41 2.2.1 Overview of hardware design ...... 41 2.2.2 The EEC amplifier and headbox ...... 43 2.2.3 Analogue to digital converter ...... 44 2.2.4 Synchronization and the programmable pulse generator . . . 48 2.2.5 Auditory, somatosensory and visual stimulators and switch­ ing hardware 49 2.2.6 The Tramscope bedside monitor ...... 53 2.2.7 The annotations term inal ...... 54 2.2.8 Electrodes and setup for te sts ...... 54 2.3 -Acquisition and review software design ...... 55 2.3.1 Overview ...... 55 2.3.2 Acquisition subsystem...... 55 2.3.3 Display subsystem ...... 55 2.3.4 Analysis subsystem...... 57 2.4 Web-based interface software design ...... 57 2.5 Archive software design ...... 62 2.6 Patient data files and fo rm a ts ...... 64 2.7 Summary...... 68

3. The NICU monitoring system: Description of operation ...... 69

3.1 Introduction ...... 69 3.2 S e t u p ...... 73 3.2.1 Scheduling tests ...... 73 3.2.2 Editing test definitions ...... 76 3.3 Data acquisition...... 82 3.3.1 -Acquiring d a t a ...... 82 3.3.2 -Adding annotations ...... 85 3.3.3 Local real-time display ...... 89 3.3.4 Paging interface ...... 90 3.4 Review and analysis ...... 91 3.4.1 Viewing raw data ...... 91 3.4.2 Deriving parameters from raw d a ta ...... 97 3.4.3 Creating parameter lists and physiological state vectors . . . 103 3.4.4 Viewing and editing derived d a ta ...... 106 3.4.5 -Analyzing the dynamics of the physiological state vector . . 107 3.4.6 Generating and viewing state space diagram s ...... 109

XI 3.5 Remote web-based monitoring and review ...... I l l 3.5.1 Navigating the web-based monitoring and review interface . Ill 3.5.2 Patient History and Problem form s ...... 115 3.5.3 Flowchart, CT and TCD form s ...... 118 3.5.4 Electrophysiological data review ...... 118 3.5.5 Vector data review ...... 123 3.6 Archiving patient d ata ...... 123 3.7 Summary...... 127

4. Raw data analysis ...... 129

4.1 Introduction ...... 129 4.2 Averaging and artifact removal in evoked potentials ...... 131 4.3 Modeling the averaged evoked potential ...... 140 4.4 Peak detection in evoked potentials...... 148 4.4.1 The zero-crossings wavelet representation ...... 150 4.4.2 Feature detection using the zero-crossings representation . . 153 4.4.3 Determining the value of the confidence vector ...... 156 4.4.4 Test results for brainstem auditory evoked potentials .... 157 4.4.5 Tracking results for somatosensory evoked potentials .... 161 4.4.6 Evaluation of the discrete detector ...... 167 4.5 Calculation of spectra and power in the electroencephalogram . . . 172 4.5.1 Spectral estimation by means of the periodogram ...... 172 4.5.2 Spectral estimation by the covariance m ethod ...... 173 4.6 Burst detection and counting in the electroencephalogram ...... 177 4.6.1 Burst detection in the time-domain ...... 177 4.6.2 Matching pursuit ...... 178 4.7 Summary...... 182

5. List and vector data analysis ...... 183

5.1 Introduction ...... 183 5.2 List analysis...... 184 5.2.1 Extrapolation ...... 184 5.2.2 Variability analysis ...... 185 5.3 Interpolating a list to form a uniformly sampled vector ...... 187 5.4 Vector analysis ...... 189 5.4.1 State space representations and attractors ...... 189 5.4.2 Dimensionality of multiparameter d a ta ...... 191 5.5 Summarv...... 193

XU 6. Monitoring protocols for experiments ...... 195

6.1 Introduction ...... 195 6.2 Standard test definitions ...... 196 6.3 Standard schedules ...... 196 6.4 Standard form s ...... 200 6.5 Standard procedures ...... 205 6.6 Standard analysis ...... 206 6.7 Summary...... 207

7. Neurointensive care unit study...... 209

7.1 Introduction ...... 209 7.2 Research protocol ...... 209 7.2.1 Patient selection criteria ...... 209 7.2.2 Parameters recorded initially ...... 210 7.2.3 Monitored p a ra m e te rs ...... 210 7.2.4 Information recorded after monitoring ...... 211 7.3 Overview of patient d ata ...... 211 7.4 .Additional analysis results ...... 227 7.4.1 Variability a n a ly sis ...... 234 7.4.2 Extrapolation analysis ...... 235 7.4.3 State space diagrams ...... 236 7.5 S u m m a ry...... 238

8. Interventional neuroradiologj'’ study ...... 241

8.1 Introduction ...... 241 8.2 Research protocol ...... 242 8.3 Overview of patient results ...... 244 8.4 Summary ...... 255

9. Conclusion...... 260

9.1 Contributions...... 260 9.2 Limitations and future s tu d y ...... 264

Bibliography ...... 265

xui LIST OF TABLES

Table Page

1.1 Summary of brainstem auditory evoked potential generators (PSP = post-synaptic potential) ...... 10

2.1 Summary of patient data file types used by the NICU system. Pa- tientname, parlistname, vectomame and imagename can be any string of text characters containing no periods or spaces. Modality is one of BAEP, SEP, VEP, EEC, Phys and Spectra; for stats files it may also be one of CT, Flowchart and TCD (web-interface data) ...... 65

3.1 Parameters and their valid ranges defined in the test definition for an evoked potential or electroencephalogram test. The last block, which contains the montage definition, is repeated as many times as there are channels ...... 77

3.2 Parameters available in principal for inclusion in the physiological test definition. These parameters are sampled at the sample rate, which is also part of the definition ...... 78

3.3 Pager codes issued by the system during or after acquisition ...... 90

3.4 Parameters derived from the brainstem auditory evoked potential wave­ forms...... 101

3.5 Parameters derived from the somatosensory evoked potential waveforms. 101

3.6 Parameters derived from the visual evoked potential waveforms. . . . 102

3.7 Parameters derived from the electroencephalogram and its spectra. These parameters are calculated for each EEC channel (F3, CP3, Ol, Oz, Fz, F4, CP4 and 02) ...... 103

XIV 3.8 Parameters that can be reviewed, edited and created using the remote web-based monitoring and review interface ...... 112

4.1 Artifact removal parameters and their possible values ...... 133

4.2 Mean and standard deviation values in milliseconds for latencies of clinically important BAEP components [CCF EP laboratory' normative data with waves II and lA' interpolated] ...... 160

4.3 Mean and standard deviation values for latencies in milliseconds of clinically important SEP components [CCF EP laboratory normative data with P25 extrapolated, normally expressed as 3 x std dev.j. . . . 164

4.4 EEC frequency band definitions. *fiow for the delta band is limited by the acquisition system’s high-pass filter ...... 172

6.1 Standard values for the test definition parameters for the evoked po­ tential and electroencephalogram tests ...... 197

6.2 Standard montages for the electrophysiological tests ...... 198

6.3 Standard selection of physiological parameters measured at the bedside for physiological test type 1 ...... 198

7.1 Summary of monitored patients (one normal and seven pathological). *DOB = date of birth. **Mod. = modality (B=BAEP, S=SEP, E=EEG,P=physiological). Lt = left, rt = right, del. = delayed, abs. = absent...... 213

8.1 Summary of monitored patients (one normal and one pathological). *DOB = date of birth. **Mod. = modality (B=BAEP, S=SEP, E=EEG,P=physiological) ...... 244

9.1 Summary of system features verified in the results. Missing parameters do not imply that the features are untested, only that no explanatory figure is included in this dissertation or that the feature is irrelevant (for example abnormal and normal review are identical) ...... 261

XV LIST OF FIGURES

Figure Page

0.1 Conceptual overview of the dissertation research ...... xxvii

0.2 Overview of research project ...... xxix

0.3 Conceptual design of monitoring system ...... xxx

1.1 Brainstem auditory evoked potential elicited by left ear stimulation in a normal subject. The top trace (Al-Cz) is derived from the electrode on the left side and the bottom trace (A2-Cz) is derived from the electrode on the right side. Waves II and IV are not identified by the NICU monitoring system ...... 9

1.2 The auditory pathways and structures involved in the brainstem au­ ditor}' evoked potential (reproduced with permission from The Wash­ ington University School of Medicine Neuroscience Tutorial [88]). . . 11

1.3 Somatosensory evoked potential elicited by left median nerve stimula­ tion in a normal subject...... 13

1.4 The pathways and structures involved in the somatosensory evoked po­ tential (reproduced with permission from The Washington University School of Medicine Neuroscience Tutorial [88]) ...... 15

1.5 Visual evoked potential elicited by left visual stimulation in a normal subject (Fz electrode was not attached during this recording). Major positivity at 100ms is PlOO peak ...... 19

1.6 The pathways and structures involved in the visual evoked potential (reproduced with permission from The Washington University School of Medicine Neuroscience Tutorial [88]) ...... 20

XVI 1.7 The relationship between cerebral blood flow (CBF) and cerebral per­ fusion pressure (CPP). Autoregulation keeps CBF constant when GPP is in the range 50 mmHg to 150 mmHg ...... 30

1.8 The vector of eleven physiological parameters of Siegel et al. (adapted from [20|) displayed on a circle diagram. States A to D represent four stable pathological states ...... 32

2.1 Networking diagram for the monitoring system with remote review and monitoring terminals ...... 40

2.2 Overall software block diagram for the NICU monitoring system. . . . 41

2.3 Hardware block diagram of the acquisition unit ...... 42

2.4 Scheme used to minimize reprogramming of EEC amplifier montages, A/D amplifier sampling parameters and pulse generator timing be­ tween tests of different types and modalities within a single schedule. * denotes reprogramming of memory in the peripheral hardware - other peripheral hardware reconfigurations are simply recalls from peripheral hardware memor}" ...... 45

2.5 Timing diagram for programmable pulse generator channels 1 and 2 during acquisition of a standard test tv^pe 1 B .\E P ...... 49

2.6 Interconnection diagram of acquisition hardware modules including the switching unit ...... 50

2.7 Internal schematic of the switching hardware ...... 51

2.8 Block diagram for the acquisition and review software ...... 56

2.9 Director}' structure of the web-based interface ...... 59

2.10 Posting scheme for the electrotrace CGI script, enabling it to generate one of four responses appropriate to the browser’s request...... 63

2.11 File formats for raw data files and their specifications ...... 66

2.12 File formats for stats files and their specifications ...... 67

xvu 3.1 Main menu window of NICU monitoring system ...... 70

3.2 Schedule editing window of NICU monitoring system ...... 73

3.3 Test definition editing window for electrophysiological tests ...... 79

3.4 Test definition editing window for physiological tests ...... 80

3.5 Data acquisition window of NICU monitoring system ...... 83

3.6 Acquisition progress report window of NICU monitoring system. . . . 84

3.7 The four screen displays of the annotations terminal ...... 86

3.8 Raw data display of the NICU monitoring system. This display shows the results of a BAEP test in a subject with a subarachnoid hem­ orrhage. Two traces are superimposed to verify repeatability of the acquired waveform...... 94

3.9 Compressed array display window of the review system ...... 96

3.10 Stacked array display window of the review system ...... 98

3.11 Derived parameters menu window of the NICU monitoring system. . 99

3.12 Compile parameter list window of the NICU system ...... 104

3.13 Display list/vector window of the NICU system ...... 105

3.14 Run further analysis window of the NICU system ...... 107

3.15 State space diagram showing the state space with coordinates con­ sisting of the ipsilateral wave I, wave III and wave V latencies of the BAEP. Points within the marked area are from seven BAEPs measured in a healthy subject and points outside the marked area are from seven BAEPs measured in a patient with subarachnoid hemorrhage ...... 110

3.16 Main menu of the web-based interface ...... 113

3.17 Overview of pages on the web-based monitoring and review interface. 114

xviii 3.18 Patient selection page of the web-based interface. The patient’s name has been replaced by the word “Abnormal” ...... 115

3.19 Data type selection page of the web-based interface. The patient’s name has been replaced by the word “Abnormal” ...... 116

3.20 Web-based interface form for editing Patient History, for the patient called “Abnormal” ...... 117

3.21 Time selection page of the web-based interface for flowchart, CT and TCD forms, for the patient called “Abnormal” ...... 119

3.22 Web-based interface form for editing CT data, for the patient called “Abnormal”...... 120

3.23 Electrophysiological test modality selection page of the web-based in­ terface, for the patient called “.A.bnormal” ...... 121

3.24 Derived parameters display page of the web-based interface, for the patient called “Abnormal” ...... 122

3.25 Raw data time selection page of the web-based interface, for the patient called “.A.bnormal” ...... 124

3.26 Single raw data trace display of the web-based interface, for the patient called “Abnormal” ...... 125

3.27 Stacked raw data display of the web-based interface, for the patient called “.A-bnormal” ...... 126

3.28 Vector display of the web-based interface ...... 126

3.29 Main menu of the patient data archiving interface ...... 127

4.1 Summarj" of raw data types and types of analysis ...... 130

4.2 The four artifact removal types. The dashed lines indicate the rejection level, which may be static or dynam ic ...... 134

XIX 4.3 “True” left and right BAEPs obtained by grand averaging a series of averaged left and right BAEPs recorded from a healthy subject. . . . 142

4.4 ‘True” left and right SEPs obtained by grand averaging a series of averaged left and right SEPs recorded from a healthy subject ...... 143

4.5 Akaike information criterion (AIC) vs model order (number of poles) for the BAEP and SEP background EEG noise. In both cases the AIC is minimized at approximately 24, and this is therefore the chosen model order for the noise ...... 145

4.6 Model of observed averaged EP incorporating variably delayed true EP and colored EEG background noise ...... 145

4.7 Pole positions for four realizations of the .\R spectral estimate for the background EEG noise of the BAEP ...... 147

4.8 Estimated spectra for the background EEG noise of the BAEP. . . . 147

4.9 Eight octave wavelet transform for the normal BAEP after median- mean filtering ...... 158

4.10 Mean feature template for peak III...... 159

4.11 Distance function h for peak III in the normal BAEP. -A. peak latency of 4 ms is identified ...... 160

4.12 Normal clinically-obtained BAEP with well-defined peaks I - V detect­ ed automatically using the zero-crossings wavelet method. Results are unprocessed ...... 161

4.13 Normal BAEP with poorly-defined peaks I - V detected automatically using the zero-crossings wavelet method ...... 162

4.14 Detected feature position vs. SNR for SEP using algorithm with a priori information on expected feature positions. Diamonds denote detected N20 latencies. Boxes denote detected P25 latencies. Dotted lines indicate true positions of N20 and P25 features. When the SNR exceeds 2, detection becomes acceptable ...... 163

XX 4.15 Detected feature position vs. delay for artificaily delayed noisy SEP using algorithm without tracking information. Diamonds denote de­ tected N20 latencies. Boxes denote detected P25 latencies. Dotted lines indicate true positions of N20 and P25 features. The slight mis­ match in positions even at high SNR is due to multiple peaks that could equally well be interpreted as the true peak ...... 165

4.16 Detected feature position vs. delay for artificaily delayed noisy SEP us­ ing algorithm with tracking information on expected feature positions. Diamonds denote detected N20 latencies. Boxes denote detected P25 latencies. Dotted lines indicate true positions of N20 and P25 features. 166

4.17 For each value in a range of SNRs, a set of synthetic EP signals is generated using the EP model, and the peak latencies are found using hill-climbing and wavelet peak detection. The MSE is calculated as the average squared error of the latencies over ever}" detected peak in each signal in the set ...... 168

4.18 A set of synthetic left BAEP signals for the Al-Cz electrode pair with the SNR value for each trace marked on the right ...... 169

4.19 The first 5 seconds (z-axis) of a synthetic left BAEP signal superim­ posed on its discrete wavelet representation ...... 170

4.20 MSE vs SNR for hill-climbing and wavelet peak detection in the left BAEP. The MSE is calculated for each SNR value as the average squared error of the detected latencies for five peaks in each of 25 synthesized left BAEP signals ...... 171

4.21 Sixty second epoch of EEG with quadratic polynomial fitted to reduce baseline drift ...... 175

4.22 Autoregressive power spectral estimates and pole diagrams for a ten second sample of EEG obtained from four electrodes in a normal patient. 176

4.23 Multichannel EEG recording in a patient exhibiting burst-suppression a c tiv itv ...... ISO

XXI 4.24 Time-frequency representation obtained by performing a matching pur­ suit on an 82 second sample of EEG containing burst-suppression activ­ ity recorded between the FT and Pz electrodes. The pursuit matched 50 signal structures to describe 95% of the energy using the Coiflet wavepacket librarj"...... 181

5.1 Heart rate signal collected in the NICU from a stroke patient, together with cubic polynomial fit and extrapolation. The polynomial was fitted to the first hour of the data and the second hour is a prediction obtained by extending the evaluated points of the polynomial ...... 185

5.2 Results of a variability analysis with a window length of 20 samples performed on the heart rate signal from a stroke patient ...... 187

5.3 BAEP wave I latency data set interpolated by cubic splines. BAEPs were recorded in a normal subject. The jagged trace in each panel represents the original set of latencies and the smooth trace represents the set of interpolated points ...... 188

5.4 Phase portraits of EEG burst-suppression activity during induced co­ ma. The burst phase is shown on the left and the suppression phase on the right ...... 191

6.1 Standard montage and timing diagram for BAEP test type 1 ...... 197

6.2 Standard montages and tim ing diagrams for SEP test tv-pes 1 and 2. . 199

6.3 Standard montage and timing diagram for VEP test type 1 ...... 199

6.4 Standard montage and timing diagram for EEG test type 1 ...... 200

6.5 Standard one-hour schedule incorporating BAEP, SEP, EEG and phys­ iological tests 201

6.6 Patient History Forms...... 202

6.7 Flow Chart Form ...... 203

6.8 CT Form ...... 204

6.9 TCD Form ...... 204

x x ii 7.1 Normal left and right BAEPs ...... 214

7.2 Normal 8-channel EEG ...... 214

7.3 Left and right BAEP results for the patient "EP". Waves III and V are delayed bilaterally as a consequence of his thalamic injuries. In each case, the results of two tests are superimposed to verify repeatability. 216

7.4 Schedule of recorded tests for BL. The schedule begins on 8/26/98 and extends over three days ...... 218

7.5 Stacked SEP array obtained in response to stimulationof the right me­ dian nerve from BL during the morning of 8/27/98. She herniated at approximately 3 a.m. Left stack shows EP and subcortical P14 and N18 responses. Right stack shows cortical N20 and P25 components. Time scale on the y-axis is 13:15 to 7:19 ...... 219

7.6 Two overlaid SEP tests results obtained from HV in response to s- timulation of the left median nerve. The N20 and P25 peaks in the cortical (CPc-CPi) response are absent, consistent with her ICH and left hemiplegia ...... 222

7.7 Two overlaid SEP tests results obtained from HV in response to stim­ ulation on the right. All components are present and none delayed. . 223

7.8 Left and right BAEP results for AH. The right waves III and V are absent, but the system automatically marks spurious peaks in an at­ tempt to find the closest match. The absence of waves III and V is consistent with right cerebellar stroke and herniation ...... 225

7.9 A physiological parameter list plot for AH. The left BAEP components are undelayed, while the right components are absent, but marked spuriously on the plot. The cardiac parameters are fairly constant throughout the recording, except at 17:15 when patient restlessness gives rise to a disturbance in the measurements ...... 226

7.10 Left and right BAEP stacks for HK. Left waves III and V are absent. This patient had hypoxic encephalopathy...... 228

X X lll 7.11 Left SEP stacks for HK. The cortical (CPc-CPi) N20 and P25 compo­ nents are absent, consistent with severe hypoxic encephalopathy. . . . 229

7.12 Right SEP stacks for HK. The cortical (CPc-CPi) N20 and P25 com­ ponents are absent, consistent with severe hypoxic encephalopathy. . 230

7.13 Parameter list showing the EEG energy in the delta, alpha and be­ ta bands at the midline occipital, left and right parietal and midline frontal electrodes. The energy in the alpha and beta bands is severe­ ly attenuated at all electrodes, with only the beta activity containing significant energy. This low frequency EEG is consistent with diffuse slowing ...... 231

7.14 Parameter list showing brainstem related parameters in HK. The left waves HI and V are absent. This patient has hypoxic encephalopathy. 232

7.15 Parameter list showing somatosensor\' related parameters in HK. The N20 and P25 components were absent bilaterally, consistent with severe hypoxic encephalopathy...... 233

7.16 Heart rate, respiration rate and mean arterial blood pressure plots for AH...... 234

7.17 Heart rate variability, respiration rate variability and mean arterial blood pressure variability plots for AH. The variability of physiological parameters may be expected to decrease in critically-ill patients [20]. 235

7.18 Left and right BAEP wave I, III and V latency plots for HK ...... 236

7.19 Two hour extrapolation of left and right BAEP wave I, III and V latency plots for HK. Polynomial is fit to data between 16:10 and 18:05. Extrapolation is from 18:05 to 19:57 ...... 237

7.20 State space diagram of BAEP components in which an attempt is made to show a transition from one region of state space to another. The data are from BL before, during and after the period of herniation. The lighter colored region corresponds to the period after herniation. 239

8.1 Stacked SEP response to left median nerve stimulation in the normal volunteer...... 246

XXIV 8.2 Stacked SEP response to right median nerve stimulation in the normal volunteer...... 247

8.3 Frontal EEG spectral stacks recorded for the normal volunteer. The low frequency activity increases in the last 20 minutes when the subject sleeps ...... 248

8.4 Parietal EEG spectral stacks recorded for the normal volunteer. . . . 248

8.5 Occipital EEG spectral stacks recorded for the normal volunteer. The alpha activity that predominates during the first half of the recording is replaced in the last 20 minutes by delta activity as the subject sleeps. 249

8.6 Parietal EEG energy in the alpha, beta and delta bands on the left and right sides for the normal volunteer. The beta activity is minimal throughout most of the recording as the subject relaxes with closed eyes. Delta energy increases significantly in the last 20 minutes as the subject falls asleep...... 250

8.7 Occipital and frontal EEG energy- in the alpha, beta and delta bands along the midline for the normal volunteer. Low frequency activity increases toward the end of the recording, when the subject sleeps. . . 251

8.8 Angiogram showing the left carotid artery' circulation in WE before balloon inflation. iN'ote the outlines of seven of the electrodes and their leads belonging to the NICÜ monitoring system ...... 253

8.9 Angiogram showing the vertebral artery circulation in WE during in­ flation of the balloon. The outline of the electrodes is present along with the outline of the transcranial doppler probe. The patient’s al­ most unaffected performance on the neurological examination and the limited electrophysiological changes during inflation of the balloon may be attributed to the good vertebral circulation shown in this figure. . 254

8.10 Energy in the alpha, beta and delta bands of the left and right frontal EEG in WE. Delta activity increases throughout the procedure and beta activity is attenuated slightly at the time of occlusion ...... 255

8.11 Energy in the alpha, beta and delta bands of the left and right parietal EEG in WE. Delta activity increases throughout the procedure. . . . 256

XXV 8.12 Energy in the alpha, beta and delta bands of the left and right occipital EEG in WE. Delta activity increases throughout the procedure. . . . 257

8.13 Energy in the alpha, beta and delta bands at the frontal and occip­ ital midline EEG electrodes in WE. Delta activity increases at both electrodes as the procedure progresses, while higher frequency activity attenuates...... 258

XXVI Preface

Figure 0.1 shows the three overlapping aspects of this research project. These are the logistical organization of the project, the design of the monitoring system and the documentation of the research in the form of this dissertation. Each of these aspects shares the same procedural flow, namely background and setup, collection of data and analysis and presentation of results. These aspects were deliberately developed in parallel to emphasize this common thread, which is also reflected in the structure of the main menu of the system, shown in Figure 3.1. Further details of each aspect are illustrated in the following three sections of this foreword.

Overview o f research project _____organization ______

Monitoring system - Layout of conceptual design dissertation

Figure 0.1: Conceptual over\dew of the dissertation research.

x x v i i Organization of research project

The organization of the research project is illustrated in Figure 0.2. The project begins with the identification of the problem, investigation of possible solutions, and definition of a chosen approach to solving the problem. The problem is to continu­ ously monitor the condition of the CNS in the NTCU patient. The most appropriate solution is multimodality electrophysiological monitoring. The task is thus to design a monitoring system to collect electrophysiological data of various modalities in real time in the NICÜ, with minimal supervision, and present easily interpretable results to the clinician who may not be trained in clinical neurophysiology. The system de­ sign requires hardware and software development. The design is validated by testing it in real patients, and the utility of this type of monitoring in the NICU is evaluated.

Conceptual design of monitoring system

The overall design of the monitoring system mimics the fiow of the dissertation research. This is in turn reflected in the chapter-for-chapter layout of the dissertation.

The system menus are also designed to parallel the flow of the dissertation research.

The main system menu therefore includes arrows and is quite similar to the conceptual system design illustrated here in Figure 0.3. The main steps in the monitoring process are setup, data acquisition, data review and data analysis. In addition, aspects of data entr>', in particular of annotations, and displaying results locally and elsewhere on the network have been considered.

xxvm Background

Problem identification: A monitor for continuously assessing brain function in the neurological intensive care unit is required.

Investigate possible solutions: Electrophysiological monitoring (electroencephalogram, evoked potentials), physiological monitoring (cardiac, respiratory, cerebral blood flow), imaging (CT, MRI).

Define overall approach: Design a single monitoring system which collects the appropriate data chosen from above in real time. Present usefully processed display to clinician. Test in appropriate population of patients (stroke patients).

System design

^ Conceptual system design

^ Hardware design

^ Software design

Testing

Experimental design

Collect and present experimental data

Derive and present conclusions from data

Figure 0.2: Overview of research project.

XXIX Setup

Set up test definitions and schedules

Acquisition Local interface Electrophysiological signal acquisition (EEG/EP) (local hardware) Local real-time display

Physiological signal acquisition (Tramscope bedside monitor) Local splash-proof terminal for annotations

Annotations (local dedicated terminal for nurses)

R em ote interface Additional Irregular time-varying data entered by hand from various sources Paging interface (page to indicate bad or terminated acquisition)

R eview Web interface (near real-time review of patient data on browser)

Display and edit raw traces and arrays of raw traces

Display and edit parameters derived directly from raw waveforms

Analysis

Automatic feature detection from raw waveforms (peak amplitudes and latencies)

Display time-varying (vector) trends

Final analysis (correlation and state-space representation)

Figure 0.3: Conceptual liesign of monitoring system.

XXX Layout of dissertation

Each chapter begins with an introductory section and ends with a summary sec­ tion. The summary includes references to any further relevant details in conference or other papers generated during the research. Chapter 1 provides the background.

Chapters 2 to 5 deal with the engineering aspects of the monitoring system and

Chapters 6 to 8 provide the experimental procedures and results.

XXXI CHAPTER 1

Background to multimodality neurological intensive care unit monitoring

1.1 Introduction

The modern intensive care unit (ICU) is equipped with a large range of monitoring

equipment. This includes equipment for monitoring many aspects of the functioning of

the circulator}' system, such as blood pressure, heart rate and the electrocardiogram.

Other systems such as the respiratory and renal systems are also well covered. One

of the most important organ systems, the central nervous system (CNS). is however

largely neglected in the ICU and the operating room [17]. Even in the neurointensive care unit (NICU) where patients are admitted specifically for critical neurological in­ juries, monitoring of the function of the CNS is limited. This project aims to address this deficiency. The most important ongoing test of neurological function is the neuro­ logical examination in which the patient's responses to various stimuli, applied by the

NICU nurse, are recorded every half hour or according to the clinician’s orders [83].

Such parameters as pupil dilation, pain response and verbal response to questioning are recorded [135]. Intracranial pressure may be monitored continuously. This is an invasive technique that involves the placement of a screw that penetrates the patient’s skull, and it is avoided if possible to reduce the chance of infection [125]. Magnetic

1 resonance imaging (MRI) and computed tomography (CT) are powerful techniques for examining the structure of the brain. Until a portable version of such a scanner is available for imaging at the bedside in the intensive care unit, the patient must be transported to the imaging center for scanning. Removal from the NICU may endanger the critically ill patient, and imaging resources are at a premium. Scanning is not repeated regularly unless it is absolutely necessary. A lot of information about the CNS may be obtained from electrophysiological data collected from electrodes attached to the patient's scalp. However, electrophysiological testing is not routinely applied in the NICU [98|[I08|. The principal reason for this is that it is difficult for a clinician untrained in clinical neurophysiology to interpret the electrophysiological data. Also, attaching and maintaining the electrodes requires a qualified technician.

The NICU monitoring system addresses this problem by automatically interpreting the electrophysiological tests, and making the data remotely available by means of a web-interface. In this way, the need for review by a trained clinical neurophysiologist is minimized, and if necessary can be done at his or her convenience on the desk­ top computer in the office. At present, the most common use of electrophysiological testing in the NICU is in comatose patients in whom subclinical seizure activity may be occuring, coma in which burst-suppression activity must be regulated, testing for brain death and testing to assist decision-making in patients who may or may not recover after an extended period of coma [132|. Electrophysiological testing is an in­ expensive, noninvasive and continuous method of monitoring the function of the CNS at the bedside, and its full potential to improve patient care in the NICU has not been realized. It is the purpose of this dissertation to report on the development and testing of a multimodality electrophysiological monitoring system for use in the neu- rointensive care unit. In addition, the system presented in this dissertation combines other physiological data with the electrophysiological information. Several authors have emphasized the value of using a combination of modalities to comprehensively monitor the critically-ill patient [26|[83l[114]. The NICU monitoring system described in this work provides tools for gathering and analyzing multimodality data. Another important feature is that it monitors continuously and allows the gathered data to be analyzed to find trends. Early changes in one or more of several parameters can provide an early indication of CNS deterioration in the NICU patient before physical signs are manifested [22] [83], and pathophysiology may thus be detected and possi­ bly reversed at an early stage. The system also provides some non-traditional tools from the complex systems approach, as the need to look at time-varying nonlinear physiological data in new ways has been highlighted by several authors [11[[89].

To limit the scope of this project, a target patient population had to be chosen, along with a set of appropriate physiological parameters to be monitored. Stroke pa­ tients were chosen for several reasons. This is a large group of patients that are treated at the Cleveland Clinic Foundation. Trauma patients might have been chosen, but they more frequently present at other primarv-care hospitals in the Cleveland area.

More specifically the group of stroke patients in w^hom edema or hemorrhage may give rise to tissue shifts or herniation was chosen, because the anatomical changes in their central nervous systems may be reflected electrophysiologically. The parameter- s chosen for monitoring consist primarily of the electrophysiological signals - evoked potentials and the electroencephalogram. Evoked potentials are the response of the brain to stimulation, and their value in detecting subtle changes in the condition of the brain has been shown by many authors [26|[128j[139|[159|. The monitoring system was designed to measure the response to auditory clicks (brainstem auditory evoked potentials or BAEPs), electrical stimulation of the peripheral nerves (somatosensory evoked potentials or SEPs) and visual flashes (visual evoked potentials or VEPs).

The system also records the electroencephalogram (EEG) at regular intervals, along with other physiological parameters recorded by the exisiting bedside monitoring e- quipment in the NICU. Continuous EEC is valuable in detecting ischemia, status epilepticus and disorders associated with tumors and coma and can be an early indi­ cator of CNS deterioration [22]. Due to the limited number of appropriate patients in the NICU, data were also collected from patients undergoing balloon angioplasty and angioplasty with stent placement in the interventional neuroradiology (INR) suite.

These patients undergo known changes over a short period of time, thus providing data with dynamical changes that may in principal be observed over a longer period of time in NICU patients. The electrophysiological data along with other physiological data are assembled into a state vector. It has been observed that the healthy state frequently corresponds with a varying physiological signal. For example heart rate should be somewhat variable in the healthy subject [42]. If it is too stable, this indi­ cates a pathology'. Conversely the pathological condition may correspond to a stable physiological state. In this work, the dynamics of the physiological state observed in

NICU and INR patients are investigated in the context of these observations.

1.2 Stroke monitoring in the neurointensive care unit

Of the various types of patients in the neurointensive care unit (NICU), the group of patients who suffered a stroke was chosen for this study. More specifically, the group of stroke patients in whom edema or hemorrhage gives rise to tissue shifts with

possible risk of herniation was chosen. This group was chosen because it is a group of

significant size and thus represents a patient population in which a real contribution

can potentially be made to their treatment. In fact, stroke is the third leading cause

of death in the United States, and the leading cause of adult disability [52|.

Stroke is a disorder of cerebrovascular origin in which brain cells die as a result

of compromised blood supply. The effects may be extensive, such as when the heart

fails to supply sufficient blood or sufficiently oxygenated blood to the brain (hypoxia)

and when the occlusion of a large vessel gives rise to widespread ischemia. They may

also be more localized - emboli may occlude the small cortical vessels, arterioles may

thicken and close causing small infarctions, and hemorrhaging may occur. Eighty

percent of strokes are ischemic and the remaining 20% are hemorrhagic, consisting of

13% intracranical hemorrhage (ICH) and 7% subarachnoid hemorrhage (S.A.H).

Evoked potentials (EPs) and the electroencephalogram (EEG) are discussed in

detail in Section 1.3 and 1.4 respectively. WTiile non-specific, the EEG is a very

sensitive detector of hypoxia and ischemia and detects dysfunction at a reversible

stage [45]. The most common use of ordinary EEGs in the NICU is to identify whether

subclinical seizures are occurring. The most common use of EPs in the NICU is the

prognosis of coma. For example, the bilateral absence of cortical components in the

SEP indicates a poor prognosis. It has been shown that the SEP is not useful in

monitoring focal cerebral ischemia [50]. Abnormal EPs may prompt early action or

further investigation, which would otherwise not have been done until more obvious symptoms manifested themselves, such as greatly elevated intracranial pressure (ICP)

in the case of hemorrhage. The most common cause of early death in stroke is cerebral edema, 24 to 96 hours after onset. One way to ascertain the success of this study is

to determine how often the system could potentially catalyze early action which is

beneficial to the patient and the extent of this benefit, as seen against the cost of the

system (including the cost of false alarms). An example of such action in the case of

elevated ICP is surgical decompression.

1.3 Evoked potentials

An evoked potential is the electrical response of the brain to stimulation. For

example, the brainstem auditory evoked potential is the response of the brainstem

to auditory stimulation of the ear. When a tone is applied to the ear, it takes about

a millisecond to be conducted along the outer ear and structures of the middle ear

and to be transducted in the cochlea to produce an impulse in the acoustic nerve.

WTien the impulse is conducted across a synapse or possibly when it is conducted

along the nerve itself, volume conduction transmits this electrical activity to the skin

surface where it may be recorded as a very low amplitude signal. The amplitude of

the response is typically a few microvolts superimposed on background EEG activity

of which the amplitude is one or two orders of magnitude higher. After another

millisecond, the impulse has travelled to the cochlear nucleus in the brainstem and the

response is again refiected at the skin surface as a small deviation. Several structures

along the auditor}'- pathway produce such responses. In order to extract these signals

from the background noise, the responses are averaged over hundreds of cycles of stimulation. After averaging, a well-defined and repeatable response is obtained.

The time between peaks in this response corresponds with the time that the impulse

takes to travel between structures in the brain. If the pathways are compromised. the resulting conduction delay will be reflected in the evoked response. By applying various types of stimulation that are conducted along various pathways within the brain, the function of different subsystems of the brain may be tested. It should be noted however that these are rather narrow neural pathways - if the evoked response is compromised, a lesion affecting the pathway may be suspected. But it is quite possible to have a lesion that happens not to affect any of the evoked potential pathways, in w'-hich case the lesion will not be detected by evoked potential monitoring.

Nuwer et al. point out that evoked potentials are an established method for de­ tecting changes in neurology. The normal limits are well-established and the tests are objective, sensitive and robust [103|. The novelty of this research lies in the inves­ tigation of regularly recorded evoked potentials over an extended period, obtaining the clinically significant features, and tracking their progress over time. It is not the absolute value (latency or amplitude) of the features, but their trends which are most important when monitoring the NICU patient [85|[139|. Improvement or dete­ rioration in the evoked response could be prognostically useful and indicate the need for intervention at an early stage. In this sense, patients can be their own baselines, and features other than those that are traditionally considered as clinically important may be useful. Approximate localisation of a lesion or region of abnormality (neural tract injury or compression) may be achieved not only by identifying abnormalities in corresponding features (peaks and troughs) in the EPs, but by asymmetry between the left and right responses. This is the motivation for unilateral stimulation. Visu­ al, auditory and somatosensor}- evoked potentials are useful in detecting defects in different regions: visual - optic nerve, subcortical and cortical; auditory - auditory nerve, cochlear nucleus, lower/upper brainstem, cerebral connections; somatosenso­ ry - brainstem (lateral lemnisci), thalamus. Each of these modalities is discussed separately below and a diagram showing the relevant anatomy is included in each case.

1.3.1 Brainstem auditory evoked potentials

The auditory evoked potential (AEP) is typically elicited by broadband click t\-pe stimulation. Earphones are used, and in this study tubal insert earphones were used for convenience in the NICU. Stimulation is monaural (unilateral) - this is important since as\Tnmetries in response are clinically significant in the AEP. White background masking noise in the ear contralateral to stimulation is used to isolate that ear from stimulation conducted from the stimulated side. Typical stimulation parameters in use at the Cleveland Clinic Foundation (CCF) in recording brainstem auditory evoked potentials (B.\EPs) are 100 ^s click-type stimulation (rarefaction or condensation) at a rate of 11.71 Hz averaged over 2000 events. It may be necessary to adapt the number of events dynamically depending on the situation. Left and right recordings may be made from .\1 and A2 respectively, referenced to Cz. The experimental protocol relating to the clinical experiments in this study concerning BAEPs is described in

Section 6.2.

The BAEP is the averaged electrical response for the first 10 ms after stimulation.

A typical BAEP response to stimulation to the left ear in a normal subject is shown in

Figure 1.1. The BAEP is a faithful indicator of neurological condition, which may be obtained regardless of the state of attentiveness or position of the patient. It reflects acoustic nerve and brainstem condition. Middle latency AEPs (the response 10 ms

8 Pabenc Normal

T e st SA £P{type 1) left (1000 c y d es a t 11.76 Hz )

Time: Tua Mar 3 16:36:56 1998

0.10

Channel i

SOHz-iSOOHz

50uV (FS)

Channel 2

50HZ-1500HZ

50uV(FS)

Time (ms) 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 10.0C

Figure 1.1: Brainstem auditory- evoked potential elicited by left ear stimulation in a normal subject. The top trace (Al-Cz) is derived from the electrode on the left side and the bottom trace (A2-Cz) is derived from the electrode on the right side. Waves II and IV are not identified by the XICU monitoring system.

to 50 ms after stimulation) reflect early cortical excitation and are traditionally more useful for audiological evaluations. Long latency AEPs (the response later than 50 ms after stimulation) are relatively inconstant (varying with attention) and correspond to later cortical excitation. Only the BAEP will be considered in this study. Five well-defined and repeatable peaks characterize the BAEP. They are denoted by the

Roman numerals I to V. Peaks I, III and V are of particular clinical significance.

The brainstem auditory^ evoked potential generators have been fairly well char­ acterized, with a considerable amount of work done also in animal subjects. Moore BAEP feature Generator Wave I Graded potentials at dendritic terminals of cochlear nerve Wave II PSPs of the cochlear nucleus Wave III PSPs of the medial superior olivary nucleus and input to trapezoid body Wave IV PSPs of the ventral nucleus of the lateral lemniscus Wave V Deep ventrolateral part of the inferior colliculus

Table 1.1: Summary' of brainstem auditory evoked potential generators (PSP = post- synaptic potential).

[90] makes some important points - BAEP waves reflect volume-conducted potentials from the brainstem. Only neurons are responsible (not glia). The summation of graded post-synaptic potentials (PSPs) and not action potentials is responsible for the components in B.A.EPs [90]. Misulis [86] states that EPs are generated not only by synaptic potentials but also by action potentials propagating in nerve tracts. Misulis attributes cortical EPs to synaptic transmission in the cortex and action potential propagation in thalamocortical projections. Such responses occur in SEPs and VEPs.

Except for the late components which are not used clinically. .A.EPs are of exclusively subcortical origin. The generators of the five positive peaks of the B.A.EP are given in Table 1.1, and the corresponding anatomy is illustrated in Figure 1.2.

Misulis [86] observes the following changes to the BAEP in the abnormal case

- lower brainstem (acoustic neirve to lower pons) lesions result in increased I-III la­ tencies. Upper brainstem (lower pons to midbrain) lesions result in increased wave

III-V latencies. Central lesions give rise to a I-V latency difference between sides.

Coma of metabolic or structural origin may cause BAEP abnormalities. The BAEP is not affected by ordinary anesthetics. Alcohol affects the latencies of wave II and beyond. Transient ischemic attacks involving the auditory tracts of the brainstem

1 0 auditors cortex ^

medial geniculate inferiOT / colliculus

^ lateral lemniscus

to inferior colliculus to infenor colliculus dorsal cochlear dorsal cochlear lateral lemniscus' ■vr-^nudeus ^lateral lemniscus «^..-^nudeus ventral ventral cochlear cochlear nucleus nucleus superior olive

VUIth v m th nerve nerve

Figure 1.2: The auditor)" pathways and structures involved in the brainstem auditor)' evoked potential (reproduced with permission from The Washington University School of Medicine Neuroscience Tutorial [88]).

11 give rise to transient BAEP abnormalities. If the auditory tracts are not involved,

the BAEP may be normal, even though there could be a major brainstem infarction

-lesions must involve the low midbrain structures or below for the BAEP to reflect

the abnormality [50]. Intracranial pressure (ICP) in excess of 30 mmHg and signif­

icantly prolonged wave V latency indicates that surgical/medical decompression of

ICP should be performed [50|. The age, sex, body temperature and possible hearing

loss of the patient should be taken into consideration when interpreting the BAEP.

Circadian variations in BAEPs may be due to circadian variations in body tempera­

ture. Miskiel et al. used automatic peak detection to identify peaks I, III and V in

the BAEP in human subjects in the ICU and trended the latencies. They concluded

that automated long-term BAEP monitoring has clinical potential in the ICU [85|.

1.3.2 Somatosensory evoked potentials

The somatosensory evoked potential (SEP) is the brain’s response to stimulation

of a peripheral sensory nerve. In this study, the scalp SEP elicited by stimulation of

the median nerve at the wrist was used. Since the cortical response was of interest,

and this can be elicited by either posterior tibial nerve stimulation at the ankle or

median nerve stimulation at the wrist, the latter was chosen. The response may be

obtained with a smaller stimulating current in the median nerv’e case and therefore

results in less patient discomfort. .A. typical SEP response to stimulation of the left

median nerve in a normal subject is shown in Figure 1.3.

The scalp SEP response to median nerve stimulation reflects thalamic and other subcortical activity. Parameters typically used in recording SEPs at CCF are 2.71 Hz

1 2 PaMnc Normal

Test SEP (type left (250 cycles a: 27i Hz )

Time: Tue Mar 3 17:31:56 19S8

1 SO

Channel i

C 3 - - E P 1

5Hz-t500Hz

50uV (FS)

1-50

C 4 - E P 2

50uV (FS)

- 1 5 0

1 5 0

C hannel 3

C S S - F z

SOuV (FS)

-1.50

C hannel *

SHz-ISOOHz

50uV (FS)

Time (m$) 1 000 500 1000 1 5 0 0 20 0 0 25 00 30 00 3 5 00 * o 'o o *5 0 0 S O o J

Figure 1.3: Somatosensory evoked potential elicited by left median nerve stimulation in a normal subject.

13 electrical stimulation averaged over 2000 events. The experimental protocol relating to the clinical experiments in this study concerning SEPs is described in Section 6.2.

The SEP measured on the scalp consists of a few low-amplitude far-held compo­ nents, such as P15, P16 and P18, followed by the near-held components N20, P25,

P30 and others [117]. P15, P16 and P18 originate in subcortical structures along the sensory pathway, viz. the ventral posterolateral nucleus of the thalamus (P15) and the thalamocortical tracts (P16 and P18). N20 is of controversial origin [86|. It probably originates in the parietal cortex. Later components also originate in the motor cortex and somatosensory cortex. Figure 1.3 shows the traces obtained during a typical SEP recording. Further labelled examples are given later in this document

(see for example Figure 4.4). The anatomy of the sensory pathways is illustrated in

Figure 1.4.

.A. number of investigations have been reported in which the prognostic value of the SEP in comatose patients is evaluated. The experimental design is similar in all of them. .A. group of unresponsive or unconscious patients is selected, or pa­ tients with low Glasgow coma scale values are chosen. The SEP to median nerve stimulation at the wrist and possibly other EPs are recorded on a single occasion a few days after injury. The SEP is classified numerically according to some grading system - usually ranging from abnormally high cortical N20 latency to bilateral ab­ sence of the N20 peak. The patient's final outcome is classified at the appropriate time according to some scale such as the Glasgow outcome scale, or often a simple favourable/unfavourable classification. The SEP grades and final patient outcomes are compared to find a correlation.

14 cerebral cortex

midbrain

pons

rostral medulla

caudal medulla

cervical cord from arm

lumbar cord fi-oni leg

Figure 1.4: The pathways and structures involved in the somatosensory evoked poten­ tial (reproduced with permission from The Washington University School of Medicine Neuroscience Tutorial [88|).

15 All studies seem to agree that the SEP is accurate in predicting an unfavourable outcome, and useful because it is resistant to medications which suppress the EEG.

For example, Nau et al. [98| reported that loss of the N20 peak corresponds with brain death. Anderson et al. [9] found that 100% of 23 patients with unilateral or bilateral absence of the N2Ü component of the SEP had an unfavourable outcome.

In this study an unfavourable outcome was defined as severe disability, vegetative state or death. Madl et al. [72] reported that 100% of 86 comatose patients with bilateral loss of the N20 peak eventually died. Ying et al. [157] reported that SEPs were abolished simultaneously and bilaterally only in brain dead patients.

Other EP modalities (VEPs and BAEPs) are reliable indicators of an unfavourable outcome, but only the SEP is also a fairly reliable predictor of a favourable outcome

[9|. -\nderson et al. found that of 11 comatose patients with a fairly normal SEP, with at worst abnormal latencies, 9 had a favourable outcome. The BAEP and VEP were less reliable at predicting a favourable outcome. Madl et al. [72] reported that in

355 comatose patients with preserved N20 components, 42% survived and 58% died, thus suggesting that the SEP is fairly unreliable in predicting a favourable outcome, at least if only the presence of the N20 component is considered. De Giorgio et al.

[32] suggest that a combination of the SEP and the P300 event-related potential are better than the EEG for predicting a favourable outcome, whereas a combination of the SEP and the EEG are better than the P300 for predicting an unfavourable outcome. In all of these studies, the patient groups contained a diversity of brain injuries - brain death, trauma, anoxia, encephalitis and infarction. Madl’s study [72] concentrated only on nontraumatic coma and he cautions that their conclusions do not apply to children or patients with traumatic head injury. The N20 component

16 was of greatest interest in these studies. This originates in the parietal cortex or thalamocortical tracts. An injury involving this region or the subcortical structures along the sensory pathways may result in N20 latency and amplitude changes.

It is mostly agreed that the SEP is not often falsely pessimistic, although Schwarz reports four patients who recovered favorably after bilateral loss of the N20 and P25 components of the SEP. Schwarz ascribes this to the fact that there was no permanent structural damage in these patients [126|. In any patient, and especially in potential survivors, it may be sensible to consider other clinical and electrophysiological data before coming to any conclusions [157|. There seem to be few data on continuous SEP monitoring and its possible value in prognostication in comatose patients and this is explored in Chapters 7 and 8. Yang et al. describe serial studies in patients with hepatic encephalopathy. This condition is associated with cortical dysfunction - the

N20 and P25 components of the SEP are affected. In their studies, they correlated the course of the disease with the increase in latencies and later disappearance of the N20 and P25 components, and found that an additional increase in the N13-N20 latency bears a poor prognosis [155|. Thakor et al. tracked the progression of the SEP in cats subjected to hj'poxia and humans at various levels of anesthesia and found a correlation [139|.

Misulis [86j observed the following changes to the SEP in the abnormal case - the

N20 and P25 features in the scalp SEP may be significant. Lesions in a single so­ matosensory pathway (including decreased peripheral conduction) result in increased

SEP latencies and a latency difference between sides. Defects above the lower medulla and at or below the somatosensory" cortex and peripheral nerve lesions can result in absence of the N20 waveform. In “locked-in syndrome” in stroke, the SEP is abnormal

17 if the pontine lesion extends into the medial lemnisci of the tegmentum. Thalamic and brainstem lesions result in an N20 delay. Parietal infarcts result in N20 abnor­ mality (absence or reduced amplitude). In stroke patients, a decrease in amplitude of the N20 wave bears a poor prognosis. Long-term monitoring of SEP during imminent brain death has been reported by Buchner et al., according to Hilz et al. [51|. The age, limb length, sex (related to limb length differences), limb temperature and sensa­ tion (numbness) of the patient should be taken into consideration when interpreting the SEP.

1.3.3 Visual evoked potentials

The visual evoked potential is the brain’s electrical response to visual stimulation.

Most stimulation types (checkerboard, Ganzfeld etc.) are inappropriate for regular monitoring in an NICU patient. Only diffuse light stimulation by means of LED goggles is practical. Unilateral stimulation at a rate less than 2 Hz is used. To ex­ ploit the information obtained by unilateral stimulation, left, right and mid-occipital electrodes are appropriate.

The VEP is considered because stimulation is straightforward and the response amplitude is large, so that only a few stimulus periods are needed to obtain a mean­ ingful average. The transient VEP response to diffuse stimulation in a normal subject is shown in Figure 1.5. The entire response is rather inconstant, but major features may be useful. Responses before 50 ms are of subcortical origin. These are not well- defined. Responses between 50 ms and 250 ms are of cortical origin and may be useful.

The rhythmic after-discharge occurring beyond 250 ms is clinically insignificant. Two peaks are used viz. peaks III and IV (or possibly peak II instead of IV). Peak III is

18 Patient: Test: VEP (type l)letl(100 cycles at 2.11 Hz) Time: Mon Jul 19 162856 1999

6.00

Channel 1 01 - MF 1HZ-70HZ SOuV (FS)

- 6.00 6.00

Channel 2 0 2 -M F 1HZ-70HZ SOuV (FS)

- 6.00 6 .0 0

Channel 3 O z-M F 1HZ-70HZ SOuV (FS)

- 6.00 6.00

Channel 4 P z-M F 1HZ-70HZ SOuV (FS)

- 6.00 Time (ms) 0.0 20.0 4().o 60.0 ad.o loo.o 120.0 140.0 160.0 isb.o 200.0

Figure 1.5: Visual evoked potential elicited by left visual stimulation in a normal subject (Pz electrode was not attached during this recording). Major positivity at 100ms is PlOO peak.

19 right Kemifield

optic nerve Meyer's loop

optic tract—

LGM

optic radiations

VI occipital poles

Figure 1.6: The pathways and structures involved in the visual evoked potential (reproduced with permission from The Washington University School of Medicine Neuroscience Tutorial [88|).

the major positive peak with latency between 50 and 100 ms. Peak IV is the major negative peak following peak III with latency between 100 and 250 ms. Peak II is the major negative peak preceding peak III with latency greater than 50 ms.

Although only the VEP to diffuse light stimulation is considered in this study, more is known about the origins of the visual evoked response to checkerboard stim­ ulation. It may be helpful to review this here. The VEP to checkerboard stimulation

2 0 consists of three main peaks - a major positive peak at around 100 ms (named PlOO,

PI or Cl) and two negative peaks at around 75 and 145 ms named N75 and N145 or

N1 and N2 respectively. There has been some controversy over the years as to the origin of these peaks and the story is chronicled in Regan [117]. Halliday et al. [117] state that all peaks originate in cortical areas 18 or 19, not the striate cortex. Lesèvre et al. [117] state that NI originates in striate area 17, and that the 2nd component

(PI) consists of two subcomponents at 100 ms and 120 ms originating from areas 18 and 19 respectively. In any case, certain cortical areas in the brain's occipital region

- areas which do not seem to be agreed upon exactly - give rise to the VEP. The anatomy of the visual pathways is illustrated in Figure 1.6.

If only one half of the visual field is stimulated by the checkerboard pattern, the peak response varies asymetrically in amplitude with the position at which it is recorded relative to the midline. Since unilateral stimulation is conducted contralat- erally by the visual pathways, it is paradoxical that the greatest positivity in the main

(PI) component of the VEP occurs on the ipsilateral side. This phenomenon is called

“paradoxical lateralization” and illustrates the principle of “paradoxical localization”, which applies to an area smaller than a hemisphere. The Queen Square group [117] suggested that paradoxical lateralization occurs because the contralateral generators are situated on the medial and posteromedial surface of the visual cortex and are transversely oriented. As a result of this orientation, the best lateral (surface) record­ ing of their activity is obtained from the ipsilateral side, even though the generators themselves are contralaterally situated. Surface EP distributions do not necessarily reflect the positions of the underlying cortical generators in a straightforward manner.

2 1 Misulis [86) observes the following changes to the VEP in the abnormal case -

optic nerve compression gives rise to VEP changes. Intracranial pressure increase

and hydrocephalus result in increased latencies in VEPs. These last two points may

be related inasmuch as an increase in ICP may result in optic nerve compression.

Acute cerebral anoxia results in VEP deterioration and disappearance. Brain death

results in VEP absence, but VEP absence doesn't imply brain death, since brainstem

activities required for survival don't influence the VEP. The age and pupil size of the

patient should be taken into consideration when interpreting the VEP.

1.4 The electroencephalogram

The electroencephalogram (EEG) and evoked potentials (EPs) change in a pre­

dictable way with the progression of cerebral ischemia. The EEG is a particularly

sensitive though nonspecific detector of ischemia. The sensitivity of the EEG to is­

chemia can be attributed to the fact that the pyramidal cells of cortical layers 3 and

5 are most susceptible to infarction, and they are also the primarv' generators of the

EEG. EEG is also used in the NICU to identify nonconvulsive seizures, nonconvul-

sive status epilepticus and burst-suppression and in the operating room for monitoring

perfusion, sedation and depth of anesthesia [17]. Seizures occur in a surprising num­

ber of NICU patients - Minahan et al. [83] report that 29 of 100 comatose patients

in a particular study had nonconvulsive seizures. Ongoing EEG monitoring to detect

the early stages of ischemia may be helpful, for example, in subarachnoid hemorrhage

(SAH) patients - Minahan et al. [83] observed that delayed cerebral ischemia occurs in 32% of SAH patients. Although the amplitude of the EEG and EP waveforms

22 is quite variable between subjects, it may still be helpful if monitored continuous­ ly in the ICU. While the absolute amplitude in any specific individual may not be meaningful, changes in amplitude that occur in response to treatment or the natural progression of the illness may be diagnostically and even prognostically useful. The

EEG changes predictably with a decrease in cerebral blood flow (CBF). The EEG slows when CBF falls in the range 16 to 22 ml/100 g/min. Its amplitude decreases when CBF decreases further to the range 11 to 19 ml/100 g/min. When CBF falls be­ low 10 ml/100 g/min, EEG activity is minimal or absent altogether. Severe ischemia, defined as CBF below 10 ml/100 g/min, gives rise to cell death within minutes.

Slowing of the EEG may be assessed by spectral methods. Quantitative EEG

(QEEG) is a representation of the Fourier transform (spectrum) of the EEG. It is sometimes displayed as the EEG power vs time in different frequency bands. This has the advantage that it is easier to interpret for non-electroencephalographers. Con­ versely, however, it is more difficult to separate artifact from true changes when presented only with the QEEG. Trained clinical neurophysiologists will seldom be satisifed with only the QEEG record, and will insist on at least a sample of the raw

EEG. In this work, autoregression and Fourier transformation are used to estimate the spectrum of the EEG. The approach is described in Section 4.5. Care should be taken to distinguish pathological shifts in the EEG spectrum from normal shifts due to sleep and alertness. Suzuki et al. [136] have used the percentage time of delta activity in the EEG as a diagnostic indicator. Furthermore, short-term changes in the EEG should not be ignored - Hal Unwin et al. emphasize that varying EEG pat­ terns are associated with a favourable prognosis, while monotonous patterns imply a poor prognosis [141]. This ties in with the notion that homeostasis or the healthy

23 “normal” state is a dynamic or even chaotic state rather than a static one. This idea is discussed further in Section 1.7.

Since the useful information in the EEG is generally at the lower end of the frequency range, it is possible in principal to measure EEG continuously, even while

BAEP’s and SEP’s are being measured. EEG monitoring during VEP’s would not be possible by this method, since the spectra overlap. Hilz et al. applied such frequency selective monitoring in their monitoring system [51].

Other features of the EEG may be prognostically significant, such as patterns of suppression, burst suppression, seizures and cycling [50] [141]. Ideally detectors for these types of features should be implemented in the NICU monitoring system, but this is beyond the scope of this study. Detectors for various EEG background states by frequency band analysis, and spindle, blink and spike detectors are described by

Gotman and Wang [43|.

Continuous monitoring of EEG in the ICU is supported by several authors. Y- oung notes that continuous EEG detects nonconvulsive seizures (which lead to death in 33% of patients) and nonconvulsive status epilepticus (which leads to death in 57% of patients) [158], and Chiappa states that continuous EEG reveals non-convulsive seizures in a larger number of patients than traditionally suspected. Chiappa encour­ ages the careful integration of BAEPs, SEPs and the EEC for improved monitoring in the ICU. Vespa et al. observe that continuous EEG detects vasospasm after SAH in ICU patients [148] and report that the use of continuous EEG (CEEG) is becoming more widespread in the NICU [149]. Apart from the increasing availability of suit­ able technology, the realization that CEEG has the potential to prevent secondary neuronal injury in the NICU has contributed to this tendency. Secondary neuronal

24 injury refers to nonconvulsive seizures and ischemia that may arise during the course of the NICU patient’s treatment. These may have serious consequences for the pa­ tient if they go undetected. Vespa et al. point out that EEG and EP monitoring in the operating room have been proved to save more than four times their own cost in prevented injury, and suggest that this may also be true in the NICU [149]. .Jordan reports that the value of CEEG has been well established in diagnosing and managing noncomuilsive status epilepticus in the NICU [58]. He suggests that it may also be useful in monitoring cases of precarious cerebral ischemia and post-SAH vasospasm.

Rampil demonstrates that the bispectral index, derived from the spectral content of the EEG, correlates with the level of sedation induced by nitrous oxide in human subjects [115]. Quinonez states that EEG has excellent temporal resolution and as such it is a unique method for detecting early deterioration at a reversible stage in the OR, ICU and INR [113|. Buzea notes that computerized EEG detects ischemia and other conditions before physical signs appear [22] and Si implemented an expert system that monitors EEG and raises the alarm if an abnormality is detected that requires the attention of a trained clinical neurophysiologist. Kay cautions that while continuous EEG is important for ICU monitoring, controlled studies are still required to assess its role [611.

1.5 Endovascular techniques

Interventional neuroradiology (INR) is defined by Young et al. as "treatmen- t of central nervous system (CNS) disease by endovascular access for the purpose of delivering therapeutic agents including both drugs and devices” [159). INR is a combination of neurosurgery and neuroradiology which is gaining popularity.

25 The motivation for studying INR patients in this research is that they represent a group of patients in whom known, induced changes to the CNS occur over a short period of time. Although INR could potentially benefit from electrophysiological monitoring, the involvement of INR patients in this research is simply to use them as a model for the dynamic changes that may occur in stroke patients over a longer period of time.

In INR, a special catheter is inserted in the femoral artery and guided within the vessel to the artery of interest in the head or neck. The surgeon may guide the catheter by means of a wire or allow the blood flow to direct the catheter. A “road­ map” of the cerebral vasculature is obtained by injecting radio-opaque material into the artery by means of the catheter, and then capturing an x-ray image of the head in the manner of conventional angiography. The catheter used to perform the INR procedure is radio-opaque. As it is guided to the site of surgery, x-ray images show its position. Since the radio-opaque material has long since dissipated, the vasculature is no longer visible. Provided that the patient has not moved, digital image subtraction can be used to superimpose the image of the catheter on the road-map.

Typically, the neurological examination is sufficient to monitor the condition of the CNS in INR patients. Of particular concern is ischemia that may be planned or may inadvertently occur due to vascular occlusion during the procedure. Young et al. suggest that EEG, SEP, TCD and Xe CBF monitoring may also be beneficial, especially in anesthetized patients [159]. EEG is a sensitive and rapid detector of cortical ischemia and the SEP is sensitive to ischemia of deeper structures. TCD may be used to detect vasospasm in subarachnoid hemorrhage, vessel narrowing and brain death and is used in INR to evaluate the occlusion of a vessel [76].

2 6 The group of INR patients of interest in this study are those undergoing balloon

angioplasty and angioplasty with stent placement. These procedures are for treating

occluded vessels - most commonly the vertebral arteries. A balloon is ahgned with the

area of stenosis within the vessel and inflated. In some cases a metal stent is placed

in the stenotic area and expanded to stretch the vessel. In either case, the distal

circulation should improve after a successful procedure. The improvement may be

reflected in the electrophysiological measurements. These may fluctuate during the

procedure too, when hypotension distal to the balloon is induced momentarily while it is inflated. It is important to take into account the activities of the anesthetist who may manipulate the patient’s systemic blood pressure during the procedure to compensate for the induced changes in cerebral blood pressure.

1.6 Other physiological parameters

Minahan et al. reviewed the crucial variables monitored in the intensive care unit in patients with cerebrovascular disease [83|. They divided these variables into five categories, namely: a) the clinically apparent function of the system (eg. neurological examination), b) physical and mechanical variables of the system (eg. intracranial pressure), c) circulation or perfusion of the system (eg. cerebral blood flow), d) bio­ electric measures of system function (evoked potentials and electroencephalogram), and e) biochemical measures of system or cellular function (eg. microdialysis). In the research described in this dissertation, the idea of a state vector representing the neurological state of the ICU patient is pursued (see Section 1.7). A comprehensive

27 neurological state vector would include not only the electroneurophysiological vari­ ables discussed in Sections 1.3 and 1.4, but also the relevant variables from the other categories described by Minahan et al. [83].

The neurologic examination is a test which is performed regularly to monitor the state of a patient and detect any signs of deterioration [135]. In awake, cooperative patients, it represents a comprehensive test of the various functions of the nervous system and is sensitive to any changes in these functions. In comatose or sedated patients, or patients with muscle-relaxants, it is less sensitive, and this is the case in which a combination of Minahan's other parameters may be especially useful. The

Glascow coma scale is an example of a simple and useful neurological examination which is commonly used to monitor the progress of critically ill NICU patients. It is a scalar assessment of the patient state which takes on a value in the range of integers from 1 to 9, combining eye opening, best verbal and best motor response.

Other functions (cortical, motor and brainstem) may be included in a more detailed neurologic examination.

Although it is commonly agreed that intracranial pressure (ICP) monitoring is pre­ dictive of outcome after severe brain injury, its value in guiding treatment is doubtful

[125]. Allen describes a protocl for management of increased intracranial pressure [8|.

Frank [37] demonstrates how indiscriminate ICP-altering inter\-entions may even ex­ acerbate injury by increasing pressure differentials which can give rise to tissue shifts, with concomitant tissue distortion and herniation. In such cases, he argues in favour of surgical decompression.

28 Normal resting ICP is below 15 mmHg [83]. Sustained values in excess of 15 mmHg can be tolerated if they develop chronically, such as in the case of a tumor, but in the acute case, such as hemorrhage, they portend a poor outcome.

Cerebral perfusion is driven by the intracranial-mean arterial pressure difference, or cerebral perfusion pressure, defined as

CPP = M AP - ICP where CPP is the cerebral perfusion pressure, MAP is the mean arterial pressure and ICP is the intracranial pressure. The CPP gives rise to flow - the cerebral blood flow (CBF) - but the relationship is not linear. The brain’s autoregulatory mechanism keeps CBF constant provided that the CPP remains between 50 mmHg and 150 mmHg. Outside this range, the autoregulatory mechanism breaks down -

CBF decreases when CPP falls below 50 mmHg and increases when CPP exceeds

150 mmHg. This is illustrated in Figure 1.7. The autoregulation mechanism is also influenced by the availability of oxygen, which is in turn related to blood oxygen saturation, and indirectly to respiration rate. ICP increases if the volume of the contents of the cranial vault increases, since the cranium is rigid. The presence of a mass, such as hemorrhage or edema, can increase ICP. The standard method for relieving elevated ICP is to drain the cerebrospinal fluid using an intraventricular catheter. The fear exists that elevated ICP may give rise to herniation or tissue shifts. For this reason, ICP control is considered necessary to prevent secondary cerebral injury [83]. In certain cases, Frank [37] and Rieke et al. [119] would suggest that surgical decompression is a better therapy.

Measuring CBF directly requires sophisticated techniques, such as single pho­ ton emission computed tomography (SPECT) which is not convenient for bedside 29 CBF

. CPP 50 150 (mmHg)

Figure 1.7: The relationship between cerebral blood flow (CBF) and cerebral perfu­ sion pressure (CPP). Autoregulation keeps CBF constant when CPP is in the range 50 mmHg to 150 mmHg.

monitoring in the ICU. This information is useful in assessing the state of the au­ toregulatory mechanism in the patient and the risk of ischemia. Transcranial doppler sonography (TCD) is the most convenient method for indirectly measuring CBF in the ICU. TCD measures cerebral blood flow velocity. An ultrasonic probe is applied to the skin at one of a few “ultrasonic windows” (gaps or thin regions in the skull) and directed so that the intracranial arterj^ of interest is within its beam. The depth is selected by means of the pulsed Doppler principle. The velocity of blood flow can be measured in several arteries, in particular the anterior, middle and posterior cerebral arteries (ACA, MCA and PCA) and basilar artery. The pulsatility index (PI) is an important parameter in TCD sonography. Although it is defined in several ways, it is typically some version of the range of flow velocities (minimum to maximum) occur­ ring in the artery during a cardiac cycle, normalized to the mean flow velocity during

30 the cycle. TCD is used to detect vasospasm in the NICU, particularly in patients with

SAH. It is also used to assess vessel narrowing and verify brain death [76]. Jordan states that CEEG can detect vasospasm-induced cerebral dysfunction before TCD

[58|. Presumably the dysfunction is detected early with CEEG, but the specificity that would have been afforded by TCD is lost. Krieger suggests that a combination of serial EPs and continuous TCD may provide good coverage to identify early changes in NICU patients that would otherwise go unnoticed [65].

Recently it has become possible to study the neurochemistry of the extracellular fluid or cerebrospinal fluid in \ivo by means of microdialysis. Potentially, this may reflect ischemia at an early stage. Unfortunately this is beyond the scope of this dissertation.

1.7 Physiological state representations

Several authors point out the inadequacies of traditional methods for analyzing physiological data [20][41][118]. For example, Rezek et al. state th at traditional frequency and amplitude analysis are insufficient for physiological signal analysis, and consider four stochastic-complexity features including spectral entropy. They apply these to Cheyne-Stokes respiration, anesthesia and sleep analysis [118]. The

NICU system provides a few tools for this type of analysis of the electrophysiological and other physiological data collected in the NICU.

The derangement of the physiological parameters measured in a patient with a particular disease process was first analyzed rigorously by John Siegel et al. [20].

They selected seventeen physiological parameters and represented the values of these

31 V p 0 2 58.00 HR 90.00 AVDIF A STATE: 4 j B STATE: 3.1 MAP 105.00 D/A RATIO: 1.84 OB RATIO: 2.60

VpH 7.41

V p C 0 2 38.00 D STATE: 83 C STATE: 8.0

Figure 1.8: The vector of eleven physiological parameters of Siegel et al. (adapt­ ed from [20|) displayed on a circle diagram. States A to D represent four stable pathological states.

parameters on a circle diagram as showm in Figure 1.8. The parameters were nor­ malized so that the mean was at a fixed distance from the origin and the standard deviation was represented by a fixed length for all parameters. In this way a cir­ cle represented the healthy state and any deviations from the healthy state would result in a distorted circle. Siegel et al. observed that rather than resulting in arbi­ trary* distortions, the pathological states mapped to specific repeated constellations

(distortion patterns) characterizing the disorder. Furthermore, these pathological s- tates were stable (robust to perturbations in some variables). The existence of the characteristic states was quantified by means of cluster analysis [20|.

In this work, a neurophysiological state vector is developed to represent the neu­ rological state of the subject, and the existence of multiple stable neurophysiological states is examined (see Chapter 7).

32 Buchman [20] points out that the idea of multiple stable states refutes the hypoth­

esis that there is only a single healthy state for the patient and that other conditions

are “unstable” (this is the principle of homeostasis, first proposed by Walter Cannon

in 1939). Furthermore, the variables representing the single “stable healthy state”

may not be constant in value at all, nor indicate a healthy condition if they are. .-\n

example of this which has received a lot of attention is heart rate. Goldberger [42]

describes how heart rate is a fractal phenomenon i.e. the heart rate is variable and

unpredictable, but statistically self-similar over a reasonable range of scales. Fur­

thermore, if the heart rate loses its variability, becoming periodic and predictable

(and thus losing information), this indicates a state of disease. The statistics be­

come dominated by a particular scale or frequency. Some ways of quantifying the

uniformity of scaling are the Fano factor and Allan factor and detrended fiuctua-

tion analysis. A number of studies confirm the notion that heart-rate variability

decreases in the disease state - multiple organ dysfunction syndrome [38|, hepatic coma and the period before brain death [41]. Goldberger also cites examples of other

physiological parameters which become more periodic and predictable in the disease state - the electroencephalogram in coma, Jacob-Creutzfeld disease, alternans and the

Wenckebach phenomenon in the electrocardiogram, tremors in Parkinsonian patients and Cheyne-Stokes breathing (characteristic of some strokes and severe heart failure)

[411-

Godin and Buchman have proposed a mechanism for the complexity-loss which typifies the heart-rate in multiple organ dysfunction syndrome (MODS) which fre­ quently precedes death in the ICU [38j. They propose that the various healthy organs behave as biological oscillators, and that an orderly coupling between these systems

33 is maintained by the communications network of the neural, humoral and cj'tokine systems. In the coupled state the oscillators are chaotic i.e. the system is driven to a stable state which is far from equilibrium and limit cycles. When the coupling is removed, the oscillators run freely to produce ordered and predictable oscillations.

This behaviour has been demonstrated with model networks of oscillators. Godin and

Buchman suggest that the systemic inflammatory response syndrome (SIRS) which precedes MODS disrupts communication between the organ systems and eflTectively decouples the organs. MODS results. Relieving the SIRS and recovering the cou­ pling does not necessarily return the system to the original healthy state - there are several stable states (attractors). This has been observed in ICU patients recover­ ing from MODS - although the organs have returned to their healthy states and the communications have been stabilized, recovery still takes several days. In this study, various candidates for a time-varying neurophysiological state vector are evaluated.

The usefulness of the neurophysiological state vector is based on its value in predicting short-term changes and forecasting the long-term outcome of the patient.

Nakao et al. studied the dynamics of a multivariate representation of the car­ diovascular system in various states [97|. They model the system as a time-varying autoregressive (AR) process and represent the dynamics in system space. They de­ scribe the geometry of the transitions from the awake state to slow-wave sleep (SWS) and from the rapid eye movement (REM) stage of sleep to SWS. The NICU system provides the mechanism for generating state space diagrams.

34 1.8 Continuous multimodality intensive care monitoring

The earliest comprehensive monitoring system for the CNS was reported in 1979 by Ackmann et al. [2]. They used a microcomputer to collect EPs and monitor the

EEG in NICU patients. They also collected other physiological parameters. They gathered data from 50 patients and concluded that the technique showed potential for detecting adverse trends for warning purposes. This was a stand-alone unit without the advantages of networking and the fast computing technology that now exists for extracting information from the electrophysiological data. Ackmann et al. later re­ ported that the traditionally-monitored parameters are late indicators of deterioration in ICU patients, and that the clinical neurological examination has thus remained the cornerstone in monitoring. They suggest that a combination of parameters including the SEP, ICP and cardiac information would be more useful [3][4j. In 1984, Miskiel et al. designed a microcomputer-based BAEP monitor for monitoring BAEPs inthe ICU

[84]. They observed the usefulness of BAEP monitoring during surgery to prevent hearing loss and the sensitivity of the BAEP to mild hypoxia, and extended this utility to the ICU. In 1987, Miskiel et al. evaluated trending of the latencies of BAEP waves

I, III and V in ICU patients and identified the clinical potential of long-term moni­ toring of this sort [85]. Boston identifies several problems that make multimodality electrophysiological monitoring a challenge, in particular the low signal-to-noise ratio of the signals and the problem of automated waveform interpretation [19]. In 1988,

Moberg et al. pointed out the lack of appropriate technology for continuously mon­ itoring the EEG in the ICU and noted its importance, especially for trauma, stroke and SAH patients [87]. They also noted the importance of automated algorithms for detecting and predicting trends. Moberg et al. designed a prototype portable ICU

35 monitor for trending EEG parameters in the ICU. Si describes an expert system that

interprets the EEG and alerts a trained clinical neurophysiologist if irregular changes

occur [131]. Rosenblatt et al. used a multimodality system for monitoring the BAEP,

SEP and VEP post-operatively [121j.

An effort has thus been made by several researchers toward building a compre­

hensive monitoring system for the NICU. In this dissertation, a unique system is

described that combines the results of previous research and introduces new tools

and facilities. The desirable features of the new system are: it is a multimodality

monitor; it monitors continuously; it provides trending; it incorporates traditional

and non-traditional analysis tools; minimal supervision is required; and, if anomalous

results occur, they may be perused at the convenience of the clinical neurophysiologist

from his or her desk or anywhere in the hospital using the web-based interface.

Many authors suggest the utility of a unified brain monitoring approach incor­

porating multimodality data. This has been done in cardiology, where Lage et al.

combine EGG, hemodynamic data and cardiac output to manage the critical care of

patients with cardiovascular disorders [66]. Bidaut et al. describe a set of techniques

and protocols for combining MRI, MRS, CT, SPECT, PET and other information

[14]. Their method also integrates electrophysiological data. Jordan supports the

idea that monitoring a combination of parameters is necessary for optimal patient care in the NICU [57|.

Continuous monitoring is important in the NICU because changes may occur at any time. Vespa reports that continuous EEG monitoring in the ICU reduces costs and improves outcome [149]. Thakor observed changes in serially-recorded SEPs in cats during middle cerebral artery occlusion, and recovery once the occlusion was

36 eliminated. He observed similar changes in the SEP in humans during and after

anesthesia and concludes that early detection of neurological trauma may be possible

if these techniques are applied in the NICU [139|.

Finally, Prior [110| cautions that for neurophysiological monitoring in the ICU

to be as effective as possible, careful study of a lot of data together with accurate

outcome information is necessary'.

1.9 Summary

The population of stroke patients in the NICU is an important one in which continual CNS monitoring could be very beneficial. Rather than collecting data of a single modality only, the approach in this research is to combine several views of the functioning of the CNS, so that an overall assessment of the health of the organ system is made. Ideally, many more parameters should be included than what are within the scope of this dissertation. Nevertheless, a broad-ranging set of electrophysiological and other parameters are measured at regular intervals in this research, in order to give a good idea of the state of several subsystems of the CNS as they vary with time. It is hoped that by investigating the data collected in the NICU and the INR suite, the following questions may be answered: What parameters form a meaningful time-varying neurophysiological state vector for evaluating the current state of the patient and for prognostic purposes? How should the state vectors and the parameters which contribute to them be displayed to clinical personnel to optimize information transfer? What are the practical implications of trying to collect these data in the

NICU? What are the characteristic dynamics of the state vector in the healthy and pathological state; i.e. is it characterized by a single fixed point or multiple fixed

37 points representing the healthy and the pathological states, or is there an attractor or multiple attractors in the state space? Can the neurophysiological state space be meaningfully divided into regions to represent the current state of the patient? Will an extrapolation of the dynamics give any idea of the short-term changes that may be expected in the patient and the eventual prognosis of the patient?

38 CHAPTER 2

The NICU monitoring system: Design and implementation

2.1 Design overview

Figure 2.1 shows the network layout of the NICU monitoring system. The overall

software block diagram for the networked system is shown in Figure 2.2. .A.t the

patient's bedside in the NICU itself is the data acquisition workstation together with

the acquisition and stimulation hardware. The acquisition workstation is an HP

Apollo workstation running the HPUX operating system and the acquisition and

review software described in Section 2.3. The acquisition and stimulation hardware is

described in Section 2.2. This part of the system provides the capability of stimulating

the patient using any of the three modalities - auditory, somatosensory^ or visual -

and acquiring and averaging the electrical response of the central nervous system.

It also collects whatever data are selected from what is available on the Tramscope

bedside monitor, such as heart rate and intracranial pressure, along with annotations

entered by the nurse on the handheld annotations terminal. Although the acquisition workstation is capable of running as a standalone system in the NICU without any network connection, it is preferable to have it on the network during acquisition if remote access is desired. The network is also necessary if the monitor is to page the user when there is a problem or when the monitoring schedule has been completed

39 -H- -H-

HP Apollo ^ Personal ^ Remote review^ acquisition computer web platform with workstation server HTML browser

HP Apollo ^ Local acquisi­ review Hospital paging system tion/stimulation workstation hardware

Patient J N euroIC U Neurology local area network Internet (hospital intranet)

Figure 2.1: Networking diagram for the monitoring system with remote review and monitoring terminals.

successfully. Paging is achieved by sending the appropriate information to a specific

URL as an HTML request. The actual paging hardware in association with the URL is provided as a service to employees by CCF.

Elsewhere on the Department of Neurology’s local area network are one or more review workstations. These are HP workstations running the HPUX operating system

(not necessarily the same version) together with the acquisition and review software.

All of the review and analysis features described in Section 2.3 are available on these machines, but the acquisition part of the software is not functional as the hardware is absent. On those HP workstations where a magneto-optical disk drive is available, the patient data archiving software may be used. The implementation is described in

Section 2.5.

40 Acquisitioii Acquisition Common Authorized and review and review Gateway remote internet software software Interface viewing (CGI) software (standard (acquisition (review browser) machine) machine) (web-server)

local area network

Patient data Archive stored on software networked Magneto­ drive optical disk (review machine)

Figure 2.2: Overall software block diagram for the NICU monitoring system.

A personal computer with an Intel or compatible processor running the Linux operating system is used to provide the web-based interface. The free Linux software,

Apache, is used to provide the web-serving capabilities. This executes numerous com­ mon gateway interface (CGI) scripts which were developed for the NICU monitoring project. These are described in Section 2.4. The web-browser can be accessed by any computer on the hospital’s wide-area intranet network, regardless of the platform, provided that it is running standard HTML browsing software.

2.2 Data acquisition hardware design 2.2.1 Overview of hardware design

Figure 2.3 shows the hardware block diagram of the acquisition unit. The acqui­ sition computer has a general purpose information bus (GPIB) to communicate with

41 (IEEE bus) BEG headbox ^ EEC amplifier A/Dcoovertec (network)

Stimulators

6 1 / VEP M

Tram scope ^ (serial) bedside monitor Remote Panent hand-held annotations Patient and bedsde equipment (scnal) terminal

Patient cubicle in neurological ICU

Figure 2.3: Hardware block diagram of the acquisition unit.

the analogue to digital (A/D) converter and EEG amplifier. The EEC amplifier in turn communicates with the headbox which incorporates a local preamplifier. The electrodes affixed to the patient’s head are connected to the headbox. The computer has 3 serial RS232 ports to communicate with the programmable pulse generator,

Tramscope bedside monitor and the annotations terminal. The A/D converter has a built-in digital input/output (I/O) port, which the computer uses to send TTL compatible logic signals to the signal switching unit i.e. the computer puts the in­ struction on the IEEE bus and the A/D converter catches it and sets its logic levels accordingly. The switching unit receives these signals and activates the appropriate stimulator on the appropriate side of the body. All units are commercially available products, except for the signal switching unit which is original hardware designed and built by the author. Most of the cables were constructed by the author. The

42 interconnection between the EEG acquisition units (the IEEE bus interconnection) is inspired by the Vangard epilepsy monitoring system. The rest of the architecture was designed by the author. The interface in contact with the patient is in all cases commercial, unmodified, and in the form approved by the United States Food and

Drug Administration (FDA) for patient safety and legal reasons.

2.2.2 The EEG amplifier and headbox

A Nicolet SM2000 or 6R12 Biomedical Multi-channel Amplifier provides the ac­ quisition front-end to the system. The amplifier consists of an EEG headbox and the

EEG amplifier. The EEG electrodes are affixed to the patient’s head with collodion glue and using conductive gel, and the electrode leads plug into the headbox. The headbox provides preamplification as near as possible to the patient so as to minimize the length of electrode leads required and thus minimize noise pickup. The headbox also provides optical isolation for patient safety and an electrode impedance check to test that the electrodes have been adequately attached to the patient.

The EEG amplifier contains a microprocessor in combination with a number of programmable switches, filters and an extensive front-panel. The HP workstation communicates with the amplifier through the GPIB. By means of ASCII command codes, the workstation is able to set overall parameters such as 60 Hz noise filtering on all channels, as well as specific information pertaining to each channel. The montage is defined as the totality of filter band limits, headbox electrode number and sensitivity or gain associated with each of the used channels. For example, the ASCII string

“#910204030204030F0F0F0F0F0F.../” sets channels 1 and 2 to pass frequencies in

43 the band 0.5 Hz to 250 Hz with a full-scale sensitivity of 200 microvolts and turns off the 30 remaining channels.

The amplifier is able to store 7 montages. The acquisition software programs four of these at the start of a schedule to correspond to the test type 1 for the BAEP,

SEP, VEP and EEG tests, where the characteristics of test type 1 are defined by the acquisition and review software (see Section 3.2.2). During testing, this means that every time a new test begins, the amplifier need only execute a simple command such as “#810B/” to select the VEP test type 1 montage. If during acquisition a test of a type other than 1 is performed, the corresponding montage is reprogrammed just before the test begins. If, for example, montage 3 were reprogrammed because the VEP test to be performed was of type 2, then from this point on the “#810B/” command would recall the new VEP montage type 2. In this w^ay reprogramming of the EEG amplifier is optimized in a first-order sense i.e. if it is assumed that most of the time adjacent tests of the same modality are of the same test type, the amount of reprogramming of the four amplifier channels is minimized. This scheme is illustrated in Figure 2.4.

2.2.3 Analogue to digital converter

-A.n lOtech ADC488 analogue to digital (A/D) converter provides conversion of the analogue outputs of the EEG amplifier to digital form so that they can be processed by the acquisition workstation. The device provides 16 single-ended analog input channels and 16 bit resolution sampling. The minimum sampling period per channel is 10 microseconds, so that a single channel can be sampled at 100 kHz, two channels at 50 kHz each, etc. The ADC488 unit used in this acquisition system is fitted with a

44 S c h e d u le

Modality, side, type BAEPR 1 EEG I BAEP L 1 EEG I SEP L2 SEPR I Test time

Amplifier, pulse generator and A/D converter

C ontents BAEPI : A ctive? C ontents SEP 1 : SEP 2* SEP 1 A ctive? C ontents VEP 1 A ctive? C ontents EEG 1 Active?

Switching unit Left Side Right BAEP Modality SEP VËP EEG Noise Active?

Figure 2.4: Scheme used to minimize reprogramming of EEG amplifier montages, A/D amplifier sampling parameters and pulse generator timing between tests of different types and modalities within a single schedule. * denotes reprogramming of memory in the peripheral hardware - other peripheral hardware reconfigurations are simply recalls from peripheral hardware memory.

45 memory expansion module so that it can store up to 131,072 samples before having to transfer them to the acquisition workstation. Communication between the acquisition workstation and the A/D converter is by means of the IEEE 488 bus. Simple ASCII commands are used to program the sampling rate, type of buffering, method of data transfer, type of triggering and digital I/O values. For example the command “D255” asserts all digital outputs.

The standard test type 1 for the BAEP modality, as defined in Section 6.2, is the most demanding of the standard tests in terms of the use of A/D conversion resources as well as workstation resources. It will be used as an example to explain the data acquisition scheme.

The BAEP test requires two EEG channels - one for the electrical activity at the left earlobe (Al) and one for the activity at the right earlobe (A2). The electrical activity at these points is recorded relative to that at Cz, the point on the scalp at the center of the midline. The patient is stimulated by a series of auditory clicks presented at a rate of 11.76 Hz or every 85 ms. The amplifier detects electrical activity at the two electrodes within the frequency range of 50 Hz to 1500 Hz and with a full-scale amplitude of 50 //V, and provides the filtered and amplified signal to the A/D converter. The A/D converter acquires 12 samples at a rate of 50 kHz on each channel before stimulation and 500 samples at 50 kHz after stimulation. This means that every 85 ms the system acquires 0.24 ms of EEG trace before the patient hears the click and 10 ms of EEG trace after the patient has heard the click.

The A/D converter transfers data across the IEEE bus in raw binary mode (2 bytes per sample). In this mode it can transfer data faster than 200 kB/s, so that in principle it could transfer the BAEP response in real time since two channels sampled at 50

46 kHz each with a resolution of 16 bits amounts to exactly 200 kB of data per second i.e. the data can be transferred in 10.24 ms. This leaves 74.76 ms for trace averaging and artifact rejection. Some samples are required before stimulation occurs and the programmable pulse generator provides a trigger pulse to the A/D converter 0.24 ms before providing a trigger pulse to the switching unit to be sent to the appropriate stimulator. The A/D converter is used in burst mode. In this mode it waits for the external trigger produced by the pulse generator before acquiring 512 samples at 50 kHz from each of two channels. When it has obtained the 512th sample (512 samples or 10.24 ms after the start-acquisition trigger) it informs the acquisition workstation by means of an interrupt that the trace has been obtained. The acquisition machine requests transfer of the data and obtains the 512 trace samples over the next 10.24 ms or less. It then processes the trace as explained in Section 4.2 (averaging and artifact rejection). It must complete processing within 74.76 ms. The HP Apollo workstation is a relatively slow machine by today’s standards. While it is acquiring data, the acquisition process is given real-time priority by the HPUX operating system. Despite this, if the machine is locally displaying the acquisition results graphically and at the same time acquiring a BAEP with trace rejection and physiological data from the

Tramscope, it will occasionally miss a trace, so that it is not uncommon for the system to acquire 98% instead of 100% of the 1500 traces that were intended. It is rare for more than 5% of BAEP traces to be lost in this way, and the BAEP is the most resource-intensive of the tests which the monitoring system performs. The averaging routine only includes successfully acquired traces in the average.

The A/D converter has a digital I/O port which is used to configure the switching hardware described in Section 2.2.5.

47 All communication protocols at a higher level than the basic IEEE 488 commands such as “address to listen” are implemented by the author.

2.2.4 Synchronization and the programmable pulse generator

A Model 400A Pulse Generator from Berkeley Nucleonics provides the accurate timing needed by the system to synchronize stimulation and acquisition. This pulse generator was chosen because it can be programmed remotely by means of commands on its serial port, it provides very accurate timing in the required range for this application and it provides pulses on two separate but synchronized channels.

The HP Apollo workstation has three RS232 serial ports, committed to com­ municating with the pulse generator, Tramscope bedside monitor and annotations terminal. It communicates with the pulse generator by means of simple ASCII serial commands such as ‘Tl:1500” to set the pulse width of the pulse on pulse output channel 1 to 1500 /j.s.

The pulse generator is used in burst mode to provide a train of pulses on each of two channels, one to trigger the A/D converter to start acquisition of a single trace, and one to trigger the stimulator to stimulate the patient. To use the example of the

BAEP once again, the timing diagram is shown in Figure 2.5. It can be seen that on channel 1 the pulse generator produces 1500 pulses of duration 100 microseconds each and separated by 85 ms. These pulses are fed to the external trigger of the

A/D converter, which also operates in burst mode. Everj' time it receives an external trigger pulse, the A/D converter captures 512 samples of the amplified EEG signal over a period of 10.24 ms on each EEG channel. Channel 2 of the pulse generator provides the same train of pulses, but delayed by 0.24 ms. These pulses are fed to

48 For each stimulus:

■<----- 100 u s

10.24 m s (512 sam ples)

For com plete te s t

127 s (1500 stim. c ydes)

Figure 2.5: Timing diagram for programmable pulse generator channels 1 and 2 during acquisition of a standard test type 1 BAEP.

the switching hardware which channels them to the appropriate stimulator according to the selection specified by the acquisition workstation through the digital I/O port of the A/D converter. For the BAEP this means that 0.24 ms after acquisition of a trace begins, the patient hears the 100 microsecond click, and acquisition continues for a further 10 ms after stimulation.

2.2.5 Auditory, somatosensory and visual stimulators and switch­ ing hardware

Figure 2.6 shows the interconnection diagram of the modules making up the ac­ quisition system, including the switching unit, and Figure 2.7 shows the internal schematic of the switching unit itself. The switching unit routes the timing puls­ es and the stimulation pulses appropriately as dictated by its digital inputs. These are provided by the digital I/O port of the A/D converter which is programmed by

49 IEEE WORKSTATION ADC EEG Amplifier F- D IG OUT %

XI Power Programmable Supply Pulse Generator 12

POWER DIG IN PULSE IN

Switching Unit EEG electrodes BAEP SEP ASU IN ASU OUT X . n _ inserts q q L ED / NOISE/ TRIGGER goggles O O CLICK A dhesive Stimulus C lick-Tone electrodes isol. unit Generator

Figure 2.6: Interconnection diagram of acquisition hardware modules including the switching unit.

50 > left — 300E transducers + 12V + I2V >■ right

HE3321 47k HE3321

- 4 noise CA3262 act right 4 click noise/click ^ >■ trigger CA3252

47k

activate BAEP clicks

74LS08

+5V

digital I/O 47k

>- left somatosensory 74LS08 L.VI324 activate stimulus right SEP 12k isolation unit >- right 74LS08 activate LM324 left VEP ► -► left 74LS08 L.M324 LED activate ^ goggles right VEP right 74LS08 L.V1324 pulse stimulus generator pulse

Figure 2.7: Internal schematic of the switching hardware.

51 the acquisition workstation. The switching unit is powered by a supply of 4-5V and

+/-12V .

Auditory stimulation

The auditory stimulation section of the switching unit makes up the top half of

Figure 2.7. Activating the left or right noise/click by asserting the appropriate line of the A/D converter’s digital I/O port causes the left or right HE3321 reed relay to be activated. To provide sufficient current to activate the relay coils, the digital inputs are buffered by means of CA3262 high current operational amplifiers. The reed relays route the noise or clicks appropriately for each channel from the Grass auditory click tone generator to the switching unit’s 5 mm stereo jack socket into which the 300Q

Nicolet auditory transducers are plugged (through a splitter that converts the single stereo jack socket to two mono jack sockets). Plastic tubes carry the sound from the transducers to the patient’s ears. In this study, the tubes were chosen such that the sound takes 1 ms to traverse their length.

A Grass Click-Tone Control Module Model SlOCTCM provides the wideband masking noise and the click tones to the switching unit. .A. click is provided by the click tone generator if the trigger pulse from the pulse generator is routed to the trigger of the click tone generator. This routing is done by an AND gate (74LS08) operating on the trigger pulse and the “activate BAEP clicks” line of the A/D converter’s digital

I/O port. The noise and click amplitudes are selected on the front panel of the click- tone generator. Typically a 30 dB masking noise signal and a 90 dB rarefaction-type click are chosen. A pulse width of 100 fj,s for the click is also chosen from the front panel.

52 Somatosensory stimulation

The trigger pulse from the pulse generator is combined by an AND operation with the left or right “activate SEP” line of the A/D converter’s digital I/O port.

The resulting signals are buffered by means of LM324 operational amplifiers and the outputs are current-limited by series resistors. These signals activate the left and right channels of the Axon stimulus isolation units (SIUs). The SIUs ensure that the patient is electrically isolated from the main circuit of the monitoring system.

Disposable adhesive stimulating electrodes deliver the stimulus to the patient.

Visual stimulation

LED goggles are used to provide visual stimulation to the patient. Each LED is driven by the current-limited (12 mA) output of a buffering operational amplifier, and the logic is the same as for somatosensoiy stimulation.

2.2.6 The Tramscope bedside monitor

Physiological monitoring in the NICU at CCF is done by Marquette Monitoring equipment. The monitors are Tramscope or Eagle monitors, which are equivalent in terms of asynchronous communications. The Tramscope monitor has an RS422 serial port, and this is converted to RS232 by an Asynchronous 232 to 422/530 con­ verter from Black Box Corporation so that the acquisition workstation can commu­ nicate serially with the Tramscope. Structured packets are used to communicate commands and data between the computer and the Tramscope. These structures are defined in C header files provided by Marquette and used in the acquisition software.

Typically the acquisition machine sends a “READ” function code and a “PARAM-

ETER REQUEST” subfunction code in the header of a packet to the monitor at

53 intervals specified by the user. The monitor responds to each request with a response packet containing an “SBEDSIDE_MSG_DEF” structure, an “SBEDSIDE_FLOAT” structure and a number of “SPAR_FLOAT” structures. This variable length struc­ ture must be captured and confirmed by the acquisition machine and then parsed to extract the physiological parameter values. Acquisition software written by the au­ thor obtains the data, parses it and masks out the unwanted parameters to leave only those parameters that are required for the particular physiological test type. This is done at the rate specified in the physiological test type definition. The physiological data acquisition process is a real-time process which provides its own timing. .A.s such, physiological parameters cannot be acquired at a high rate. A sampling period of 10 seconds is reasonable.

2.2.7 The annotations terminal

For the purposes of entering annotations at the bedside, a ProTerm 45R2-1 .\SCII terminal from Two Technologies is used. Communications are by RS232 serial. The annotations terminal is used in the mode where its screen displays 32x16 characters.

Simple escape sequences from the acquisition workstation are used to issue the com­ mands to display the appropriate menus on the screen. For example the sequence

ESC[PnA moves the cursor up one position. The keyboard echoes ASCII characters back to the acquisition computer, which interprets the keystrokes appropriately.

2.2.8 Electrodes and setup for tests

Cold-plated reusable miniature disk electrodes with conductive electrode jelly and collodion glue are used to record the EEC from the scalp. The setup for the studies is described in Section 6.5 as part of the protocol. A kit of required paraphernalia

54 accompanies the system. Surgical tape, alcohol prep swabs, gauze, EEG electrode gender adapters, Omni prep skin preparing paste, a measuring tape and skin marking pen, an air flow meter, extension tubes for the ear inserts, ER3-14A ear tips, network and serial cables and a network terminator are included in the kit.

2.3 Acquisition and review software design 2.3.1 Overview

Figure 2.8 shows the block diagram for the acquisition and review software. Each block in this diagram represents a module - either a Tcl/Tk file or a single executable compiled from one or more C source files. The subsystems are called appropriate­ ly by the main program (NICUsystem) at the users request from the main menu.

NICUsystem is written in Tcl/Tk.

2.3.2 Acquisition subsystem

The acquisition subsystem consists of five Tcl/Tk files together with its precom­ piled C routines (rtroutines - real time routines) (precompiled only for the HPUX-9 platform) and a single executable compiled from a number of separate C source files.

2.3.3 Display subsystem

The display subsystem consists of five Tcl/Tk files together with its precompiled

C routines (dxroutines - display routines) (precompiled for HPUX -9 and Linux (Red

Hat 5.2) platforms and implementing little-endian to big-endian conversion on the

Intel version).

0 0 Overall block diagram

Display/dxroudneV

DeiivedParametcrs.td dxroutines (O

Display/ DisplayRawTraces.td Display Airay.id

DisplaySdicdulc.td Display Défini tions.id DisplayStalistics.td

NICUsystem (Id) r

AcqDisplay.td CreateStateV ector.td

acquire (C) createvector (C)

Annotatetd page.td

Acquisitioa/rtroutiGes/

Figure 2.8: Block diagram for the acquisition and review software.

56 2.3.4 Analysis subsystem

The analysis subsystem consists of three sections - the (raw) analysis section, the vector analysis section and the final analysis section. The interfaces are written in

Tcl/Tk and the actual analysis is done by an executable implemented in C or C4-4-.

The mathematical details are given in Chapters 4 and 5.

Raw data analysis

Raw analysis encompasses finding peak amplitudes and latencies for EPs, calcu­ lating the spectral properties of the EEG and separating out physiological values.

Vector analysis

Vector analysis encompasses smoothing and interpolating irregularly spaced data for a data set that is easier to represent graphically and examine as a continuous

(regularly sampled) vector.

Final analysis

Final analysis encompasses extrapolating the state vector, calculating the vari­ ability of parameters within the state vector, calculating the dimensionality of com­ ponents of the state vector and generating state space diagrams. A separate C exe­ cutable performs each type of analysis. These executables are called by the Tcl/Tk script “FinaLAnalysis.tcl”. They are not mentioned in Figure 2.8.

2.4 Web-based interface software design

Nenov identifies the significant demand by physicians for remote access to medical monitoring results, specifically NICU monitoring results [99][100], and points out

57 that world-wide web (WWW) technology and client/server applications now make

it possible to respond to that need. Lee et al. [68] present such a real-time patient

monitoring system using the W^VW. Smith notes further that the tools of information

systems now make it possible to automatically gather data, seamlessly integrate it

amd make it transparently available [133]. Murchison also emphasizes the utility

of information systems in uniting bedside monitors with other equipment and the

hospital information system for remote access and improved patient care [92j.

Figure 3.17 shows the layout of the web-based interface of the NICU monitoring

system from the point of view of the user. Each page in this figure is labeled with

the name of the HTML file or CGI (common gateway interface) script that generates

it. Related pages are generated by the same script. For example, the electrotrace.cgi

script generates the HTML corresponding to the selection and display of single and

multiple raw traces of electrophysiological data. The directory structure for the web-

based interface implementation is illustrated in Figure 2.9. The layout of the site from

the user’s point of view and details of how to navigate the site are given in Section

3.5.

The main menu and protocol are simple HTML pages. All the other pages are generated at the moment of access by CGI scripts written in Tcl/Tk. As shown in

Figure 2.9, the CGI scripts are all included in the subdirectory MEPstudy. The cgi.tcl file contains routines for writing CGI scripts in Tel and was written entirely by Don

Libes [71[. These routines take simple arguments and generate appropriate HTML.

The cgi-local.tcl file contains definitions required by cgi.tcl, but specific to the site.

This includes the paths to the scripts, common headers and footers, and the email address of the administrator (the author). The administrator is emailed with the

58 pub[ic_btml

index.html

html

footer

icons

ccf-logo.sif

jump—arrow.g if

fonts dxroutines

trebucd.rtf I ihdxrou tines.so

MEPstudy gdtclft

cgLtcl: cgi'-locaLtci

cgi-mep.icl: mep-display.tcl

protocol.html

datechoice.cgi: displayformdata.cgi: electrochoice.cgi: electrotrace.cgi: protochoice.cgi: review.cgi: submitrecord.cgi: vectorchoice.cgi: vectortrace.cgi

Figure 2.9: Directory structure of the web-based interface.

59 Tcl/Tk error message if the system should encounter an error while it is executing one of the CGI scripts, cgi-local.tcl is supplied by Don Libes and edited by the author.

All other CGI scripts were created by the author.

The CGI scripts have two special capabilities. They read data that the NICU monitoring system has created, and they create graphical interchange format (GIF) images in response to requests from the browser. These capabilities are provided by routines contained in two statically hnked libraries compiled from C.

The first statically linked library is libdxroutines.so which provides functions for reading and formatting the data acquired and stored by the NICU monitoring sys­ tem. This is exactly the same source code as described in Section 2.3.3, where it is used by the review station software for retrieving data. The only difference is the compilation - when the library is compiled for the review station, the executable is for the HPUX-9 platform (Motorola 68000 processor), whereas for the web-server it is compiled for a Linux (Red Hat 5.2) platform (Intel 586 processor). This gives rise to a further subtlety - the Intel integer representation is little-endian whereas the

Motorola representation is big-endian. compiler directive sets a switch in the code to include a short section of code that does the conversion if necessary.

The second statically linked library is Gdtclft.so which provides functions for cre­ ating GIF images. This code was written entirely by Thomas Boutell and is available at http://wwrw.boutell.com/gd. The routines provide low-level drawing functions for lines and polygons as well as the capability of including text strings in the image. The font chosen is Trebuchet MS, a font designed by Vincent Connare, available as part of the free core fonts for the web provided by Microsoft Corporation. A set of general but higher-level graphics generating routines was written by the author, and these

60 are included in the file mep-display.tcl (multimodality evoked potential display). The routines provide such functions as creating a ruler marked in seconds, justifying text and drawing traces from lists of data.

All web-pages have the same basic layout. The headers and footers are included by routines in the cgi.tcl script. The left-hand column of links is generated dynamically by a routine in the file cgi-mep.tcl, written by the author. This file also includes other routines related to page formatting, such as menu layout, whitespace for page layout and link generation.

The web-server software chosen for the web-based interface is Apache. This is a free web-server available for the Linux operating system. The server is started by typ­ ing /usr/local/sbin/httpd -d /net/home/vandera/apache on the UNIX command line on svr2.eeg.ccf.org. The httpd.conf file in 'vandera/apache defines the port (6500), user name and group (vandera and usrs for security - the server is not run as root) and the administrator's email address for posting error messages if necessary.

Some of the CGI scripts generate several different pages depending on the data passed to them by the browser. Data can be passed as part of the URL or invisibly as a separate post operation. In both cases the CGI scripts accept the variable values and determine the appropriate response. For example, if electrotrace.cgi receives a patient name and modality as part of the URL, it returns the raw selection page for the patient - this page includes a list of dates and times when tests of that modality were done for the particular patient - and the appropriate mechanism to repost the patient name and modality to electrotrace.cgi, along with the chosen date and time.

If a patient name, modality and a single date are posted to electrotrace.cgi, it returns a page containing a reference to a GIF image which is created at the moment the

61 request is made and contains a graphical rendition of the appropriate trace along

with labeling, rulers etc. If multiple dates are posted, electrotrace.cgi generates a

list of channels appropriate to the specified modality. A selection can be made, and

the repost mechanism posts all the data back to electrotrace.cgi together with the

chosen channel. If multiple dates and a channel selection are posted,electrotrace.cgi

generates a graphical rendition of the appropriate stacked display.

If only the patient name and modality are posted, posting is done as part of

the URL so that it is clear to the user what is happening. When one or more

dates and times must be included, the list becomes excessively long for the URL

and is posted instead as a separate POST operation to the server. The scheme

is illustrated diagrammatically in Figure 2.10. The other CGI scripts use similar

schemes to respond to browser requests.

Every time a GIF image is created, it is given a unique name to prevent the

browser from cacheing it. The name incorporates the time of the request (UTC) and

it is assumed that two requests will not be separated by less than a second. Since

the process of generating and transmitting the image lasts longer than a second this

is a reasonable assumption. The interface has not been designed for the situation in

which there are several simultaneous users.

2.5 Archive software design

The patient data archiving software consists of a single Tcl/Tk program called

"NICUArchive". This program was not written as root and is not executed as root on the HPUX platform on which it runs. Since only root has permission to mount mag­ netooptical (MO) disks, the archive software is unable to do this. Some other software

62 fiom eitxtrochoke.cyii Raw Data (Trace) Display

URL=dcctroirace.cgi'’panenm«*ne=# &moda!i^=#

Raw Data Select Time uRLsdecgotrace.cgi

POST panent name, .‘iLBTvuemtMa : modality, angle dace and time

Raw Data Channel Select Raw Data (Array) Display URL=dectro trace cgi

POST panent name electrotrace.cgi modaliQr.mulnple dattsand nmes URL=dectro tracent ► POST panoitname modality, mulnple daces and nmes. channel

electrotracecgi electro trace.cgi

Figure 2.10: Posting scheme for the electrotrace CGI script, enabling it to generate one of four responses appropriate to the browser’s request.

63 such as the Vangard system archive services must execute the mount command before the NICU archive software will recognize the MO.

Section 2.6 explains the structure of the individual patient data files. The directory structure for patient data is very simple. The “PatientData” directory contains a separate subdirectory for each patient. Within the patient subdirectories the structure is flat, that is there are no further subdirectories - all data files are together at the same level within the patient subdirectory. This makes it easy for the archive software to archive or restore patient data - the software simply creates the patient directory on the target medium and copies over all the files within the directory.

2.6 Patient data files and formats

Patient data are stored in a director}^ with the name of the patient, in the Pa­ tientData subdirectory of the NICUsystem directory. Below this level the structure is flat i.e. all data files are stored at the same level, and they are distinguished from one another by file name. The naming scheme is outlined in Table 2.1. All but one of the file types is written in ASCII text format. This was done to facilitate system development, portability of the data and editing of the data. In the case of the raw acquired data, however, it is impractical to use a text format because of the large volume of data, and these files are written in binary format. Figure 2.11 shows the file formats for the raw data and specification files and Figure 2.12 shows the file formats for the stats and specifications files.

The pending schedule files consist of a list of tests to be performed. Each line of the file corresponds to a test. For example, the line “210 BAEP left 1” in a pending.rel schedule file signifies that a left BAEP of type 1 is to be performed 210 seconds after

64 File name Contents (non time-stamped data) ErrorReport.txt Contains the error message that caused system termination (and possibly triggered a page). patientname.pvohlems.txt Description of logistical and other problems en­ countered during monitoring; entered by web- interface. patientname.H.istory Patient histor}^ entered using web-interface. patientname.imagename.pov POV-ray script describing state space diagram. File name Contents (schedule-related, time-stamped data) patientname.sched.past List of completed tests, and number of intended cycles, cycles actually acquired and cycles includ­ ed in average. pafieniname.sched.pending.abs List of tests in monitoring schedule. Times are absolute UTC. paizeniname.sched.pending.rel List of tests in monitoring schedule. Times are relative to start (seconds). File name Contents (monitoring results, time-stamped) patientname.annot Annotations entered on the annotations keyboard during monitoring. patientname.modality.traces Raw traces data. This is the only file stored in binary format rather than plain text format. patientname. modality.traces Specification file for test types. Defines acquisi­ .specs tion and stimulation parameters and montage. patientname.modality.stats Values of parameters derived from raw traces da­ ta. patientname.modality.stats.specs List of names of parameters derived from raw traces data. patientname.parWst.parlistname Parameter list values. patientname.parWst.parlistname Specification file for parameter list defining how .specs it is obtained from the stats files. patientname.vector .vectomame Vector values.

Table 2.1: Summary of patient data file types used by the NICU system. Patient- name, parlistname, vectomame and imagename can be any string of text characters containing no periods or spaces. Modality is one of BAEP, SEP, VEP, EEG, Phys and Spectra: for stats files it may also be one of CT, Flowchart and TCD (web-interface data).

65 patientname.modality.spec

nochannels sampleraie nocycles puisedur notchfilter rejectfilter rejectlevel I I masknoise nosamples interstimper samtostim refelect thresh type rejectstart

i i i 1 U i i i i i il 2 Yes 50000 512 1500 85000 100 240 Off STD Trace-reject Dynamic 20 62 Type I 23 A l l8Cz50 1500 50 definition 24 A2 18 Cz50 1500 50

Type 2 T definition 1_

etc. patientname.modality.tTaces

test total trace Test time (utc) side type length header

Test sample 0. nosamples data channel 0 nochannels

cycle 0 nocvcles

Figure 2.11: File formats for raw data files and their specifications.

6 6 paâentname.modality.siats.spec

parameter name unit ; ; EC G - Heart rate (/min) EC G - PVC count (/min)

SP02 - Per. pul. rate (/min) ■

patienIname.modality.siats

▼ Ÿ 913097347 73 2 70 913097947 71 I 68

UTC

Figure 2.12: File formats for stats files and their specifications.

the system is started. The past schedule file lists completed tests and some statistics related to how the test progressed. For example, the line "913066034 BAEP right 1 okay 1443, 1469, 1500 (ave. 96%, acq. 97%)” in a past schedule file indicates that at the time 913066034 (UTC), a right BAEP of type one was performed, that it completed okay, that 1443 single responses to stimulation were included in the trace average, 1469 responses were acquired and that these comprised 96% and 97% of the intended 1500 responses respectively.

Parameter lists and vectors are stored in text files, where each line has significance as follows: The first line gives the name of the parameter and the unit is included in brackets at the end of the line, for example “BAEP Ipsilateral I latency (ms)”. The second line gives the number of available samples n of this parameter. The next n

67 lines list the times (UTC) and parameter values. This is repeated until all parameters in the list are completely qualified.

Annot files follow the same format as stats files, but there is no spec file, and the lines consist of only the time (UTC) and annotation, for example ‘*913066711

Phénobarbital increase”. Problems.txt files are free text files with the standard footer,

'‘’{patientname / patientname.prohlems.txty'.

2.7 Summary

This chapter introduces the NICU monitoring system and describes its implemen­ tation. Details of how to operate the system are given in Chapter 3. Implementing this design required several hardware modules to be integrated. The software was also designed as a set of modules that were integrated on HP and Intel platforms running different operating systems and versions of the operating systems. Various communications protocols were implemented to allow the software and hardware to interact. .A. scheme for storing, retrieving and archiving the raw and processed data was implemented. Algorithms for processing the data were developed. local GUI interface and web-based interface for reviewing the results was designed and imple­ mented. This work was presented in part at a poster session of the Edward F. Hayes

Graduate Research Forum at the Ohio State University in 1998 and the .-Annual Gon- ference of the American Clinical Neurophysiolog}* Society in New Orleans in October

1998 [144].

68 CHAPTER 3

The NICU monitoring system: Description of operation

3.1 Introduction

The system developed to monitor the parameters described in Chapter 1 is unique not in the details of the individual tests it performs, but in the combination and scheduling of tests which it makes possible. This chapter describes how the system is used to set up and perform a schedule of tests and analyze the results. Further details on the methods of analysis and their application in particular studies are included in subsequent chapters.

When the system is started, the menu shown in Figure 3.1 is displayed. This menu remains active throughout execution of the software. The main menu is laid out to illustrate the flow of data and control through the system. The first step is to select a patient name using the "Select new patient” button. A pull-down menu allows the user to choose either a completely new patient, or a patient for whom data collection has previously been set up. The next step is to set up a schedule of tests for that patient. A test schedule is a chronological list of tests and the times at which they must be performed, such as: BAEP at 10:30 on Monday, March 18,

1998; SEP at 10:45 etc. The graphical scheduler is described in detail in Section

3.2.1. Associated with each scheduled test and its modality is a test type, such as

6 9 Fattentnamer Unda-Zh

Select new patient

Eitt test sciieduie B it test delMtians

Uspiajr CBfcaiats derived parameters

Ho 1st / vector cuRsntly selected

Select new Sst / vector

parameter ro<

Create state vector

D tsptav ü s l / v e c to r

Run further analysis

Bdt

Figure 3.1: Main menu window of NICU monitoring system.

70 SEP type 1. The type number refers to a set of preset properties for the BAEP, such as the stimulation rate, filter settings, montage and type of rejection. The interface for setting test definitions is activated by pressing the "Edit test definitions” button and is described in detail in Section 3.2.2. Once the test schedule and test definitions are finalized, data collection is started by pressing the "Run acquisition” button. A display of the live data during acquisition is provided, and has several features which are described further in Section 3.3.1. Once some or all of the data has been collected, the raw results from any previous test may be displayed by pressing the “Display raw test results” button. Also, certain parameters which are automatically determined from the raw results may be calculated by pressing “Calculate derived parameters”.

This calculates, for example, the positions of waves I, III and V of the BAEP, their amplitudes and some other details. These values may be illustrated on the raw trace display. Details of the “Raw trace display” and “Calculate derived parameters” operations are included in Sections 3.4.1 and 3.4.2, respectively. The purpose of the derived parameters is to facilitate interpretation of the electrophysiological results by medical personnel untrained in clinical neurophysiology. Nursing personnel may be hindered by many other duties and cannot learn the details of interpreting raw EEC or EP waveforms. Clinical neurophysiologists are seldom on hand in the NICU to read the data. A monitor that automatically extracts the salient information from the raw waveforms and displays this clearly would be much more readily accepted by these personnel. This system was designed to extract the important information, but also store the raw waveforms so that the automatic analysis can be reviewed and possibly corrected by a qualified physician later if necessary.

71 The second block of functions in the main menu is concerned with data analysis.

A state-space approach has been taken, based on recent interest in the literature in

modeling physiological state by means of a state vector [20| and the author’s interest

in extending this model to a time-varying state-vector approach. Data analysis is

a three-step process. The first step is to analyze the raw data to obtain derived

parameters such as peak latencies. This is described in Section 4. The next step is to

take the resulting lists of parameters and construct regularly sampled vectors from

them. Since there are so many parameters which could potentially be included in

the vector, the system gives the user the freedom to choose which to include. Every

choice of state vector is given a name. For example, a state vector containing the

parameters heart rate, blood pressure and BAEP latencies could be called “cardiac-

brainstem”. This is constructed by pressing the "Select new list/vector” button on the

main menu and choosing the name. Once a new name is chosen the list of parameters

is selected by pressing the "Compile parameter list” button. This feature is described in Section 3.4.3. A list is an irregularly sampled list of parameter values. By pressing the "Create state vector” button, the list is transformed into a regularly sampled set of vector values by interpolation. The mathematical details of this operation are given in Section 5.3. The final results (list and superimposed vector) are displayed by clicking the "Display list/ vector” button.

Further analysis of the state vector is done by pressing the "Run further analysis” button. This is described in Section 3.4.5. The mathematical details are given in

Chapter 5.

72 : TWMmylznmaoi

Figure 3.2: Schedule editing window of XICU monitoring system.

3.2 Setup 3.2.1 Scheduling tests

The system provides a graphical interface whereby the user can design a program or schedule of tests to be performed on the patient. Figure 3.2 shows the layout of the window used to edit the schedule. Xine horizontal slots are labelled with a test modality on the y-axis and the time (absolute and relative) on the x-axis. Each scheduled test is represented by a solid bar in the appropriate slot. The length of the bar corresponds to the duration of the test. The test modalities are:

BAEP: brainstem auditory evoked potential,

73 SEP: somatosensory evoked potential,

VEP: visual evoked potential,

EEG: electroencephalogram,

Phys.: other physiological parameters.

There are two slots for each of the evoked potential tests, to represent stimulation on the left and right sides. The two slots of the EEG test denote the time and frequency domain representations of the test results. An EEG test cannot be scheduled in the frequency domain. When the raw results are analyzed and the spectra are calculated, the frequency domain slot of the EEG test is used to hold the bars representing the spectral estimates, which may be called up for display as described in Section 3.4.1.

The x-axis is labelled in absolute and relative time. The bar representing a test is filled in one of two colors, depending on whether the test is scheduled to be per­ formed at an absolute or a relative time. Absolute times are labelled using the format hours : minut es zseconds past midnight on the date selected. The date for the test sched­ ule is displayed in the top-left corner of the schedule window. If the test schedule extends over more than one day, this indicates the first day of the schedule. Relative times are labelled using the format hoursrminutesrseconds past the time at which the test was started (defined as three seconds after the time that the start button was pressed to begin acquisition, as described in Section 3.3.1).

The vertical red line is a cursor for editing the schedule. It is moved by clicking on it with the left mouse button and dragging it while the button is pressed. Pressing the right mouse button while the mouse pointer is in the range of the schedule display causes the cursor to jump to the start of the nearest scheduled test. The cursor

74 represents the start of a test, and it must be aligned exactly with the left side of a test bar to edit the test. The absolute and relative times of the cursor are displayed on the left below the schedule display.

At the position of the cursor, tests may be added, removed or modified by pressing the appropriate button below the schedule display. A series of a particular type of test or a program of tests may also be added. The details of a test which are represented in the schedule itself are the following:

Modality: BAEP, SEP, VEP, EEG or Phys;

Side: left, right or both (this detail is applicable only to evoked potential tests);

Test type: test definition number;

Time: absolute or relative; date and time (hours:minutes:seconds).

Further details, such as the sampling parameters and montage are described in the test definition. Each individual test in a schedule may have a different test definition number (test type), provided that the test has been defined as described in Section

3.2.2. The test number is unique to the modality, i.e. there are separate definitions for test type 1 of the BAEP and test type 1 of the SEP. If a test is scheduled for which the definition does not exist, the number in the bar on the schedule appears in parentheses. It is assumed that all test types which appear in the schedule will be defined before acquisition begins. If a series of tests is added to a schedule, the test modality, side, test type and time are again required, along with the repetition interval and number of repeats. Attempting to acquire patient data using a schedule for which some test types have not been defined will result in an acquisition error.

75 The “Add event program” button adds a preprogrammed program of tests to the

schedule. The test modalities and number of hours must be specified. For example,

selecting the modalities BAEP, SEP and EEG and a duration of 3 hours will generate

a schedule of interleaved BAEP and SEP tests on the left and right sides lasting 3

hours with a test ever}' 15 minutes, and one minute samples of EEG inbetween. This

is the quickest way to generate a standard schedule of tests of the type described in

Chapter 6.3.

The schedule must be saved upon exiting by pressing “update and exit”. .A. facility

for printing is provided - the part of the schedule shown in the entire scrollable area

of the schedule display (including that which is not currently displayed) is printed

across two printed pages if the “Print schedule” button is pressed.

3.2.2 Editing test definitions

The test schedule defines the sequence of tests and the modality and side of each

test. Further details of each test are defined in the test definitions. For each modality there can be several test definitions. There are test definitions for electrophysiological modalities and physiological tests. Table 3.1 lists the parameters which are defined in the test definition for each electrophysiological modality. The test definition edit­ ing window for electrophysiological tests which reflects these parameters is shown in Figure 3.3. The physiological parameter test definition consists simply of a list of parameters that must be recorded, along with the sampling rate and number of samples. The full list of possible physiological parameters is given in Table 3.2 and these are reflected in the test definition editing window for physiological tests, shown in Figure 3.4. The subset of available physiological parameters is dictated by the

76 Param eter Range B.AEP VEP SEPEEG Masking noise off/on y - -- Number of channels 1-8 y y y y Sample rate ls/s-50ks/s y y y y Number of samples 2-2048 y y y y Number of cycles > = 1 y y y y Interstim. period 1-10* us y y y y Stim. pulse durât. 1-10* us y y y - -\rtifact rejection none/TR/PC/PZ y y y - Threshold type static/dynamic y y y - Reject level > = 0 y y y - Discarded init. samples >=0 y y y - Notch filter off/on y y y y Reference electrode STD/AVE/LE y y y y Active (number) 1-38 y y y y Active (name) Any label y y y y Reference (number) 1-38 y y y y Reference (name) Any label y y y y Low freq. cutoff O.lHz-lOkHz y y y y High freq. cutoff lHz-50kHz y y y y Sensitivity 50uV-lV y y y y

Table 3.1: Parameters and their valid ranges defined in the test definition for an evoked potential or electroencephalogram test. The last block, which contains the montage definition, is repeated as many times as there are channels.

77 Parameter group Param eter ECG Heart rate PVC count Respiration Inferior ST Lateral ST Anterior ST Blood press. 1 -4 Vlean Systolic Diastolic Blood press. B1 - B4 Vlean Systolic Diastolic SP02 Sat. 02 press. Per. pul. rate Temperature Temperature 1 Temperature 2 Temp, change Temperature 5 -8 Temperature 1 Temperature 2 Temp, change Cardiac output Blood temp. Cardiac output IT Non-invas. BP Vlean Systolic Diastolic C 02 Expired C02 Inspired C02 Resp. rate SV02 SV02 Ventilation Resp. rate PEEP Vlean volume Gas 1- 3 Type Expired Inspired

Table 3.2: Parameters available in principal for inclusion in the physiological test definition. These parameters are sampled at the sample rate, which is also part of the definition.

78 'M (tôôôo 100 pîT— 'M « “ 0 |»3aaP°° |ïîô~

jriKt-nlfCt ^ nmoMinK tPyo^ ^ n#Kllm,0l (nrtr. V«0; pô W lôT ISID- - Zn *1 MOmO’umùm} te»f(immy f)BC.o»Mr) nr.O"omy uwavo^entarr Hÿiftoo^cMBfr SMKBr I » ~ | ë » ~ ^ al» - ^ r 41^ M -^(5- ■*S»~ ai" M 4|pôôo a i” a P5~ pc ±lp ±|pooo a i” a *r ai a =IF= ' èt a l a M 'Æ ’Sr a ■±r a r ¥ a Ll

Figure 3.3: Test definition editing window for electrophysiological tests.

configuration of the TramScope and its accompanying sensors at the bedside, from

which the NICU system samples the physiological data. The default values for the

test definitions, chosen to match the standard values for the clinical tests in the ac­

companying experimental research, are listed in Figures 6.1 to 6.4 and Tables 6.1 to

6.3.

A utility called “tramtest” is included that can be run from the UNIX command

prompt when the system is connected to the active Tramscope in the NICU. This

provides basic interaction with the Tramscope and can be used to obtain a list of

the parameters which are available from it. This is useful, for example, when blood

pressure is being measured on one or more channels by internal catheters and the channels which are used are not obvious from a simple inspection of the setup.

79 Testparaowtes; Motfalty: jPhy» TVPK r Ttofcigpiwnataii: Sam ptarats: |0.1 Hunter of samples: l « Recorded panroetenr

ECG Btaad|ims.B1 T em peratm S co z ; m Heart rata J Mb» J Temperature 1 J Expired COZ ! j IVC count j Systole J TemperahanZ _I Inspired COZ ST j Diaxloic J Temp, change _l Resp. rate j j feitetarST aood prass. BZ Temperature E SVtJZ i j Lateral ST _] M on J Temperatml J SVOZ ! j Anterior ST J Systole J TemperadureZ ventlatkm \ Respirattan u Diastole J Temp.chaaige J Itesp.rats J j Resp. rate Btood piBSs. B3 TemperatimT J PEEP j _i Apnea _l M e» J Temperature 1 J Me»voiisne | ShxMlpress.1 J sy sto le J Temperature Z Gas j _r Mean J Diastole J Temp, change J Type j _i Systofcr Btood press. B4 TamperatisaS J Expired u Dtastoec J M e» J Temperatiml J Inspired Btoodpress.2 Systole _l Temperature Z C a s t i _i Moan J Diastole J Temp, change J Type u systoBc SP02 cartlae output J Expired 1 _f OtastoSc ■ S at.O Z p n ss. _I Bkndtenip. J kispired Blood|r«s4S.3 _l Per. puL rate J cartlae output G asZ j _i Mean Teirperabra _l IT J Type ; j Systoflc J Temperature 1 Hm-lnvas. BP J BqXred _! Dfastolc _l Temperature 2 K M e» J Inspired ! Stood press.4 J Temp, change ■ Systole j Mean K Diastole j Systote _t Diastole

fceept record | Exit without ifdatSig [ Update and exit |

Figure 3.4: Test definition editing window for physiological tests.

80 For electrophysiological tests the parameters are blocked into 6 sections as in

Table 3.1, according to which hardware is aflFected by the settings. The hardware block diagram is shown in Figure 2.3. The parameter blocks are as follows:

Test parameters: Modality, test type and masking noise, for setting up the circuitry

in the switching unit.

Sampling parameters: Number of channels, sample rate and number of samples, for

setting up the analogue to digital (A/D) converter. These parameters

relate to the single sweep of sampling that occurs everj^ time the patient

is stimulated.

Timing parameters: Number of cycles (of stimulation), interstimulus period, stim­

ulus pulse duration and sample to stimulus period, for setting up the

programmable pulse generator. These parameters relate to the timing of

the complete series of stimulation cycles that make up an averaged evoked

potential test. For EEG tests, more than one cycle may be recorded per

test, but no stimulation is applied and the results are not averaged - they

are stored in their raw form.

Artifact rejection: Rejection type, threshold type, reject level and number of dis­

carded initial samples, for setting up the averaging code in the computer.

Artifact rejection and the meaning of these parameters is discussed further

in Section 4.2.

Montage (two blocks): Notch filter, overall reference electrode number and the ac­

tive and reference electrode number, sensitivity and filter frequencies for

each electrode in the montage, for setting up the EEG amplifier.

81 A facility for graphing and printing the timing diagram and montage for the test is provided. The diagrams are only updated when either the “Display (electrode names)” or the “Display (electrode numbers)” button is pressed. The former displays the names of the electrodes on the montage diagram while the latter displays the numbers which are printed on the EEG electrode headbox used with the system.

The “Accept record” button updates the record for the selected test type in mem­ ory, but does not exit nor save to disk. When all the test definitions for a particular modality have been updated, they may be saved by pressing “Update and exit”.

3.3 Data acquisition 3.3.1 Acquiring data

Figure 3.5 shows the window that appears when acquisition is selected. Simulta­ neously another window labelled ‘Acquisition progress report” is created and this is shown in Figure 3.6. Monitoring does not begin until the “Start” button is pressed.

This starts the acquisition process which reads the schedule and performs the tests accordingly. Messages from the acquisition process appear in the status line of the display. The last few messages may be viewed by clicking on the pull-down button of the status line. Messages are also copied to the acquisition progress report window which allows scrolling through the entire histor\' of the current acquisition session.

The first messages that appear in the status line confirm that the hardware compo­ nents have been successfully initialized. Subsequent messages are concerned mainly with the scheduling of tests - the test currently being performed, next in line to be performed, or completed. When the entire schedule is complete, this is reflected in

82 ♦ &ctrapiiy5.awetage y Becbopbys. In/e ^ Hiyribkqfcal | ir Active tSsptrf J Display threshoM» _j

Status; Waiting to psfform - EEG h 210 soimnds . — Total: |l72 Accspk |I7Z RojKt; |0 %

patent: Tea Tea: BAEP (type I) 194(1500 cyclocat It J 6 H z/ O.DS s) with trace-rcject (dynamicthreaisld] Tine: Sun Jun 27 18:18:55 139S S4S,

cnannei i AT - cz SCHi-150Qir 50UV (pg

-5 .4 5 ' SiKt

cnannei z A2- Cz SCHl-1500Hz 50UV [FS)

- 6.0 2 ' Tina (ns)

3cn>5 chaiaieis zoom fe> Zoom out Overlay last trace Cancel Ause Annotai» Exit

Figure 3.5: Data acquisition window of NICU monitoring system.

83 Phys. te a t (Phys - 1) s ta r te d Sun.;Jan. 27 18:22:25 1999 EEG tèst (BAEP le ft 1) started Sun Jun 27 18:22:25 1999 Insufficient bytes avaiüble Scon TRhX bedside monitor, insufficient bytes available from TRbK bedside monitor. 107 of 1500 cycles acquired. EEG te st completed (stop 108, 107, 1500 (ave. 7%, acq. 7$)). Scheduler detected EEG test coi^plete.. Phys. test completed (quit 2 of 54).

Bdt

Figure 3.6: Acquisition progress report window of NICU monitoring system.

the status line. Pauses and user-requested terminations are also shown on the status line and in the acquisition progress report window.

-Acquisition may be stopped at any time by pressing the '‘Stop” button. This immediately stops acquisition, even if all the cycles of a particular test have not been averaged or collected. For physiological tests and EEG tests, the remaining samples for that test will be set to zero. For evoked potential tests, only the traces collected up until that point will be averaged, i.e. even if only 10 stimulation cycles occurred, the 10 single traces of EEG response will be averaged and saved. The number of cycles acquired and actually included in the average are stored in the past schedule file for later reference (see 2.6). If the results from an interrupted test appear to be useless, they may be deleted manually from the patient record as described in Section

3.4.1. Pressing “Pause” similarly saves only what has been sampled so far in the

84 current test, so as to stop testing immediately, but allows for testing to be resumed later by pressing “Unpause”. Testing will continue at the point in the schedule that would have been reached had the system not been interrupted i.e. what remains of the current test and the subsequent tests that would have been performed during the interruption are skipped.

At any time during acquisition, the “Annotate Menu” button may be pressed to activate the annotations keyboard. This is a separate splash-proof hand-held terminal with its o^vn screen and keyboard that is placed at the bedside for nursing personnel to enter annotations. Annotations may describe significant changes in the patient’s condition or actions which may affect the integrity of the test results at the time of the annotation such as examination, movement of the patient or pausing of the system. Adding a comment does not interrupt testing. Pause and stop functions are also provided from the annotations terminal. Operation of the annotations terminal is described in more detail in Section 3.3.2.

During testing, partial test results are displayed in the acquisition window, as shown in Figure 3.5. Manipulation of this local real-time display is described further in Section 3.3.3.

3.3.2 Adding annotations

The annotations terminal is a small hand-held terminal which is connected to one of the acquisition computer’s serial ports. It has a small screen and keyboard combined into a single unit and is used by the nursing personnel to enter annotations which may be important later when the acquired data must be interpreted. Figure

3.7 shows the four displays which may be presented on the screen of the terminal. The

85 (II Suctioning (1) Phénobarbital increase (2) Coughing (2) Phénobarbital decrease (3) Restlessness (3) Hyperventilation increase (4) Turning (4) Hyperventilation decrease (5) Leaving bed (5) Hypothermia increase (6) Diagnostic test (6) Hypothermia decrease (7) Examination

<— status message — > <— status message — >

(FI) Interference menu (FI) Interference menu (F2) Treatment change menu (F2) Treatment change menu (F3) Other annotation (F3) Other annotation (F4) System stop/pause/page (F4) System stop/pause/page

OTHER AMÎ'IOTrtTIOK

Enter annotation and press enter (1) Stop when complete (ESC to cancel) : (2) Pause (3) Unpause (4) Page

<— status message — > <— status message — > (FI) Interference menu (FI) Interference menu (F2) Treatment change menu (F2) Treatment change menu (F3) Other annotation (F3) Other annotation (F4) System stop/pause/page (F4) System stop/pause/page

Figure 3.7: The four screen displays of the annotations terminal.

86 layout is designed to minimize tj-’ping by the nurses, so that they can enter relevant details about the patient with very little hinderance to their normal duties.

The main menu is shown at the bottom of the annotations screen at all times. This prompts the user to select between the four screen displays by pressing FI to F4 at any time on the unit. The appropriate annotation is chosen from a menu by pressing the corresponding number on the terminal’s keyboard. This writes the annotation to the annotations file for that patient along with a time-stamp. A confirmation also appears in the acquisition progress report window. The four screen displays are as follows:

Interference: Annotations that pertain to patient movement are included in this

menu. Movement of the patient and indirectly of the EEG recording

electrodes may cause artifacts in the recorded EEG and contaminate the

record. It is important to note such actions as suctioning, patient rest­

lessness etc. so that when the recordings are interpreted later, they may

be attributed to noise rather than a change in the patient’s condition.

Treatment change: The three most important changes in the treatment of the NICU

patient which may influence the EEG and evoked potential results con­

cern phénobarbital, hyperventilation and hypothermia. Any increase or

decrease in these levels may be annotated by choosing the appropriate

item on the menu.

Other annotation: Annotations that do not appear in the previous two menus must

be entered by hand. What is appropriate is at the discretion of the nurse.

If the system is paused or stopped, an annotation explaining the reason is

87 helpful. If the author is paged by request of the nurse from the terminal it

is also helpful if an explanatory annotation is included - this may be read

from the annotations file remotely, for example by modem from home.

Once this menu has been selected, an annotation must be entered or ESC

must be pressed before one of the keys FI to F4 will select a different

menu.

System stop/(un)pause/page: The system may be stopped or paused in case of an

emergency or extended interference with the patient. The option is se­

lected from the terminal by choosing “Stop" or “Pause" from the system

menu. The system is restarted after a pause by choosing “Unpause”. If

the system is not running (schedule complete, stopped or never started)

it is not possible to start acquisition from the terminal. This must be

done from the keyboard of the acquisition computer. The “Page” option

on the system menu activates the pager held by the author with a code

to indicate that the nurse has requested attention. The paging interface

is described further in Section 3.3.4.

All interactions with the terminal are visually and audibly confirmed. The confirma­

tion appears on the “status line” part of the display shown in Figure 3.7.

The standard annotations are hard-coded in the system and cannot be readily changed. The “other annotation” button may be used liberally if the context is not the NICU. Alternatively the code associated with the annotations terminal must be altered if a different set of annotations from the standard set described here is necessary. A data file containing the annotations screens may be more convenient.

88 and implementing such a scheme would be an improvement to the system that could be made in the future.

3.3.3 Local real-time display

Figure 3.5 shows the data acquisition (real-time display) window. A radio button above the display area is used to select one of three display options:

Averaged data: For evoked potential tests, this sets the display to show the averaged

evoked potential so far. For EEG tests, there is no difference between this

display and the “live data” display.

Live data: This sets the display to show the unaveraged single traces of EEG. This is

useful in evoked potential tests for checking for the presence of a legitimate

signal or artifacts.

Phys. data: This switches the display to show the physiological data being collected

concurrently with the electrophysiological data.

If no data are currently available in the display mode selected, because either an electrophysiological test or a physiological test is not in progress, this will be reflected in the status line when the display switch is made. The display may be switched off entirely by unchecking the “Display active” checkbox.

Checking the “Display thresholds” box superimposes on the display red lines de­ noting the rejection thresholds. If the thresholds are dynamic, they are updated in real time. The pager checkbox is checked to enable automatic paging as described in

Section 3.3.4.

89 Codes generated when acquisition terminates: 1-1 Schedule completed successfully. 1-2 Schedule cancelled at user’s request. 1-3 Fatal error occurred during acquisition. Codes signalling non-terminal conditions: 2-1 Page requested by user from annotations terminal. 2-2 Number of rejected traces exceeded 50%.

Table 3.3: Pager codes issued by the system during or after acquisition.

Non-graphical status information provided by the acquisition process is displayed

immediately above the trace area. This includes the status line which was described

in Section 3.3.1. The total number of traces acquired so far during the current test is

displayed along with the number actually included in the average and the percentage

rejected.

Below the trace area are “Zoom in" and “Zoom out” buttons for adjusting the y-axis scaling of the display. The “scroll channels” button is used for EEG and physiological

tests. A maximum of four traces are displayed at a time and this button scrolls

between the available channels in sets of four.

3.3.4 Paging interface

The monitoring system interfaces with the hospital’s network to allow it to page any CCF pager. In practice it only pages a single number - that of the author. Since the author carries a numerical pager, a numerical code is issued to indicate the reason for the page. The codes are listed in Table 3.3. Codes are divided into two categories - those issued during acquisition (non-terminal codes) and those issued when acquisition is terminated for some reason. If a page is requested by the nurse from the annotations

90 terminal, the code is 2-1. This is the only page which is not deactivated by unchecking

the “Pager” box at the top right of the acquisition display window (Figure 3.5). It

does not afifect acquisition. If “Pager” is checked, the system issues a page when the

number of rejected traces forming an evoked potential average exceeds 50%. The

system continues with the schedule despite this condition. This is only relevant if

trace rejection is selected for the particular test (part of the electrophysiological test

definition). Noisy acquisition due to a faulty electrode may result in a high number

of rejected traces and this is brought to the author’s attention by a page with the

code 2-2.

The three other page codes are issued only if acquisition has terminated. The

benign case is when the schedule is complete (code 1-1). If acquisition is terminated

from the annotations terminal or the acquisition computer directly, a code 2-2 page

is issued. Finally if the Tcl/Tk shell detects an error, a code 2-3 page is issued.

The error message is written to the patient’s “ErrorReport.txt” file. The intention

is not to detect errors in the Tcl/Tk code so much as to detect errors which cause

the acquisition code to fail. The actual acquisition (hardware interface and signal

averaging) is performed by an executable compiled from C source code. The Tcl/Tk

shell catches the condition that acquisition terminates with an error condition. This

may be caused by network errors or when the patient data disk is full.

3.4 Review and analysis 3.4.1 Viewing raw data

Once acquisition has been started and at least one test has been performed, raw data is available for viewing. This data may be viewed during or after acquisition

91 on the acquisition machine or on a remote machine. A facility for viewing data from a standard HTML browser is provided. Although the browsing machine may in principal be anywhere on the Internet, it must have access to the CCF intranet for reasons of patient confidentiality and security. Remote web-based monitoring and review are described further in Section 3.5.

Since a set of raw data is generated for every completed test in the schedule, the

“View raw data” window displays a schedule of completed tests in the same format as the schedule editor window. A red cursor is also used to select tests and is controlled in the same way as in the schedule editor. An additional “Skip to next trace” button is provided to skip forward to the next trace of the modality chosen from a pop-up menu. Test results may be viewed or deleted. There are three ways to display raw data - as a single raw trace (possibly overlaid with at most one other single raw trace), as a cascade diagram in which any number of raw traces are assembled into a three- dimensional cascade representation and as a stacked array of raw traces. The stacked array is similar to the cascade except that it is not viewed at an angle and is without perspective. The vertically stacked traces are easier to interpret than the angled cascade. Code for shading or hiding those parts of the slanted cascade representation that are below the horizontal plane is included in the system, but is compiled either on or oflf. Shading or hiding these sections improves the interpretability of the cascade representation. A better option to improve this display would be true hidden line removal. Hidden line removal in perspective plots is not a trivial operation, and it has not been implemented in this system. It is an improvement that could be implemented in future.

92 Single raw data trace display

Figure 3.8 shows the raw data display window. Zooming and panning operations by default apply to all traces simultaneously in electrophysiological test results and only the currently selected trace in physiological test results. This may be changed be checking or unchecking the “Zoom all” checkbox. A trace is selected by clicking on the background of the box containing the trace with the left mouse button. Zooming and panning affect only the y-ajds scaling, and adjust it by a preset amount. If an exact scaling is required, this may be done by clicking on either of the y-axis labels of the appropriate trace with the left mouse button. This will bring up a window in which the exact values for the y-axis extrema may be typed directly. If the “Zoom all” checkbox is checked, all axes will be adjusted to these values.

Parameters such as peak latencies and amplitudes may be calculated by selecting

“Calculate derived parameters” from the main menu. This is described in Section

3.4.2. If such parameters have been calculated for the raw trace being displayed, then clicking on the “Stats” button superimposes the results on the raw trace. The vertical lines denote the peaks of interest and the 4- marks denote the corresponding troughs, as shown for the BAEP in Figure 3.8. Although not all parameters calculated for a particular test are displayed, all the parameters are calculated from those superim­ posed on the raw display. For example, although the wave I-III interpeak latency for the BAEP is not shown on the raw display, it may be calculated directly from the wave I and wave III latencies, which are displayed. In other words, all the dependent parameters are displayed and the independent parameters are available for the pa­ rameter list, but not displayed on the raw trace display. The independent values may be changed by clicking and dragging with the left-mouse button. The “Save stats”

93 Patient' .... Test- BAEP CypeT) left (T500 cycles at 11.76 H t) Time: Wed Mar 24 17:08:261339 Wed Mar 24 1 7:18:26 1333

040

Channel 1 A1 - Cz SOHz-1500Hz 50uV (FS)

Channel 2 A2-CZ SOHz-1500Hz 50uV (FS)

- 0.10 Time (ms)

Channel: 1 Zbomout Zoom Pan up Pan down

Display stats I Update stats I Rave stsis Save ASai

.Scroll ciiaiinols _l ZtoomHian all M it trace Close

Figure 3.8: Raw data display of the XICU monitoring system. This display shows the results of a BAEP test in a subject with a subarachnoid hemorrhage. Two traces are superimposed to verify repeatability of the acquired waveform.

94 button saves the altered values of the independent parameters and recalculates the dependent parameters for use in the parameter list.

For EEG test results, “Previous cycle” and “Next cycle” buttons appear for paging between cycles of EEG data. In the case of standard test definition 1 for EEG tests, there are 6 pages of 10 seconds each making up a one minute EEG recording per test.

If there are more than four channels, a “Scroll channels” button is provided to scroll through the channels in sets of four.

Compressed array (cascade) display

Compressed array or cascade displays are displays of multiple test result traces in a three-dimensional format, as shown in Figure 3.9. The “Display compressed array” window is opened by pressing the “Display compressed array” button on the “View raw data” window. Before doing this however, a sweep of results must be selected.

This is done by moving the red cursor to the first of the test results on the schedule to be included and pressing “Mark start”. The cursor is then moved to the last of the test results to include and the “Display compressed array” button is pressed. Provided that the first and last tests are of the same modality, side and type, all tests of this modality, side and type will be included in the compressed array display. Once the display is active, the “Stats” button can be used to superimpose the independent derived parameters on the display. These cannot be edited from within this display.

Zooming and panning features are available for the y-axis, although these are slow if numerous traces are included in the display.

95 Patient Test- BAEP (type l) leit (1500 cycles 3111.76 Hz)

16:38:26

16:48:26 Channel 1 A1 - Cz 16:58:26 SOHz-1500Hz 50uV (FS) 17:08:26

17:18:26

17:28:26 T im e (m s) 0.10

Stats I Zbomout Zbcmki I Pan up Pan down

Print trace d ose

Figure 3.9: Compressed array display window of the review system

9 6 Stacked array display

Stacked array displays are generated and controlled in the same way as compressed array displays. Statistics may also be superimposed on stacked array displays. The only difiference is that the three-dimensional perspective has been eliminated in this type of display, as shown in Figure 3.10. Although the display is more plain, it is easier to identify trends when the traces are aligned vertically rather than diagonally.

For both compressed arrays and stacked arrays, a facility to print the display to any network printer or file is provided.

3.4.2 Deriving parameters from raw data

The philosophy in this project is to collect raw data, reduce it to a few meaning­ ful parameters which vary with time, assemble these parameters into a time-varying state vector and then analyze the dynamics of this vector in terms of correlations and trends. The raw data take the form of averaged evoked potential traces, electroen­ cephalogram traces and series of physiological parameters. The parameters that may be obtained from these various forms of raw data are summarized in Tables 3.4 to

3.7. The parameters derived from physiological data series are simply the separate physiological data sets (short segments in the form that they are acquired) concate­ nated into a single continuous data set (rather than the original shorter segments) by the “Derive parameters” operation.

Derived parameters can be either dependent or independent. Independent param­ eters are those obtained by mathematical operations on the raw traces themselves.

The mathematics is described in Section 4.4. Dependent parameters are calculated directly from the independent parameters. The meaning of and motivation for this

97 Tast BAEP fype 1] left 0500 cycles al T1.75 Hz)

Channel t 1Ec38:Z8 A1 - Cz / SQHz-l 500Hz SOuV (FS)

15:48:26

18:58:29

17:08:28

17:18:28

17:28:29

Tine (ms) 0.00 1Ü0 2.00 3.00 4.00 5.00 6.00 7.00 0.00 3.00 10.00

Display stals I Zoom out I Zoom in Pan up I Pan dcwn

M a t trace dose

Figure 3.10: Stacked array display window of the review system

98 Parameters: ^ BAEP peaks V- SEP peaks x^VEP peaks v ' EEG spectral properties V Physiological traces ♦ AU Test side: ■vr iBft •V- Both Right ♦ AO

Test type: [ajT

Calculate parameters

Exit

Figure 3.11: Derived parameters menu window of the NICU monitoring system.

distinction will be made clear by giving the example of the brainstem auditor}" evoked potential (BAEP). When the “Calculate derived parameters” button is pressed on the main menu, the window shown in Figure 3.11 appears. If “BAEP peaks” or “All” is selected in the “Parameters” section, and the “Calculate button” is pressed, the algorithm runs through the available and appropriate raw BAEP results and finds the values of the independent parameters listed in the right-hand column of Table

3.4. In the case of the BAEP (and the other evoked potentials), this requires a par­ ticular type of peak detection, of which the algorithm is described in Section 4.4.

The latencies and amplitudes of the salient peaks and troughs form the independent parameters. When the raw trace is displayed using “View raw data” and the “Stats” button is pressed, the latencies and amplitudes of these extrema are displayed on the

99 raw trace as the x and y coordinates respectively of the appropriately labeled crosses on the traces. The values of all of the dependent parameters can be obtained directly from the values of the independent parameters. For example, the wave I-III interpeak interval on the ipsüateral side (a dependent parameter) is obtained by subtracting the wave I latency on the side of stimulation from the wave III latency on the side of stimulation (both independent parameters). This has the benefit that if the us­ er edits the positions of the automatically-detected peaks manually, the dependent parameters may be automatically updated - they need not be separately edited by hand.

Brainstem auditory evoked potential derived parameters

For the BAEP, the latencies and amplitudes of waves I, III and V on the side ipsilateral to stimulation and the latencies and amplitudes of waves III and V on the contralateral side are automatically detected (Table 3.4). These particular peaks were chosen because they are clinically the most significant. Wave I is generated at the cochlear nerve, necessarily only on the stimulated side if stimulation is unilateral.

Wave III is generated in the brainstem and wave V in the inferior colliculus. Lesions between these points may give rise to increases in interpeak latencies. The background is given in Section 1.3.1 and provides motivation for the choice of parameters. Waves

II, IV, VI and VII are difficult to detect and clinically less significant, and have been omitted from this study.

Somatosensory evoked potential derived parameters

For the SEP, the latencies of the Erb’s point (EP) response, P14, N18, N20 and

P25 responses are automatically detected (Table 3.5). The EP response is present

100 Independent parameters Dependent parameters Wave I latency left (ms) Ipsilateral I latency (ms) Wave I potential left (uV) Ipsilateral III latency (ms) Wave I (positivity) latency left (ms) Ipsilateral V latency (ms) Wave I (positivity) potential left (uV) Ipsilateral I-III interpeak interval (ms) Wave III latency left (ms) Ipsilateral III-V interpeak interval (ms) Wave III potential left (uV) Ipsilateral I-V interpeak inter\^al (ms) Wave III (positivity) latency left (ms) Contralateral III latency (ms) Wave III (positivity) potential left (uV) Contralateral V latency (ms) Wave V latency left (ms) Contralateral III-V interpeak interval (ms) Wave V potential left (uV) Ipsilateral wave I am plitude (uV) Wave V (positivity) latency left (ms) Ipsilateral wave III amplitude (uV) Wave V (positivity) potential left (uV) Ipsilateral wave V amplitude (uV) Wave I latency right (ms) Contralateral wave III am plitude (uV) Wave I potential right (uV) Contralateral wave V amplitude (uV) Wave I (positivity) latency right (ms) - Wave I (positivity) potential right (nV) - Wave III latency right (ms) - Wave III potential right (uV) - Wave III (positivity) latency right (ms) - Wave III (positivity) potential right (uV) - Wave V latency right (ms) - Wave V potential right (uV) - Wave V (positivity) latency right (ms) - Wave V (positivity) potential right (uV) -

Table 3.4: Parameters derived from the brainstem auditor}- evoked potential wave­ forms.

Independent parameters Dependent parameters EP latency (ms) EP-P14 latency (ms) P14 latency (ms) EP-N18 latency (ms) N18 latency (ms) EP-N20 latency (ms) N20 latency (ms) EP-P25 latency (ms) P25 latency (ms) -

Table 3.5: Parameters derived from the somatosensory evoked potential waveforms.

101 Independent parameters Dependent parameters M ajor left 4-ve peak latency (ms) - Major left -ve peak latency (ms) - Major right -Hve peak latency (ms) - Major right -ve peak latency (ms) - Major middle 4-ve peak latency (ms) - Major middle -ve peak latency (ms) -

Table 3.6: Parameters derived from the visual evoked potential waveforms.

if stimulation was effective and the peripheral nerve is intact. P14 originates in the thalamus, a structure of the diencephalon. N18 originates along the thalamocortical tracts. N20 probably originates in the parietal cortex and P25 also originates in the parietal cortex. The background is given in Section 1.3.2 and provides motivation for the choice of parameters.

Together, the parameters chosen for the BAEP and the SEP cover the integrity of pathways within the brainstem, diencephalon and cortex. The fact that it is only the integrity of certain narrow paths within these regions that is observed should not be overlooked. Poor results indicate definite damage, but good results do not necessarily imply that there is no damage - only that damage is not to these narrow paths.

Visual evoked potential derived parameters

For the VEP, the latencies of the major positive and negative peaks following stimulation are automatically detected for the left (01), middle (Oz) and right (02) electrodes (Table 3.6). These peaks originate in the cortex and indicate the integrity of the visual pathways to the occipital cortex. The background is given in Section

1.3.3.

102 Independent parameters Dependent parameters Energy in delta band Betazalpha energy ratio Energy in alpha band - Energy in beta band - Burst count -

Table 3.7: Parameters derived from the electroencephalogram and its spectra. These parameters are calculated for each EEG channel (F3, CP3, 01, Oz, Fz, F4. CP4 and (32).

Parameters derived from the electroencephalogram

Section 4.5 describes the algorithm used to obtain the spectral estimate of the

EEG and the burst detection algorithm. From the spectral estimate is calculated the energy in the alpha, beta and delta bands. These are the energies in the bands 0.0 to

4.0 Hz, 8.0 Hz to 13.0 Hz and 13.0 Hz to 50.0 Hz respectively. Along with the burst count, these constitute the derived parameters for the EEG (Table 3.7). Of special interest in comatose patients is an uneven topological distribution of frequencies and slowing. Section 1.4 includes further details on the background and motivation for selecting these parameters.

3.4.3 Creating parameter lists and physiological state vectors

To create a new parameter list and physiological state vector, the procedure is to first choose “Select new list” on the main menu. This allows the user to choose a name for the new parameter list and vector. Selecting “Compile parameter list” on the main menu now causes the window shown in Figure 3.12 to appear. A physiological state vector of arbitrary dimensionality is constructed from a subset of the available parameters. The parameter number is the index for the scalar value within the vector.

103 Par.no* ModaMy Side Test type Statistic» nranetername 1 Phys al ECG - Heart rate (Ank) Heart rate (AMn) z . Riys pm afl Non-invas. BP - Mean (mmHg) Hon-kivasive Wood pressure (mean) Î Phys al Blood press.1 - Mean (mmHg) Arterial Wood pressors (mean) $ BAEP Left al - Ipsiateral l-M! kterpeak interval (ms) Left BAEP I-III interval (ms) s BAEP Right 80 %%s#ateral I-III inteipeakinteival (ms) RIgit BAEP I-III hlerval (ms)

s ...... ±1 ...... AI Al PM 1 Rzmove | Replace |

CDmpOe 1st 1 cancel 1 ______

Figure 3.12: Compile parameter list window of the NICU system.

Normally this is set to the next available integer and once the other boxes have been appropriately completed, the ”.A.dd” button adds the parameter details to the list.

If the parameter number is set to the index of an existing parameter, the “Remove” button can be used to remove the parameter from the list or the “Replace” button can be used to replace it with new details. To make the selection of parameters easier, pull­ down menus are provided to select from modalities, test sides and statistics (derived parameters) available for the particular patient. In order for this to work effectively, however, it is important to enter the details from left to right, since pull-down options to the right depend on selections to their left. The last box is the parameter name and this determines how the parameter value is labelled when the vector is displayed.

This should normally include the modality, side, statistic (derived parameter) name and unit of measurement. These details are not displayed automatically - labelling is left entirely at the user’s discretion.

104 n— iMiArrxnnusu»

tSlS9 ------i m p ------...... ------"

117m tt£«3 ...... - - ■ - ■ ■ ------

ItSM 1 . z u z

I1M® 4.74 ^SS ------— ------...... U*B«£PI-« zss w m l

a s -- - " - - ■ — ------■ - - -- •474

116 - — ------«KNBAfPk» ------IS MWVH

m a s

12 30 06 1247Ca 13 21 08 13 38 08 13 55 08 1*1208 '42908 14 4808 1503 Oel

zmWK [ iMBwtv 1 a«mbv fmim f*mm KMMwky 1

Figure 3.13: Display list, vector window of the NICU system.

Once the list of parameters has been selected, the "Compile list" button may be pressed to generate the parameter list from the acquired data. If a list of that name already exists, it will be overwritten.

If "Display list vector" is now chosen from the main menu, a display such as that shown in Figure 3.13 appears. This is a display of the list (line graph of parameter values vs time) only. The next step is to convert the list of irregularly sampled parameters into a vector of regularly and simultaneously sampled values.

105 Selecting “Create state vector” on the main menu causes a window to appear that requests a value for the interpolation interval. This is the sampling interval in seconds for the vector which is to be generated from the parameter list. Once this is set, pressing “Create vector” causes the appropriate algorithm to be executed so that the vector is generated. The mathematical details of this operation are described in

Section 5.3. If the “Display list/ vector” option is now chosen from the main menu, the displayed parameter list graphs will have the smooth vector values superimposed on them in red.

3.4.4 Viewing and editing derived data

Before derived parameters can be viewed or edited, they must be calculated using the procedure described in Section 3.4.2. If the raw data are now viewed in the manner described in Section 3.4.1, the “Stats” button may be pressed to display the values of the derived parameters relevant to the displayed raw trace. This is true of the single or overlaid raw data trace display (Figure 3.8) as well as the stacked array display (Figure 3.10). For the raw data display, there is a facility for dragging the bar representing the detected position of the EP peak with the mouse, if the left button is pressed throughout the drag. This facility is not available for stacked displays. Once the derived data for a particular trace have been altered to the user’s satisfaction, and parameters are saved to disk by pressing the “Save Stats” button. This also causes the dependent parameters for this trace to be recalculated. Parameters derived from the EEG cannot be edited graphically.

Another way to edit derived parameter values is to alter the data files by hand using a text editor, since the derived parameter data files are plain ASCII text. The

106 vector start lime: 3i3oetiaz vector end n m : 313097S47.

Analysis start time: |313DS110Z Analysis and ttm : 1913097347

Generate stale space dbgrem

ExtrapdafB Calculate varfaMty Calculate dhnensionaity

Figure 3.14: Run further analysis window of the NICU system.

philosophy in designing these file formats was to make them somewhat self-describing and as accessible to the user as possible. Therefore all files are written in plain text unless this is impractical - the raw acquired data are written in a binary format. The file formats are described in Section 2.6.

3.4.5 Analyzing the dynamics of the physiological state vector

Once the parameter list has been defined as described in Section 3.4.3 and possibly smoothed to create a state vector, further analysis is possible. The menu for further analysis shown in Figure 3.14 is obtained by pressing the "Run further analysis” button on the system’s main menu. The start and end times (UTC) for the currently selected parameter list/vector are displayed at the top of this window. Below these are the start and end times for the analysis. By default these are set the same as for the entire parameter list/vector, but the values may be edited and by doing this it is possible to analyze only a subinterval (the “analysis interval”) of a vector.

The menu for further analysis provides four types of analysis:

107 • Generate state space diagram;

• Extrapolate;

• Calculate variability;

• Calculate dimensionality.

The process of generating and viewing state space diagrams is described in Section

3.4.6.

The extrapolate analysis is provided to allow an attempt at a prediction of future

values of the parameters to be made for prognostic purposes. In particular it is

intended to be used with the EP latencies. The analysis requires that a third time,

the "Extrapolation end time", be provided. It also requires the order of the polynomial

that is to be fitted to the vector to obtain the extrapolation. The function then fits

the polynomial to the vector in the analysis interval and evaluates the function for ever}'" point in the analysis interval extended to the extrapolation end time. This is done for every parameter in the vector, and the new extrapolated vector is written

to a new vector file under a name chosen by the user. The time selection scheme allows the user to “extrapolate” over an interval for which data are already available, and in this way compare the “prediction” with what was actually recorded. Some results of this type using actual physiological data are presented in Chapter S. The mathematics of the extrapolation are described in Section 5.2.1

The variability analysis is a simple variance analysis of the vector over the anal­ ysis interval. Its purpose is to provide a tool for understanding the d^mamics of the parameters. The variability may be altered under pathological conditions. The calcu­ lation is done only for a chosen parameter, and the result can be expressed as one of

108 either a single value or a separate parameter list. The single value is simply displayed

on the screen. The parameter list is added to the vector as an additional parameter,

and in this case the smoothness of the resulting trace must be specified in terms of the

length of the moving average filter used for smoothing. The mathematics is described

in Section 5.2.2.

The dimensionality analysis provides a measure of the dimension of a chosen

parameter in the vector. This gives an idea of the number of degrees of freedom of the

parameter or the number of independent variables that would be needed to construct

a mathematical function to describe the parameter. This knowledge contributes to

understanding the system dynamics. Performing the dimensionality analysis is similar

to the variability analysis in that the result can be either a single value or a new

parameter list. The mathematics is described in Section 5.4.2.

3.4.6 Generating and viewing state space diagrams

State space diagrams are generated using the "Generate state space diagram"

option on the menu for further analysis described in Section 3.4.5. Since state space

diagrams are three-dimensional, three parameters in the parameter list must be chosen

to represent the x, ?/, and z-axes. The system then generates a text description of the

state space diagram in POV-ray (Persistence of Vision Ray Tracer) format and saves

this under the chosen file name. This file must be passed to the POV-ray interpreter

to render an image such as that shown in Figure 3.15. Extensive translation and scaling (normalization) of the axes are necessary in order to produce a reasonably

proportioned state space diagram. State space diagrams are used in this dissertation

to illustrate the concept of regions of state space that correspond to the state of health

109 Figure 3.15; State space diagram showing the state space with coordinates consisting of the ipsilateral wave I. wave III and wave V latencies of the BAEP. Points within the marked area are from seven BAEPs measured in a healthy subject and points outside the marked area are from seven BAEPs measured in a patient with subarachnoid hemorrhage.

of the individual and the idea of transitions in space between these regions that could be predicted by examining the state vector dynamics. In reality, the state vector has far more dimensions than the three that can be displayed on a state space diagram.

If the axes are judiciously selected, the state space diagram may capture important information in a way that can be verj" easily conveyed to untrained clinical personnel.

However, it is by no means intended to convey a comprehensive picture of the state of the system or its dynamics. The XICU system does not run PO\'-ray. and this operation must be performed from the command line. POV-ray is a free ray tracer

110 available for various platforms. The Linux version was used to generate the images in this research. State space diagrams are described further in Section 5.4.1.

3.5 Remote web-based monitoring and review

The web-based interface is designed to facilitate the execution of the NICU study defined in Chapter 7. It is not simply a passive review interface. It allows the user to create and alter certain types of information relating to the study. The interface is available on any machine that is connected to the CCF intranet and has the capability of browsing HTML. The intended users for the purpose of this research study are the intensive care fellows who assist in entering and interpreting the patient’s clinical data. It also allows them to review and recover analyzed data so that they too may benefit from the patient database to which they have contributed. The intensive care clinicians and neurologists may also review the data from their office computers using this interface.

The following information is available using the web-interface:

Protocol: -A. text description of the study protocol, copied from Chapter 6.

Patient data: Patient history, parameters recorded and entered during the course

of the patient’s monitoring, and a description of problems encountered

during this particular recording. Table 3.8 gives more details.

3.5.1 Navigating the web-based monitoring and review inter­ face

The web interface may be accessed within the hospital by setting the computer’s browser to access the address http://svr2.eeg.ccf.org:6500 (provided that the server is

111 Patient History Create once/edit/read Form Flowchart Create multiple/edit/read Form CT Create multiple/edit/read Form TCD Create multiple/edit/read Form Electrophys. Data BAEP Read only Single trace/stack/ table of derived parameters SEP Read only Single trace/stack/table of derived parameters VEP Read only Single trace/ stack/ 1able of derived parameters EEG Read only Single trace/table of pa­ ram eters Phys. Read only Single trace traces Problems Create once/edit/read Form

Table 3.8: Parameters that can be reviewed, edited and created using the remote web-based monitoring and review interface.

running on this machine). This brings up the site’s main menu, shown in Figure 3.16,

from which the user may choose to view the study protocol or review patient data.

-\11 pages except the protocol page are set up in the same way. The left-hand column

contains a hierarchy of links to facilitate quicker navigation within the site. At the

bottom of each page is a link to the previous page and the author’s email address

for problem reports. Since most of the pages are generated dynamically by common gateway interface (CGI) scripts, unforeseen errors are possible though unlikely, and in such a case the server will automatically email the author with the error report. Each page has a header with the name of the study, the Cleveland Clinic Foundation’s logo and a link to the Foundation’s home page. Since there is quite a lot of information on each page, it is best to set the browser’s text to the smallest size, so that evervdhing may be displayed neatly within a single screen.

112 MULTIMODALI

# W ï® ^

iUiM voft dirXoMv* tsss^tsâssjsùa 2SSSS

Figure 3.16: Main menu of the web-based interface.

Although the interface is designed to be used without additional documentation, an overview may be helpful. Figure 3.17 shows the overall structure of the site and to help in navigating it. It is possible to review data while it is being collected in the

ICU. In such a case it should be noted that simply pressing "Back" on the browser to get back to the list of times at which tests were done will not provide the most recent list. The user must press "Refresh" to ensure that the server is interrogated once again for an updated list. Also, the time of a test is the start time. Test results do not become available until a few seconds after the test has been completed. Since a test may take as long as ten minutes (for physiological tests) to complete, the user should not be surprised to see an apparent gap of up to twenty minutes, plus the time between tests, between the last available result and the present time (ten minutes from the start of the last test to the end of the last test plus ten minutes to complete the present test). Figure 3.18 shows an example of what the browser shows when "Review patient data" is chosen from the main menu. The patient's name has been replaced by the word "Abnormal" for reasons of confidentiality. Once a specific

113 M ain M enu Protocol Edit Problemi Contirm History Form Submissioii

.. index.html protocoi.htmi submitrecord-cgi

Patient Data Select Data Type FlowcbartlCT' Edit Flowchart Select Patient TCD Select Time CT TCD Form

-

proto choice.cgi datechoice.cgi di^Iayform.cgi

Vector Display Vector Select / Raw Data Array Display Channel Select 1

u:', \

vectortrace.cgi vectorchoice.cgi eiectrotrace.cgi electro trace, cgi

f T Derived Data Derived Data Electrophysio- Raw Data Raw Data D i^ Iay Select Time logicai Data Select Time IT race I Display -- ■—n ___ mo- _ ^ ~ ■ «sa t- - w. m u -

di®layformdata.cgi datet±oice.cgi eiectrochoice.cgi eiectrotrace.cgi electrotrace.cg:

Figure 3.17i Overview of pages on the web-based monitoring and review interface.

114 ?sesi5 *?W3 a a h

Figure 3.18: Patient selection page of the web-based interface. The patient's name has been replaced by the word "Abnormal".

patient is chosen, the user makes a choice as to which type of patient data to retrieve or create. The corresponding screen is shown in Figure 3.19.

3.5.2 Patient History and Problem forms

Figure 3.20 shows the top of a Patient Historv' form as it appears within the

browser. The Patient Histor}- form is a form that matches the printed history form

(Figure 6.6) described in the study protocol in Chapter 6. The Problem form consists of a single free-text box. within which comments can be made about technical issues related to the study for this particular patient. This is intended primarily to document what may be done to improve the system - such difficulties as scaling problems on the display and stimulators falling off the patient during the study may be noted here.

Patient History and Problem forms are only stored once per patient, although they may be edited as often as is necessary'.

115 MULTIMODALI#]

?<:>('«i.^v

-vHJihau,?

BttgaBttg -'• .

ggn." : Ëkmdmmdi6ùa;; '':

- '

*> B^ s ita * .

Figure 3.19: Data type selection page of the web-based interface. The patient’s name has been replaced by the word "Abnormal".

1 1 6 MULTil

n?T999( -•■-

. .. •— . . ,

Figure 3.20: Web-based interface form for editing Patient History-, for the patient called "Abnormal".

117 3.5.3 Flowchart, CT and TCD forms

Figure 3.22 shows the top of a CT form as it appears within, the browser. The

Flowchart, CT and TCD forms match the printed forms (Figures 6.7, 6.8 and 6.9) described in the study protocol in Chapter 6. These forms pertain to events which occur at irregular intervals during the monitoring of a patient. The user is presented with a page on which the relevant time must be entered before the actual form appears. Either a time may be selected from a list of previously entered forms to edit an existing form, or a new time may be entered to create a new form. The time selection page is shown in Figure 3.21.

3.5.4 Electrophysiological data review

Figure 3.23 shows the page used to select the specific modality of electrophysio­ logical data that is to be retrieved. Although it is not electrophysiological data, the physiological data from the bedside monitor is included in this list. For each modality, either the raw data or the derived parameters may be viewed. These data cannot be altered using the web-interface. The derived parameters are available only in text format. The top part of the list of derived parameters for a BAEP test is shown in

Figure 3.24. Once a modality has been selected, the user must choose the time of the test from a list of completed tests of that type which is generated by the system at the time that the page is accessed (see Figure 3.25), provided that the browser has not cached the page. If the page was reached by pressing “Back” on the browser it should be reloaded, and the request information should be reposted to the web-server. If a single test is chosen, a single raw trace (for every available channel) is generated, as shown in Figure 3.26. If multiple tests are chosen, a stacked display is generated, as

118 -5 .yi r a ^ a A r y ifrf^.WCrt, yciToM-arr- MmMcw

Rgwew l'14BiiW iMiM iiTa'r~M ~ii'Tr’im iilliP i W iirim t ■<>■■ !

Aba pro tl

.;k. .

Heetreohv»

•^OeatenerCXneordfiMrAtewnml

f U c I f f 7^~r\ S 3 ÎÎ EST 1999

| aM»| r

IfalHBdtttnnitttltelMtttorAliiieniMl Ok

Figure 3.21: Time selection page of the web-based interface for flowchart. CT and TCD forms, for the patient called ".A.bnormal".

1 1 9 M«mMcnn R»iew

Histeffg Flftwchatt Ct àÊ:ï±3^4 E» ^='SS. m m ElwTOlfff^ Vterm ...

BtopcqI 2 0 0 0 # #

A gM m taCTm dlgiteE «Æ cim fin-*liMrntf

Figure 3.22: Web-based interface form for editing CT data, for the patient called "Abnormal".

1 2 0 The CLEVELAND C lmc n Foundation k J MULTIMODALITY EVOKED POTENTIAL STUDY

M ain M enu

Review A bnorm al Abnormal: H istory

F low chart Electrophysiological C T Data TCD

Electrophys.

V ectors

Problem s RAW DERIVED TRACES RESULTS Protocol

Brainstem auditory EP 52 g

Somatosensory EP 5! 2

Visual EP □ □

Electroencephalogram 5Î □

Other physiological data a 2

EEG spectra □ □

Return to data type selection for Abnormal

Figure 3.23: Electrophysiological test modality selection page of the web-based inter­ face. for the patient called "Abnormal".

121 ;‘i!SSjisiSëie%S5®siï)s^ M an Mena Swag AbnonMl

Hiftoiv Rqwchaff CL î i ' ICQ Bt?g??hyT R«

VecTorc

Pf«h?e«wa Prmtcml

s :

Figure 3.24: Derived parameters display page of the web-based interface, for the patient called "Abnormal".

122 shown in Figure 3.27. The stack display is for a single channel only, and the channel must be selected from a list of available parameters for that modality. All graphical displays are automatically scaled and no zooming or panning facilities are provided.

3.5.5 Vector data review

Vectors must have been generated on a review workstation before they become available on the web-based interface. All vectors that have been generated on a workstation will appear on the list that is displayed when the web user chooses to view vector data. The vector appropriate to the user’s selection will then be displayed as shown in Figure 3.28. The image that is generated is too %dde for the screen, and the browser will supply a horizontal scroll bar for viewing the entire width of the display.

3.6 Archiving patient data

The monitoring system includes a simple patient data archiving interface. This can be executed on any review workstation to which a magnetooptical (MO) disk drive is attached. The archive software is started by running "NICUArchive" from the NICUsystem/.A.rchive director}". This produces the menu shown in Figure 3.29.

The user presses "Select new patient” and chooses the name of the patient of whom the data is to be archived or restored from the pull-down menu. The pull-down menu lists the names of patients of whom data is stored on the hard drive (current patients) as well as the names of patients of whom the data is stored on the currently mounted

MO disk. Once a patient is selected, the two remaining options on the archive main menu become active. The first, "Copy from mo to hard drive” restores the selected patient’s data from the MO disk. The second, “Copy from hard drive to mo” archives

123 THE CLEVELAND CLMICfn Foundation b J MULTIMODALITY EVOKED POTENTIAL STUDY

Main Menu

Review Abnormal Abnormal: BAEP History

Rowchart Raw Selection

CT

TCD

Electrophys. Mon Dec 07 15:05:02 EST 1998 left Raw Mon Dec 07 15:15:02 EST 1998 richi Select Mon Dec 07 16:20:51 EST 1998 left Mon Dec 07 16:2-1:15 EST 1998 left Vectors Mon Dec 07 16:27:14 EST 1998 richt Mon Dec 07 16:30:28 EST 1998 left Problems Mon Dec 07 16:33:18 EST 1998 richt Mon Dec 07 16:38:31 EST 1998 left Protocol Mon Dec 07 16:48:31 EST 1998 richt Mon Dec 07 16:58:31 EST 1998 left Retrieve

Return to electrophys. modality selection for .-\bnortnal

André van der Kouwe 25564vandera @ eeg.ccf.org

Figure 3.25: Raw data time selection page of the web-based interface, for the patient called "Abnormal".

124 rîiîTsfefis KÆILTIMODALIWri

CS-f'W

Btmtlin RggSmn

Figure 3.26: Single raw data trace display of the web-based interface, for the patient called "Abnormal".

125 MULTIMODALI iSSBI

Figure 3.27: Stacked raw data display of the web-based interface, for the patient called "Abnormal".

Figure 3.28: Vector display of the web-based interface.

1 2 6 Ho patient currently selected

Select new patient

Omw from mo U> harü rtnvc Cojiy from torû dnv& to nto

r- û ^ C' ''' <î ^ o

B dt

Figure 3.29: Main menu of the patient data archiving interface.

the selected patient’s data to the MO disk. When a patient’s data are archived, the local copy is not automatically deleted.

3.7 Summary

The work in this chapter represents an investigation into a user interface for a neurological monitoring system and how to integrate various aspects of a research study. system for acquiring multimodality patient data according to a predefined but highly tailorable schedule is discussed. The data are made available during ac­ quisition and for retrieval at a later date to the fellows and hospital staff in various ways using the existing network at the hospital. Basic data processing is available to extract meaningful features from the raw data to make the system more useful to untrained clinical personnel. Various forms of data analysis are available within the framework of a time varying physiological vector representation of the processed data. The implementation of the design was discussed in Chapter 2. This work

127 was prGsentad. as a, papsr and poster at the 3rd International Workshop on Biosignal

Interpretation in Chicago in June 1999 [146].

128 CHAPTER 4

Raw data analysis

4.1 Introduction

The raw data acquired directly from the patient undergo various operations to extract meaningful clinical information from them. The types of data and the opera­

tions performed on them are illustrated in Figure 4.1. The first operation for EPs is averaging the individual EEG sweeps and removing contaminating artifacts. Several authors such as Bezerianos et al. [12] and Muhler et al. [91] studied EP averaging and artifact rejection and an approach similar to that of Cluitmans et al. [29] is chosen here because their application resembles the application in this work. The resulting averaged traces are stored so that a clinical neurophysiologist can review them if nec­ essary. The EEG data before averaging are not stored because the volume of data is excessive. This means that the artifact rejection parameters cannot be changed after the patient has been monitored. The next step is processing the averaged EPs and the EEG. This includes peak detection, spectral analysis, finding the energy in the significant frequency bands of the EEG and identifying clinically significant pat­ terns in the EEG (burst-suppression in this study). EP latency detection is included because the EP latencies hold critical information regarding the state of the NIGU

129 4.5 A cquired 4:2 x;(f) data Single sweep EEG responses to Electroencephalogram. Physiological parameters. stimulation.

Processing 4.2 during Averaging and artifact acquisition rejection.

4.3 Raw data Averaged evoked potentials.

4.4 4.5 Raw data analysis Peak detection. Spectral analysis and burst counting.

4.4 4.5, 4.6 PSD{ f) D erived param eters Peak positions. EEG power spectral densities, band powers and burst counts.

P aram eter lists Parameter lists.

Figure 4.1: Summarj' of raw data types and types of analysis.

130 patient’s CNS [85] [103] [139]. Many authors promote the use of the EEG and the spec­ tral properties of the EEG for continuous monitoring of NICU patients [83|[113j[148j.

The diagnostic and prognostic utility of burst counting in critically ill subjects has also been established [128]. These analyses are considered to be ‘haw data” analyses and they give rise to “derived parameters”. It is not necessary to be trained in clinical neurophysiology to interpret the derived parameters, and this is why the automatic analysis is necessary in the bedside unit where clinical neurophysiologists may not be available. The raw data analysis algorithms are described in detail in this chapter.

The use of AR spectral estimation is common in EEG analysis [15] [44], and the use of matching pursuit is promoted by several authors [36[[151[. The derived param eters are formed into lists and vectors and further analysis of the vector dynamics is possi­ ble once this representation has been obtained. List and vector analysis is discussed in C hapter 5.

4.2 Averaging and artifact removal in evoked potentials

The EEG is the recording of the electrical activity of the brain. This signal is typically measured by electrodes on the scalp, and reflects the summated electrical activity of millions of underlying neurons. The evoked potential (EP) is the electrical response of the brain to stimulation. For example, in the case of the brainstem audi­ tor}- evoked potential (BAEP), the stimulus is a click to the ear. The auditory pulse is transducted by the cochlea and the resulting electrical signal travels along well- known neural pathways to the auditory cortex, where it is perceived as a sound. As the signal travels along the auditory pathways, it passes through nuclei at well-defined times after stimulation. When the signal jumps a synaptic gap, volume conduction

131 allows the activity to be recorded using EEG scalp electrodes. The scalp recording of the EEG for a short time after stimulation is a single-sweep EP. It is common clinical practice to average hundreds of single-sweep EPs to obtain an averaged EP in an effort to reduce the ongoing background EEG noise [117|. The EP amplitude is between 1 and 20 fiV while the EEG am plitude is between 20 and 100 fiV.

Let the single EP sweeps be denoted by {xi(i)|0 < t< T} where i = 1,.... JV. T is the duration of the single sweep, for example 10 ms for the BAEP, and N is the total number of stimulus cycles or individual sweeps.

There may be occasional noise in the EEG resulting in outlying points in the single

EP response. These outlying points are defined as artifacts i.e. Xi{t) is an artifact if |xj(t)| > Ti where is a threshold and Xi{t) has zero mean because the EEG is high-pass filtered by hardware during acquisition to remove the DC offset and drift.

This definition of artifact was also adopted by Cluitmans et al. [29|.

To reduce the effects of the artifacts on the average, they are eliminated before averaging. Several schemes have been studied by other investigators [127|. In the

NICU project, the goal is at least to mimic the behavior of commercial EP monitors, and for this reason only some more conservative artifact removal techniques have been implemented. The NICU monitoring system provides four types of artifact removal (none, artifact clipping, artifact rejection and trace rejection) for which the parameters listed in Table 4.1 must be selected. If an artifact is detected, the entire single trace may be eliminated from the average. This is the “trace reject” type of removal. .A.lternatively, just the artifact itself may be eliminated. If artifacts are set to zero, the rejection type is “artifact reject”. If artifacts are set to the nearest threshold value, either positive or negative, the rejection type is “artifact clip”. The

132 Parameter name Range of possible values Rejection tj-pe Artifact clip; artifact reject; trace reject; none Threshold type Dynamic, static Reject level absolute threshold or 10 times mean peak value Number of discarded initial samples Number of samples

Table 4.1: Artifact removal parameters and their possible values.

rejection types are illustrated in Figure 4.2. Cluitmans et al. scanned each trace for artifacts before averaging. If the artifact could be identified and removed, the trace was included in the average, otherwise it was eliminated [29].

Let (7/i(t)|0 < t < T} denote the single EP sweep after artifact removal. Then without artifact removal:

Uiit) = Zi(t)|o

For artifact clipping:

TiifXi{t)>Ti Viii) —Ti if Xi(t) < —Ti for 0 < t < r Xi{t) otherwise

For artifact rejection:

î'-W = { otheiï fo^0

And for trace rejection:

yi{t) = ( ° > n for 0 < t < r [ Xi{t) otherwise

Once the single responses have been filtered for artifacts, the averaged EP, z(t)

Zjv(t), is obtained by ensemble averaging as follows:

133 none

artifact clip

artifact reject

trace reject

Figure 4.2: The four artifact removal types, The dashed lines indicate the rejection level, which may be static or dynamic.

134 i ZjW = IE I Viii) t=i

The final choice of artifact removal technique is left to the user, and will depend on the situation. For example, in the NICU, trace rejection may be preferred since the noise is temporally spread out, being caused typically by patient movement, while in the INR the X-ray imaging equipment causes very short bursts of noise, that may be better removed with artifact rejection or artifact clipping. The “safe” option of not removing artifacts is always there, in which case a larger number of traces may have to be averaged.

The threshold value for artifact rejection can be either static or dynamic. For static artifact rejection, a fixed threshold in microvolts is specified for all traces:

Fi — T"|z=l,...,V

For dynamic artifact rejection, the threshold is updated as signal averaging pro­ gresses. It is calculated as a fixed multiple of the mean peak value of the traces. The mean peak value is the average of the absolute peak value of each of the single traces contributed to the average so far in the acquisition process:

Ti - e max o<£

where the fixed multiple e has a value chosen by the user, typically 1.2 to allow

20% fluctuation in the separate EP traces. This value was chosen because it results in a rejection rate of about 5% of traces in real patients under quiet experimental conditions. This is a clinically established reasonable rejection rate in common use.

If the situation is such that more cycles of stimulation can be tolerated, the value of

135 e may be reduced. Muhler studied averaging off-line, and found that simple clipping below 25 jj.V is almost as effective as wieghted averaging or artifact rejection with a dynamic threshold [91|.

Since zo(t) is undefined, the first few values of r are set to oo:

Ti = oo|i

In effect the first I traces are included in the average regardless of artifact content.

After some experimentation, a value for / of 5 was chosen. This is a compromise between allowing too many artifacts to be included in the average, and obtaining an average based on too few traces. If distributions for the normal EP amplitude and artifact amplitude were available, an optimal Bayesian threshold could be calculated.

This information was not available, but the threshold can nevertheless be set to an optimal level if this is calculated externally from a sufficient amount of data. The average is used for the artifact removal threshold. Another approach that was not implemented would be weighted averaging, in which:

max I , 0

where

k i- 1

and

2 i(0 ) = 1

or an adaptive least mean square exponential scheme such as that of Svensson could be used [137].

136 The choice of the constant k remains. Given the distributions of the signal and artifacts, an optimal value could be calculated using statistical decision theory to achieve a required SNR. In the NICU system, a value of k=o is chosen rather arbi­ trarily to obtain EPs that are clinically acceptable.

Bezerianos presents two algorithms that replace conventional artifact rejection.

The algorithms compute the weight for each trace depending on its similiarity to the others. It was verified experimentally that the technique reduced the number of trials required to form a reasonable average [12].

The signal to noise power ratio of the averaged EP increases in direct proportion to the number of averaged responses. By the strong law of large numbers [104]:

Form the sum z{t) of n independent identically distributed (i.i.d.) random pro­ cesses:

i = i

Then z(() is a random process with mean Tj(t) and variance cr{t) as follows:

4 iz i = l

i=i If it is assumed that rji{t) to 77,v(t) are identical and the noise variances cri(t) to cr,v(i) are equal, then the signal to noise power ratio SNRave of the average z{t), is:

= N ■ (4.1)

The assumption that the single EP sweeps are truly i.i.d. is not necessarily a good one. It is nonetheless the prevailing assumption in the literature [117]. Since

137 the “noise" is background EEG, it is not necessarily independent of the signal [147].

This is discussed further in Section 4.3. The violation of these assumptions results in (4.1) being slightly optimistic. It is nonetheless a reasonable guide in choosing the number of samples to include in the average.

Now the effect of the various types of artifact removal on the averaged EP may be qualified. For “trace reject” artifact removal, the signal to noise ratio of the average decreases so that it is in proportion to the number of remaining included traces, as specified in 4.1. Commercial EP monitors typically implement the trace rejection type of artifact removal. There is a subtle difference in the implementation, however. If a commercial EP monitor is commanded to acquire 1500 traces, it will keep stimulating the patient and adding or rejecting the responses until 1500 traces are accepted. The

NICU system acquires 1500 traces and includes in the average only the subset that was not rejected. Although this has the definite disadvantage that the signal to noise ratio of the acquired signal is not guaranteed, there are two reasons for doing it this way that overshadow the disadvantage. The first is that the system must be able to run unattended. If a recording electrode were to become detached, every trace would be noisy and rejected and the patient would be stimulated repeatedly wdthout limit. The second is that if a significant number of traces were rejected, possibly for legitimate reasons, the length of a test might exceed the allocated time in the schedule of tests, thus jeopardizing the successful completion of the next test in the schedule.

This is why the system is designed to limit EP tests to a fixed number of stimulation cycles but records the number of cycles (see Section 2.6) actually acquired so that a signal to noise ratio analysis can be done afterwards if necessary.

138 For removal of the “artifact reject” type, the distribution of the artifacts in time in the trace is important. If it is assumed that the artifacts are uniformly distribut­ ed in the trace and that the trace statistics are variance-ergodic (the time-variance and the ensemble variance of the process are equal) the signal to noise ratio of the averaged trace will again simply decrease so that it is in proportion to the ratio of accepted to total points on the trace. However, these assumptions are not necessar­ ily appropriate. One significant artifact is the stimulus artifact. This artifact is the result of electromagnetic induction or electrical conduction between the stimulating transducer and the recording electrodes, hence it appears in the trace at the moment of stimulation. It is not uniformly distributed at all - it occurs always within the first few hundred microseconds after stimulation. To eliminate this problem, the first tE milliseconds of the trace are ignored for the purposes of artifact rejection, where ts is defined in the test definition as “discarded samples”. Parsa et al. present a nonlinear adaptive filter for the SEP stimulation artifact, pointing out that it may mask the start of the artifact [106). McClean et al. developed a conceptual model for the SEP stimulus artifact and a means of reducing it [80|. No method for specifically- eliminating the stimulus artifact was implemented here as the stimulus artifact, even in the case of the SEP, was not large enough to mask the part of the response that is of interest in this work. Any other artifacts that are not time-locked to the stimu­ lus are equally likely to appear anywhere in the trace, and therefore have a uniform distribution. Since stationarity is not necessarily a reasonable assumption, neither is ergo dicity. .A.gain, stationarity is the prevalent assumption in the literature despite the fact that these biological signals tend to be non-stationary [117|. .4s for simple averaging, the result is simply that the SNR of the averaged, artifact rejected trace

139 will decrease a little less than in simple proportion to the ratio of accepted to total

points on the trace. The effectiveness depends on the characteristics of the artifacts.

For specific types of artifact, an alternative approach may be more effective. For

example, Nakamura et al. eliminated ECG from EPs measured with a non-cephalic

reference before averaging by means of a synchronized detection technique. In this

work the reference is cephalic and ECG artifact is not a problem [96]. McGillem et

al. reviewed EP filtering in 1981, discussing approaches such as Wiener and Kalman

filtering and suggested a nonlinear and adaptive pattern recognition approach based

on statistical decision theory.

The "artifact clip” type of rejection lessens the need to assume a uniform distri­

bution by “filling in” in place of the artifact w^hat is presumably close to the true

uncontaminated signal i.e. the signal saturated at the clipping limit. In the example

signal of Figure 4.2, this seems intuitively to be a reasonable assumption. If the signal

is very low in amplitude, however, and contaminated by occasional high-amplitude

bursts of noise, then it is intuitive that the “artifact clip” type of rejection would

reduce but not eliminate the effects of the noise on the average.

4.3 Modeling the averaged evoked potential

In this section, a model is developed to verify the performance of the peak detec­ tion scheme devised in Section 4.4. The recorded single-sweep and averaged EP, z{t), are often modeled as the sum of the true EP, s(t), and noise, v{t) [25][59][101|:

z(t) = s(t)-f-u(t) (4.2)

140 This is despite the fact that the EP and the “background” EEG noise may not be additive. Some reasons that this assumption might be violated are overlapping of evoked responses, locking to an external source such as the power line frequency and stimulus-locked physiological responses other than the EP itself [23]. Overlapping evoked responses are eliminated by waiting for a sufficient length of time between cycles of stimulation. For example, for the the BAEP the auditor}' clicks are applied at a rate of 11.71 Hz. Stimulus cycles are 85 ms apart, but the response is only recorded for 10 ms. The 11.71 Hz stimulation rate is chosen because it does not divide into the power line frequency. A sleeping subject or anesthetised subject will have much-attenuated muscle activity, and this will also reduce time-locked noise.

Karjalainen points out that the averaged EP is the first moment of the joint distribution of the measurements and that second order methods such as principal component analysis may yield more information [59]. In this work the goal is simply to obtain the clinically established useful information in the EP trace. Higher-order statistics are not yet used in clinical practice. Rather than use these higher order statistics, the tendency in the literature has been to concentrate on the single sweep response to an individual stimulus. The problem is to remove the background EEG from the observed response to obtain a single-trial estimate of the EP. Time-invariant digital filtering [102], adaptive filtering [25], Wiener filtering and model-based estima­ tion [101] have all been applied to this problem. The research in this dissertation is not principally concerned with the estimation of the EP itself, but rather the changes that are observed in the standard, clinically-used EP under certain conditions of neurological illness. Therefore the clinically prevailing averaged EP is modeled and

141 I’m! &ACP(iyp« i)Mt(l000cy«M4lll 7B»u ) TmC. aA£P(typ* l)n9W(I000cvaMM it 7*Mzi r«n* ruauw 31737SSi«ee rmr 3 174656 1908

3 OOQIQO 200 3 00 4 00 5 00 60C TOO 600 SOO«COd

Figure 4.3: 'True” left and right BAEPs obtained by grand averaging a series of averaged left and right BAEPs recorded from a healthy subject.

analyzed here, even though it is understood that the newer experimental methods may ultimately deliver more clinical information or the same information more quickly.

Although the literature concentrates on single-trial EPs. some of this information may be applied to modeling the averaged EP. The additive model (4.2) of Karjaleinen et al. [59| is combined with the background EEG modeling approach of Davila et al.

[311-

The true EP. s{t). is created easily by averaging a series of averaged EPs recorded in a healthy subject with no known neurological abnormalities. This averaged average, or grand average, still contains some noise, but the power SXR is reduced by the number of averages averaged, by equation (4.1). Since this is the best that can be done with the available data, it will be used as the "true” EP. The grand average was inspected visually to verify that it resembles a text-book EP and the peaks were identified by a human expert. The "true” left and right BAEPs are illustrated in

Figure 4.3 and the "true” left and right SEPs are illustrated in Figure 4.4.

142 C

*"• I OOO S 00 tGOC tsao JOOO ZSOO »00 IS00 4000

Figure 4.4: "True" left and right SEPs obtained by grand averaging a series of averaged left and right SEPs recorded from a healthy subject.

143 The next step is to model the additive noise, v{t). Davila et al. modeled the

EEG additive noise as an AR process [31]. Karjalainen et al. went a step further and

modeled the EEG as a time-varying AR process [59). In this work a fixed AR process

is chosen to model the background EEG noise. This is a reasonable assumption, since

the power spectral density of the EEG is characterized by peaks. The additive noise

in the frequency domain, V { f ) is given by the model:

= D'(/) A ( f )

where U { f ) is white Gaussian noise with zero mean and unit variance, and A { f )

provides the color to simulate the background EEG signal, is the white noise

variance and is used to adjust the SNR of the synthesized test signal.

To obtain the AR model order and pole positions, the set of EPs used to obtain the grand average was used. The grand average or “true” EP was subtracted from each of the noisy averages to obtain a trace containing only the background EEG noise.

These noise traces were then analysed. The Akaike information criterion (AIC) was obtained for the model orders 2 to 70 for each trace of both EP types on the left and right. The set of superimposed plots of the AIC vs AR model order (number of poles) is shown in Figure 4.5 for the BAEP and SEP. From these graphs a model order of

24 is chosen for both the BAEP and the SEP background EEG noise models. The final prediction error (FPE) gave similar results, and is not shown here.

The complete model for verification purposes is now constructed as shown in

Figure 4.6 and described by the equation:

z{t) = s{t — Tq) + a-[u{t) * a(t)j

144 Figure 4.5: Akaike information criterion (AIC) vs model order (number of poles) for the BAEP and SEP background EEG noise. In both cases the AIC is minimized at approximately 24, and this is therefore the chosen model order for the noise.

True EP: S(t)

-To White noise: L/(f)

EEG AR filter a(t)

Synth. EP: 2 (f)

Figure 4.6: Model of obseirved averaged EP incorporating variably delayed true EP and colored EEG background noise.

145 where s(i) is the true EP approximated by the grand average EP from a normal subject, Tq is a time-delay introduced to simulate a pathological shift, u{t) is real white

Gaussian noise with zero mean and variance cr^ and a{t) is the impulse response of the AR filter that colors the noise to match the background EEG.

Tq is a random variable with a Gaussian distribution of which the mean and variance depend on the peak that is being found. The mean and variance of each peak latency is given in Table 4.2 for the BAEP and Table 4.3 for the SEP. It is common clinical practice to model the distribution of these peaks as Gaussian [117], and also common to express the variance in terms of 3 times the standard deviation

[23|, contrary to what is typical in signal processing.

The SNR for the synthetic noisy EP z{t) is given by:

SNR = ______(T- * a{t)]dt

The AR parameters for the discrete implementation a[n] of the continuous filter a(t) were obtained using the covariance method [60|. The corresponding poles were found by factorizing the AR polynomial using Laguerre's method [109] and are illus­ trated in Figure 4.7 for the BAEP. The spectral estimation algorithms are revisited in Section 4.5.2. The PSD for the BAEP background noise is shown in Figure 4.8.

Thakor modeled the EP with orthonormal basis functions such as the Fourier and

Walsh basis functions [139|. Liberati presented two stochastic models for the single­ sweep EP that are similar to that presented here, but include time-v'ariation in the

AR model to simulate intrasweep variation [70).

146 0.8

0.6

0.4

0.2

X O

- 0.2

-0.4

- 0.6

- 0.8

- 1 - 1 - 0.8 - 0.6 -0 .4 - 0.2 0.2 0.4 0.6 0.8

Figure 4.7: Pole positions for four realizations of the AR spectral estimate for the background EEG noise of the BAEP.

15000

10000

o s

5000

5001000 1500 2000 2500 3000 3500 4000 4500 5000 frequency (Hz)

Figure 4.8: Estimated spectra for the background EEG noise of the BAEP.

147 4.4 Peak detection in evoked potentials

The monitoring system has been equipped with an algorithm for identifying the peaks of the evoked potentials. These are the peaks labelled in Figures 4.3 and 4.4 in Section 4.3. The peak positions give useful clinical information about the state of the neural pathways in the patient from whom the averaged EP w^as obtained. In its current form, the purpose of the algorithm is simply to assist the user in finding the peaks after they have been recorded. In a commercial NICÜ monitoring product, an algorithm would be included that operates in real time to accurately find the peaks.

The purpose of automatic peak detection is to allow unskilled NICU personnel to interpret the EP results without having to look at the raw waveforms. The raw waveforms are difiicult to interpret. Furthermore, in a system such as this one, manual interpretation of the repeated tests would be laborious. Automatic detection also allows the computer to sound an alarm when the values start to fall outside the normal limits. Sattar et al. modeled the EP as a template and colored noise and used a bandpass filter-bank to expand these into basis functions [124|. They matched the template with the signal to extract the single-sweep EP. This resulted in faster acquisition of the EP. The peak detection approach presented in this section is based on a somewhat similar idea. Williams et al. modeled the EP as a pole-zero filter and passed the parameters to a neural network that classified the subject as ‘injured” or

“not injured” during high-risk surgery in the OR [152|.

A feature detection algorithm for detecting singularity-like features in a one­ dimensional signal using the continuous wavelet transform (CWT) w^as originally developed in preparation for this project [143]. It is based on the zero-crossings wavelet representation introduced by Mallat [73], and is described in Sections 4.4.1 to

148 4.4.5. It involves a probabilistically-weighted cross-correlation of the transform of the signal in the various frequency bands with that of an “ideal” example feature, that is, a matched filter. As such, the colored noise that is the background EEG should be whitened before matched filtering. This is not done here. Mallat’s quadratic spline wavelet of compact support [75] was chosen for this implementation, as its form re­ sembles the evoked potential peaks. It is the derivative of the cubic spline smoothing function. It is explained in Section 4.4.1 that a wavelet which is an approximate derivative of a smoothing function is required. An obvious choice of smoothing func­ tion would be a Gaussian function. In fact, multiresolution edge detection done using the derivative of a Gaussian-smoothed signal is known as Marr-Hildreth edge detec­ tion [75j. In Section 4.4.2, it is assumed that il>{x) has compact support. This is not true of the Gaussian function in either the time or the frequency domain. Ideally, a function should be found that has compact support in both domains. For compact support, a truncated Gaussian could be used. In Section 4.4.1 a fitted quadratic spline is used.

A discrete version of the feature detection algorithm was implemented in the NICU monitoring system. The move to the discrete version was inspired by the fact that the fast wavelet transform is as efficient as and sometimes even more efficient than the fast Fourier transform algorithm [28], whereas the CWT is not very efficient [123|.

Also, the wavelet transform's standard competitor, the short-time Fourier transform, has been described as less tractable for real time analysis than wavelet decomposition

[123]. The discrete wavelet transform is therefore a reasonable choice for a real-time application. The Daubechies wavelet was used for this implementation, because it has compact support in both time and frequency, unlike the quadratic spline wavelet.

149 The Daubechies wavelet has a peculiar appearance, and is discontinuous in some derivatives. This is a necessary condition of wavelets with compact support [109].

The Daubechies wavelet of 20th order (D20) was chosen because it is smoother and more similar to the EP peak waveforms than the lower order Daubechies wavelets.

Other wavelets such as the Coiflet wavelet, mentioned again in Section 4.6.2, may also have been appropriate but were not evaluated [6j. The discrete detector was less successful than the CWT version. It is described further in Section 4.4.6, where its performance is evaluated using the model of Section 4.3.

Blinowksa et al. applied wavelets to extract the single EP from the EEG but remark that matching pursuit gives better time-frequency resolution. They demon­ strate this using sleep signals and simulated signals [16].

4.4.1 The zero-crossings wavelet representation

The zero-crossings wavelet representation gives rise to a convenient norm that is useful in feature detection [73], and for this reason it is is used in the peak detector.

Intuitively, the zero-crossings wavelet representation may be interpreted as a tabula­ tion of the zero-crossings of the second derivative of the signal smoothed at various scales. According to the Logan theorem [73], the signal may be reconstructed from this information, within certain constraints on the properties of the signal. Since this representation is unstable in the sense that small changes in the representation may give rise to large changes in the reconstruction, Mallat has included also the measure of the area between zero-crossings of the signal’s second derivative in the represen­ tation. This has the added advantage of giving rise to the simple norm in L^(5R), which will be described later and is useful in feature detection [73]. The stabilized

150 version is used in this work, though pattern recognition and not signal compression and reconstruction is the purpose here.

Let L^(3?) represent the Hilbert space of square-integrable functions of a real scalar and let f{x) represent a function on L‘^{R). Then the dyadic wavelet transform of f{x) at the scale 2-' , jeZ, is defined by

where the subscript denotes scaling as follows

1 X

If 6{x) is a smoothing function, such as a Gaussian, and 'é{x) is defined as

^ then

= / * dx^dx^ • «2-)W (4.3) so that the wavelet transform is proportional to the second derivative of the function

/(x) smoothed by 02^ (x).

The zero-crossings of the wavelet transform are represented by Zn- The area between two consecutive zero-crossings, 2„_i and Zn, is given by

en = f W2jf{x)dx (4.4) Jzn~l Finally, the zero-crossings wavelet representation is given by

---- (4.5) ^71 — 1

151 The method presented by Mallat [73] for determining e„ is not used in this ap­

proach to peak detection. Instead of finding the zero-crossings of the second deriva­

tive, the extrema of the first derivative are used in the manner described by Cvetkovic

and Vetterli [30]. It is shown by Mallat [73] from (4.3) that

— ( / * doj){Zn) — ~^{f * ^2j)(2n-l)

SO that the e„ values can be determined from the wavelet transform directly also if

the wavelet is defined as the first derivative of the smoothing function

and

= 2; [PV2;/(Z^) - and the Zn values are defined as the transform extrema positions instead of the zero- crossings positions. The equivalency of the representations derived in these two man­ ners is supported by Cvetkovic and Vetterli [30].

Mallat and Zhong also describe the use of the first derivative approach [75], but in this work they identify the wavelet modulus maxima rather than the wavelet ex­ trema. They point out that the first derivative modulus maxima correspond to inflec­ tion points of the signal where the gradient is steepest, whereas the modulus minima correspond to signal inflection points of minimum gradient and they thus motivate the irrelevance of the extrema which are not also modulus maxima and the corre­ spondingly greater efficiency of the representation. They also present an algorithm for reconstruction of the signal based on the wavelet modulus maxima. Cvetkovic and

Vetterli [30] prefer the use of extrema, since their scheme for signal reconstruction

152 from wavelet extrema is simpler than that of Mallat. Furthermore, it is a consistent estimator of the original signal, a property which arises from the convexity of the representation, referring to the property that all of the information in the represen­ tation is expressed in the reconstruction [30]. The modulus maxima representation, in contrast, is nonconvex. Also, the number of wavelet extrema is not radically larger than the number of modulus maxima for a typical signal [30].

In the continuous peak detector, singularity-like features are identified by means of the zero-crossings representation obtained from the wavelet extrema. An additional advantage of this is that wavelets which are approximate derivatives of smoothing functions occur more frequently in the literature than second derivative wavelets.

4.4.2 Feature detection using the zero-crossings representa­ tion

By employing Parseval’s theorem, it can be shown [28]]73] that

ll/w f = Ë lln-i/WlP J = — OO which is an expression of the energy* of the signal and the simple norm in referred to earlier.

It follows that the distance in T^(5R) is

OC li/W -sW ll"= E l\W2if{x)-W^g(.x)H- (4.6) J = —CO

Following the derivation of Mallat [73] for the purpose of detecting singularity-like features, consider the wavelet transform of the Dirac delta centered at u

W2i5u{x) = 5u{x) * 'iÙ2j (a:) = - u)

153 Assuming that the energy of ^(x) is concentrated over the interval [-cr, a\, which is only exactly true for wavelets with compact support over this interval, the energy of the Dirac delta is correspondingly concentrated in the wavelet transform over the interval [u-2^a, u-r2^a\.

The distance function (4.6) for a singularity in / at u and a singularity in g at zero, denoted d-u{Zf,Zg), may thus be approximated by the local function

CO d^(Zf,Zgf= •£ di(Z^f,Z,^gf (4.7) where

di(Z2 jf, Z 2 igŸ = f \Zojf{u + x) - Z23g{x)f dx (4.8) J- 2 1 e and

6 = a (4.9)

Let je{l, ..., J] represent the relevant octaves. For discrete zero-crossings repre­ sentations, the number of octaves J over which the wavelet transform is determined must be finite, and therefore the representation is only approximate. In the practical case, however, if J is chosen appropriately, the energy in the omitted levels will be negligible. In the case of the evoked potentials, the signal is band-limited by hardware.

Let the vector be comprised as follows

du = [du(Z2i/, Zgip),..., diiZojf, Z2jg)\

In the discrete case, the distance function is approximately equal to the magnitude of the vector d^. In this work, however, this metric is not used directly. In order to exploit the multiresolution property of the zero-crossings wavelet representation, the elements of du are weighted by their relative significances Sg for the particular feature

154 g{x). The scalar metric (weighted distance) h\ is thus obtained

hiU = dudiag{s'^}\^ where u is the position of the singularity and the vector Sg is a vector of probabilities, the values of which are discussed in Section 4.4.3, and the vector is defined as follows

Sg — ..., Sg j

The most probable location I of the feature g{x) within the signal f{x) is then found by simply identifying the minimum of

I = arg |m m /iu| in which case

c = l - 4 n\~ is a normalized indication of the confidence of the match, where

n = [n^{Z 2 jg),—:n-^{Z2 ig)\ and n is the mean value of n over the available training samples, and

(vPf = J ^ \\Zojf{l + x)\- + \Z2 ,g{x)\^^ dx

It can be shown [73| that

so that the confidence c is a likelihood on the interval [0,1], within some variation in practice due to the fact that the magnitudes are integrated over a finite domain.

The confidence may be helpful in determining which of two features is more prob­ able if they are both detected at the same position.

155 4.4.3 Determining the value of the confidence vector

Although the determination of the vector Sg corresponding to the feature g from a set of examples may be a classical application for neural networks [152], the approach taken in this study is to obtain the vector's value by directly analyzing examples of

“ideal” features. The “ideal” features are the best model for a real signal, since they are real signals. The use of real signals as models is discussed in Section 4.3.

Let ..., M} represent M “ideal” example features centered at x=0.

Assuming that the distributions of the zero-crossings representations of these exam­ ples are Gaussian, the feature may be characterised in terms of mean and variance.

The mean zero-crossings representation is determined using

_ 1 ^2j 9 — ^2 ^2iQTn{^) m = l

A consistent estimator (defined in [73] for this case) for the square of the distance between the example features and their mean at each scale, which corresponds to the integral of the variance of the zero-crossings representation at each scale, is

CTy _ =

W hen d^iZojgm, Z2 jg) is determined from (4.8), then the value of 9 as defined in

(4.9) is redefined as

9 = a + a (4.10) where a denotes the half-width of the feature in the time-domain.The significance vector is determined simply from the estimate of the standard deviation as follows

4 = (4.11) ^Z.J9

156 In this way, those frequency bands in which the energy is the most constant over the various examples bear the most weight in the matching process.

The confidence vector expressed in this way is not a likelihood, but it may be modified to express the approximate likelihood that the feature falls within a standard deviation of the mean. If x is the feature position, and the symbol a denotes crz^g then

4 = P{x € [-o-,cr]} so that

= 2Erf{a) - 1 and this can be approximated as follows [18|:

< = '

4.4.4 Test results for brainstem auditory evoked potentials

The first 8 octaves (J = 8) of the wavelet transform were obtained as described in Section 4.4.1 for each trace. The quadratic spline wavelet of compact support was used. The effect of the stimulus artifact in the BAEP within the first 5 ms on the wavelet transform is rather significant. This also proved to be detrimental to the detection process. For this reason, the 8 octaves of the wavelet transform for each trace were all filtered before continuing to find the extrema. The filter used was a combination of a median and a moving average filter with a rectangular window, defined by

fP^{x) = ^mean[A/(x)] + ^median[I\f{x)]

157 1

3 g 4 5 6 7 g

01 2 3 4 5 6 7g 9 10 odlliaeconds

Figure 4.9: Eight octave wavelet transform for the normal BAEP after median-mean filtering.

where

A /(a:) = {f(x)\xe[x — Ax, x + Ax]} and Ax was chosen to be 31.25 /is (2 samples at a sampling rate of 64 kHz). .A.lthough the nonlinearity of the median filter makes it effective in removing outliers, it tends to flatten the extrema in the signal, at least in the discrete implementation. This makes it diflacult to identify the exact location of the extrema. This is the reason that a 10% moving average filter was included - it provides some definition to the signal with respect to extrema locations. Figure 4.9 shows the wavelet transform for the first example B.\EF after filtering.

The zero-crossings representation for each wavelet transform was obtained next, using equation 4.5 following the method of Section 4.4.1. The location of each feature

158 1

2

3 g 4 5 6 7

g

•5 -4 ■3 •2 •1 0 1 2 3 4 5 Qullûeconds

Figure 4.10: Mean feature template for peak III.

in the three examples was identified and the corresponding mean feature template was found as described in Section 4.4.2. Figure 4.10 shows the mean feature template for peak III. The variance of the traces was determined, and the confidence vector calculated using equation 4.11. The confidence vector for peak III, for example, was:

[1037 123.1 19.83 2.63 0.612 0.363 0.225 0.200|.

Finally, the distance function defined in Section 4.4.2 between the mean feature template and the zero-crossings representation of the wavelet transform of a test signal was found for each of the five features (peaks I to V). The distance function for peak

III in a previously unseen normal test BAEP trace is shown in Figure 4.11.

In order to assist the algorithm in the case of poorly defined traces, it was given some a priori information regarding the probability distribution of the peak latencies.

This is the information in Table 4.2.

159 _L ____ i _

0 234 5 6 7 81 9 10 millisecond»

Figure 4.11; Distance function h for peak III in the normal BAEP. A peak latency of 4 ms is identified.

Peak Mean 3 X std dev. I 1.61 2.15 II 2.75 4.5 III 3.80 4.46 IV 4.98 4.5 V 5.40 6.66 I - III 2.1 2.6 III - V 1.9 2.3 I - V 4.0 4.6

Table 4.2: Mean and standard deviation values in milliseconds for latencies of clini­ cally important BAEP components [CCF EP laboratory normative data with waves II and IV interpolated].

1 6 0 I------1 -I

IV -/H y \i / '\ / ' \ /1 V / — I—\J - A V — — 1 \ y 1 1 4- - - 1 - -4 4-

4- U -À 4- 1 1 1 _l J I _L 4 5 6 8 10 milliseconds

Figure 4.12: Normal clinically-obtaiiied BAEP with well-defined peaks I - V detected automatically using the zero-crossings wavelet method. Results are unprocessed.

Figures 4.12 and 4.13 show the peaks found in two test traces. The BAEP shown in Figure 4.12 is a normal well-defined example. The BAEP in Figure 4.13 has poorly defined peaks, particularly peaks II and IV. The algorithm correctly identifies all peaks in both cases. It was only necessary to provide the a priori information described above to correctly identify peaks II and IV of the second test signal. The other peaks could be identified without this information.

4.4.5 Tracking results for somatosensory evoked potentials

Pei et al. tracked latency and amplitude changes in evoked potentials by means of an adaptive least mean squares algorithm [107]. In the work of this dissertation, the SNR for the SEP trace from a single normal subject was artifically modified by

1 6 1 — — -f — — I— m

01 2 3 4 5 6 7 8 9 10 milliseconds

Figure 4.13: Normal BAEP with poorly-defined peaks I - V detected automatically using the zero-crossings wavelet method.

1 6 2 30.0 1 ' I T T f ■

275

O O O G O ' I 25.0

225 o o o o S' o o §20.0

17.5

15.0 2 3 4 2 3 4 2 Î 4 lO-i IQO IQl 102 Signal to noise latio

Figure 4.14: Detected feature position vs. SNR for SEP using algorithm with a priori information on expected feature positions. Diamonds denote detected N20 latencies. Boxes denote detected P25 latencies. Dotted lines indicate true positions of N20 and P25 features. When the SNR exceeds 2, detection becomes acceptable.

adding white Gaussian noise (after averaging) in the ratio of n times as much noise power as original signal power. Note that the signal to which the noise is added is the averaged SEP after 1000 samples. This signal already contains some noise, so that the SNR in this case is really lower than 1/n. This definition of SNR (or “signal and noise” to noise ratio) should be taken into account when interpreting Figure 4.14. .A. better evaluation of the algorithm in terms of mean square error in colored noise, but without tracking, is discussed in Section 4.4.6.

Figure 4.14 shows the detected feature positions for the N20 and P25 features plotted against the SNR. The a priori information in Table 4.3 relating to the ex­ pected positions of the N20 and P25 features was supplied. In the absence of this

163 Peak Mean 3 X std dev. N20 19.3 22.3 P25 25.0 22.3

Table 4.3: Mean and standard deviation values for latencies in milliseconds of clin­ ically important SEP components [CCF EP laboratory normative data with P25 extrapolated, normally expressed as 3 x std dev.j.

information, performance of the algorithm was comparable at SNR values exceeding

5, but deteriorated at lower SNR values.

To obtain tracking results, noise was added to the normal SEP waveform in the ratio of 3:1, after the entire waveform had been delayed by a period reflected on the x-axis of Figure 4.16. priori information was used by the algorithm to assist it in locating the features. This information was initially kept fixed. Once the features had moved out of the ‘"normal” region deflned by these a priori data, the tracking was lost. This occured at a delay of around 4 ms.

If the a priori information were continually updated with the most recent positions of the features found by the algorithm, successful tracking persisted for the entire range of delays (delays greater than 10 milliseconds were not tested). This is shown in Figure 4.16. For comparison. Figure 4.15 shows the results without tracking.

The clinical usefulness of this method of feature tracking depends on the assumption that suflSciently frequent measurements of the evoked potential waveform are made, and that the feature positions do not change abruptly over time. It is expected that changes in the neurological conditions that determine the positions of evoked potential features in patients in the NICU should be relatively gradual. Feature tracking was not included in the NICU system implementation of peak detection, mostly for lack of

164 35.0

30.0

I 27.5

g 25.0

20.0 f -

17.5 -o-

15.0 0 1 2 34 5 6 78 9 10 Delay (milliaecoiida)

Figure 4.15: Detected feature position vs. delay for artifically delayed noisy SEP using algorithm without tracking information. Diamonds denote detected N20 latencies. Boxes denote detected P25 latencies. Dotted lines indicate true positions of N20 and P25 features. The slight mismatch in positions even at high SNR is due to multiple peaks that could equally well be interpreted as the true peak.

165 35.0

30.0 TT

- —ch

20.0

15.0 0 1 2 3 4 5 6 7 8 9 10 Delay (milliseconda)

Figure 4.16: Detected feature position vs. delay for artifically delayed noisy SEP using algorithm with tracking information on expected feature positions. Diamonds denote detected N20 latencies. Boxes denote detected P25 latencies. Dotted lines indicate true positions of N20 and P25 features.

166 time. Implementing this feature is not as straightforward as in the previous example,

since the system may irregularly acquire EPs of the appropriate type. The system

has to determine when to apply tracking and when not - if a recent EP of the same

type was acquired, tracking may be applied.

4.4.6 Evaluation of the discrete detector

A discrete version of the wavelet peak detector is implemented in the NTCU moni­

toring system. Since peak detection is simply a tool to assist the researcher in finding

the peaks in many repeated EP tests, the purpose in this project was to at least

equal the capability that commercial systems provide. “Commercial systems” refers

to commercially available systems for intra-operative monitoring, since there are no

commercially available systems for NICU monitoring. Although the methods of peak

detection in these systems are proprietar}', personal observations of their behavior sug­ gest that they use a hill-climbing algorithm, and require frequent correction. For this reason wavelet peak detection is compared with simple hill-climbing. Hill-climbing is a simple algorithm in which the nearest signal maximum to the expected (mean) peak position is found.

Figure 4.17 shows the process by which the algorithms were evaluated. For a particular signal to noise power ratio (SNR), the grand average signal was scaled and added to the appropriate amount of colored noise generated by a white noise generator filtered with the AR filter described in Section 4.3. The resulting synthetic

EP signal had a fixed amount of energy and the specified SNR. An example of a set of such signals is shown in Figure 4.18.

167 Generate noise. (AR model)

Add to grand average to achieve specified SNR/power.

Discrete wavelet Hill- climb. detection.

Calculate MSE over detected peaks.

Figure 4.17: For each, value in a range of SNRs, a set of synthetic EP signals is generated using the EP model, and the peak latencies are found using hill-climbing and wavelet peak detection. The MSE is calculated as the average squared error of the latencies over every detected peak in each signal in the set.

168 Patient: Model Examples Test: BAEP (type 1 ) left (1000 cycles at 11.76 Hz)

Channel 1 0.001

A1 - C z 0.002 50HZ-1500Hz ^ 0.004 50uV (FS) 0.008

0.016

0.032

0.064

0.128

0.256

0.512

1.024

2.048

4.096

8.192

16.38

32.77

65.54

262.1

524.3

Time (ms) o.OO 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00 10.00

Figure 4.18: .A. set of synthetic left B.A.EP signals for the Al-Cz electrode pair with the SNR value for each trace marked on the right.

169 -2

-3

0 0.51 1.52 2.5 3 3.5 4 4.5 5

Figure 4.19: The first 5 milliseconds (r-axis) of a synthetic B.A.EP superimposed on its discrete wavelet representation.

These signals were supplied to both the hill-climbing and the wavelet detection

algorithms. An example of a signal superimposed on its discrete wavelet representa­

tion is given in Figure 4.19. The mean squared error (MSE) was calculated as the sum of the squared differences between the detected latencies for the ipsilateral waves

I, III and V' and the contralateral waves III and and their true values in the signal

without noise. This MSE was averaged over 25 synthetic; signals for each SXR. The

MSE was calculated in this manner for the two algorithms, and the results are shown

in Figure 4.20. It can be seen that the MSE decreases with increasing SXR in both cases, and the performance of the wavelet transform algorithm is superior for SXRs in excess of about 3. The hill-climbing procedure is very sensitive to local maxima

170 Hill climbing V Wavelet

UJ

10''

10'' 10 ' SNR

Figure 4.20: MSE vs SNR for hill-climbing and wavelet peak detection in the left BAEP. The MSE is calculated for each SNR value as the average squared error of the detected latencies for five peaks in each of 25 synthesized left BAEP signals.

171 Name of band flow fhigh. delta > 0.0 Hz* 4.0 Hz alpha 8.0 Hz 13.0 Hz beta 13.0 Hz 50.0 Hz

Table 4.4: EEG frequency band definitions. *fiow for the delta band is limited by the acquisition system's high-pass filter.

and since it starts at approximately the expected position of the peak, it rarely wan­ ders far before being caught in a maximum. This is the reason that it outperforms the wavelet method for low SNRs. The poor performance of hill-climbing for very high SNRs is surprising and may be attributed to the algorithm climbing the smooth, noise-free signal for some distance before settling on an erroneous peak.

4.5 Calculation of spectra and power in the electroencephalo­ gram 4.5.1 Spectral estimation by means of the periodogram

The powers in the delta, alpha and beta bands are calculated from the power spectral density PSD{f) of the EEG by integrating from ftow to fkigh as follows:

r /high P = PSD{f)df flow where flow to fhigh are defined for each frequency band in Table 4.4. The power spectral density of the EEG is estimated by means of the averaged periodogram

(squared Fourier spectrum), typically averaged over 6 epochs of 10 seconds each.

.Averaging reduces the variance of the spectral estimate in proportion to the number of averages, but increases the bias [60]. Muthuswamy et al. describe the drawbacks of the EFT for spectral estimation of the EEG and suggest that an adaptive AR

172 approach be taken instead [93]. The AR method of spectral estimation is discussed in Section 4.5.2.

The powers in the three bands are calculated for each of the eight electrodes that together cover the frontal, parietal and occipital regions of the brain. For each elec­ trode, a count of the number of bursts is performed. This is only relevant in patients exhibiting burst-suppression activity. The burst counting algorithm is described in

Section 4.6.

The importance of EEG spectral information is noted by Agarwal, who presents an automatic EEG analysis method that compresses long-term EEG information into a short representation for easy review [5|. Hoyer et al. also recognize the importance of EEG spectra and present a neural network EEG classifier for use during cerebral ischemia [53]. EEG spectral changes with sedation are routinely exploited in the bispectral index [115].

4.5.2 Spectral estimation by the covariance method

The covariance method is a means of estimating the parameters of the autoregres­ sive (AR) model of a signal. The autoregressive model is an all-pole filter model and as such models peaks in the spectrum. The work of Goel et al. [40] inspired this at­ tempt to apply the method to EEG data collected in the NICU. Goel et al. observed three well-defined peaks in the EEG spectra of hypoxic piglets and reported that the ratio of the power in the two high-frequency components (approximately theta and beta bands) increased while the power in the low-frequency component (delta band) decreased with ongoing hypoxia. Pardey reviews parameteric modeling techniques

173 for analyzing time-series such as the EEG, discussing adaptive and non-adaptive AR techniques and the choice of model order [105].

The first step in analyzing the EEG is to remove the baseline drift in accordance with the approach of Goel et al. This can be considered as a type of nonlinear low- pass filtering. In practice it made verj" little apparent difference to the signal in the time or frequency domain, but was included so as to accurately mimic the approach of Goel et al.

Baseline drift in the 60-second sample of EEG was removed by fitting a quadratic polynomial to the data. Polynomial fitting of arbitrary order is revisited in Section

5.2.1. In both cases, the least-squares fit is found using the algorithm described by

Mathews [79]. The process reduces to solving a set of linear simultaneous equations and this is easily done by matrix methods. The linear matrix equation is solved by

Gauss-Jordan elimination with full pivoting [109]. Evaluating the polynomial and subtracting from the EEG results in the baseline-filtered signal (see Figure 4.21).

Next, the EEG signal was low-pass filtered at 30 Hz to remove high-frequency noise.

This operation was performed directly in the frequency domain after Fourier trans­ formation. This was done by Goel et al. in a slightly different way. Preliminary results using EEG from human patients in the NICU showed that without 30 Hz low-pass filtering, the AR poles tended to fit the high-frequency noise above 30 Hz at the expense of the desired frequency components in the delta, theta and beta bands that were observed by Goel et al. in their piglets.

Finally the covariance method was used to obtain the AR parameters using Kay's implementation of the algorithm [60]. An example of a power spectral estimate and the pole diagram for a section of EEG from a comatose patient is shown in Figure 4.22.

174 X 10

time (s)

Figure 4.21: Sixty second epoch of EEG with quadratic polynomial fitted to reduce baseline drift.

175 Tm c P o w er l o e c n f M n s ty (typo l ) TWn#. Ffi JU 24 IS .40:33 1990

IMZ-IQOMZ SOuVCFS)

SOuV(FS)

SOuV (FS)

-1

Figure 4.22: Autoregressive power spectral estimates and pole diagrams for a ten second sample of EEG obtained from four electrodes in a normal patient.

1 7 6 The poles are determined using Laguerre’s method [109| for finding the complex roots of a polynomial, in this case the AR equation. The human EEG did not display the spectral features that Goel et al. observed [40], and a higher-order spectral estimate was necessary to resolve the detail of the EEG. Pursuing this avenue of exploration further was not considered to be within the scope of this dissertation. Although the code for calculating AR poles and zeros, generating these as lists of derived parameters and calculating the corresponding PSD (the method of Press et al. for calculating the PSD from the ARMA parameters using the EFT [109]) is included in the NICU monitoring system, it is not accessible using the GUI interface. In this work, non-adaptive methods are implemented for estimating the spectrum of short segements of EEG. Goto et al. describe an on-line adaptive AR spectral estimation algorithm for nonstationary time-series that implements order selection based on the

.A.IC with a forgetting factor. They apply this to the EEG [44]. Birch et al. state that the EFT and AR spectral estimator are not always successful for calculating the spectrum of short EEG segments. They suggest prewhitening the data to overcome these difficulties [15].

4.6 Burst detection and counting in the electroencephalogram 4.6.1 Burst detection in the time-domain

As in Section 4.5.1, the first step in burst detection by the method of Sherman et al. [129] is to remove the baseline drift. A moving window with a width of 700 ms is now passed across the 60 second sample of EEG in 100 ms steps and the power in the window is assessed at each step. The power is defined as the average sum of squares of the samples in the window. This is compared with the power in the 700

177 ms window immediately preceding it. If the energy has increased more than seven­ fold, the event is counted as an energy burst. Successive bursts must be separated by a non-burst interval of at least 100 ms. This algorithm was originally described by Sherman et al. [128|[129|. They point out that their definition of bursting is not entirely consistent with authors such as Clancy [27] who require that bursts arise from a quiescent background. An alternative method of burst detection is discussed in

Section 4.6.2. Muthuswamy et al. performed higher order spectral analysis of bursts, and detected indicators of nonlinearity in the EEG generators during bursting [93].

They note that bursting in neonates portends a grave outcome.

4.6.2 Matching pursuit

It seems intuitive that the burst detection and counting problem is best handled by means of a time-frequency representation. Although this approach is not implemented in the NICU monitoring system and is not pursued further in this project, a short discussion is included here for completeness. Bursts are somewhat similar to other structures that occur in the EEG. Durka et al. applied matching pursuit to sleep spindle detection in the EEG and comment that the time-frequency resolution of this method is ‘"finer... than any other method available at present” [36]. Wang et al. [151] implemented matching pursuit with a neural network and tested this on the electrogastrogram (EGG) and other nonstationary signals. They found that it had better joint time-frequency resolution than the Wigner distribution and holds promise for EGG and EEG analysis. Akay applied matching pursuit to the problem of separating muscle artifact (EMG) from the EEG but found it to be unsuccessful during tension due to the overlapping spectra [7].

178 The dyadic wavelet transformation used for peak detection and described in Sec­

tion 4.4 maps a time-domain signal into a time-frequency plane that is completely

tiled by an orthonormal set of basis functions. In some signals, however, particular

patterns arise that may be more appropriately expressed in terms of a different dic­

tionary of basis functions. In particular, functions of which the Fourier transforms

have narrow high frequency support are poorly expressed using the transform’s set

of basis functions [74]. Mallat and Zhang [74| compare this to the spoken word - a

message may be expressed more succinctly using a dictionary of highly descriptive

but redundant words as opposed to a small vocabulary in which a single concept may

have to be described using several words (“nocturnal mammal of southern Africa,

with a tubular snout and long extendible tongue” instead of “aardvark”). The noisy

burst in the EEG burst-suppression signal is precisely a structure that has narrow

high frequency support in the time-frequency plane.

In 1993, Mallat and Zhang introduced an algorithm called “matching pursuit” for

decomposing a signal into a linear expansion of waveforms that belong to a redundant

dictionary of functions [74|. The matching pursuit attempts to match the signal

structures in the dictionary to the signal in such a way as to express as much of

the energy in the original expression with as few time-frequency elements as possible.

When the dictionary contains a variety of atomic structures, this method is well suited

to nonstationary signals such as most biological signals.

Figure 4.23 shows the EEG from a patient exhibiting burst-suppression activity.

Figure 4.24 shows the time-frequency representation of a 30 second sample of EEG

from the same patient exhibiting burst-suppression. This figure was generated using some of the routines from a set of routines for MatLab called “WaveLab”, that is

179 Bed Aug 9 20:12:08 1995 RTSE_EEG dur: 30.00 sec.

F P 1 -F 7 2 i f u V I F7-T7 I T7-P7 I P7-01

Fp2-F8 I F8-T8 T8-P8 P8-02 Fpl-F 3 F3-C3 -VM n C3-P3 P3-01 Fp2-F4 F4-C4 C4-P4 P4-02 Fz-Cz Cz-Pz

Figure 4.23: Multichannel EEG recording in a patient exhibiting burst-suppression activity.

180 Phase plane: Burst-suppression 0 .0 5 1------1------1------1------1------1------1------r

= 0.02 I

Burst-suppression EEG "T

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 4.24: Time-frequency representation obtained by performing a matching pur­ suit on an 82 second sample of EEG containing burst-suppression activity recorded between the F7 and Pz electrodes. The pursuit matched 50 signal structures to describe 95% of the energy using the Coifiet wavepacket librar\'.

181 freely downloadable from the web-server of the Department of Statistics at Stanford

University [21]. It can be seen that tall, narrow structures match the bursts. In principle these structures could be characterized and counted to obtain the number of bursts in the EEG epoch. This particular avenue is not pursued further in this research.

4.7 Summary

The EEG signal is rich in information and extracting this information is an in­ teresting challenge. The approaches that have been considered in this chapter are only a few of many that might have been chosen, and those that were implemented extract only a fraction of the information that is present in the signal. Most of the algorithms were implemented in C and C+-r as part of the NICU system, and include a GUI front-end to facilitate data analysis and visualization of the results. The peak detection algorithm was first presented at the 34th Rocky Mountain Bioengineering

Symposium in Dayton in 1996 [143]. Other aspects of the data analysis were present­ ed in part at the 3rd International Workshop on Biosignal Interpretation in Chicago in 1999 [146]. At this workshop, the importance of the emerging field of biosignal interpretation, in which medical signals are analyzed to facilitate interpretation by clinicians, was emphasized. In this chapter the emphasis was on algorithms that fa­ cilitate interpretation of electrophysiological signals by clinicians untrained in clinical neurophysiology.

182 C H A PT E R 5

List and vector data analysis

5.1 Introduction

The work in this chapter is motivated by three requirements of monitoring in the

NICU, namely the need to combine multiple parameters [26l[57|[83][114], the need to track changes over time [48|[97][158| and the need to analyze the dynamics by means of promising non-traditional methods [11|[89|[118|[133|. A parameter list is a list of the values of several derived parameters and the times at which they were measured.

Necessarily, subsets of parameters are sampled at different times, because electro- physiological tests of different modalities cannot be measured simultaneously. In this dissertation, the word vector is often used to refer to a set of parameters of which the values var}' in time. For the value at any particular instant to be a true vector, the component parameters would have to be sampled simultaneously. Since this is not possible in practice, the uniformly sampled vector must be generated synthetically by interpolating the parameter list. This chapter describes various operations that are applied to the parameter lists, the process of converting from a parameter list to a vector, and a method for visualizing the dynamics of the vector. Two types of analysis that are of interest are parameter list extrapolation and variability analysis. Extrapo­ lation is provided as a simple investigational tool for identifying linear or higher-order

183 polynomial trends in the parameter lists [89]. The variability of physiological param­ eters, in particular cardiovascular signals, has been proven to be a useful diagnostic and prognostic tool [13|[95|. Since the process of interpolating and resampling the parameter lists will almost certainly alter the outcome of these analyses, they are not available to perform on vector data and must be applied directly to the individual components of the parameter lists. Variability analysis and extrapolations are done only on the list data, not the interpolated uniformly sampled vector. Although the same argument applies to dimensionality analysis, it is discussed in the section on vector analysis (Section 5.4) because the dimensionality of an individual parameter, the dimensionality of the combined parameters and the dimensionality of the phys­ iological state space are linked. The formation of a vector is clearly necessary if a state space diagram is to be generated since every point along the trajectory in state space has associated with it a time and the value of every parameter in the vector at that instant. Authors such as Bellamy and Molenaar promote the use of state space representations, dynamical analysis and complex systems theory for dealing with the complexity and uncertainty of physiological signals [ll|[89|.

5.2 List analysis

5.2.1 Extrapolation

The same approach is used for extrapolation as was used in Section 4.5.2 for remov­ ing baseline drift from the EEG signal before AR spectral estimation. A polynomial of a specified order n is fitted to the parameter list over a specified region (typically the entire list). The least-squares fit is found using the algorithm of Mathews [79], and the resulting linear matrix equation is solved using Gauss-Jordan elimination

184 Tir*: Thu S«p 1700:^000 199e

■^oAT rate 9 * 0 0 vîrar) 86.00 ' — —

bcHxyoa oo 1Z09 oo^'eoo coaso* oo-*»(* ouBrOC 01 1M9 gvaéoa ai ais-oo at *â-t» la-oo-od

Trfne- T tv iS e o i 7 0Ct0Q;09 1998

Banpounor et 52.00

b(70ft09 00:1M9 003*09 00:3e09 00:4a.-09 0100-09 01 1?^ 012*09 31 3609 Ot *àÔft Çg-QC od

Figure 5.1: Heart rate signal collected in the NICU from a stroke patient, together with cubic polynomial fit and extrapolation. The pohmomial was fitted to the first hour of the data and the second hour is a prediction obtained by extending the evaluated points of the polynomial.

with full pivoting [109|. For evaluation purposes, the polynomial may be fitted to a subregion of the list, and extrapolation beyond this region compared with the actual

physiological results that were collected. Figure 5.1 shows the results for a heart rate signal collected in a stroke patient. This is done in 7 in an attempt to see whether the technique has any predictive value.

5.2.2 Variability analysis

Several authors have recently reported on the diagnostic and prognostic utility of heart rate variability (HRV*) and relative alpha (RA) variability. .A. feature for evaluating variability is also included in the NICU monitoring system. Nagel et al. observe that the spectral components of the HRV correlate with the operation of the cardiac control system, a function of the CNS, and show that HRV has diagnostic

185 value in animals and humans [95]. Bianchi et al. used an AR model of the HRV

spectrum to discriminate between normal subjects, patients with multiple sclerosis

and patients with Parkinson’s disease [13]. Winchell observes that HRV decreases as

ICP increases and CPP decreases, and suggests that it may be a noninvasive early

indicator of possible intracranial pathology [153]. Jiang observed that while HRV

decreases in most congestive heart failure patients, it is significantly lower in those

patients who suffer a significant cardiac event or death within 18 months of testing

[56]. Lanza observes the same correlation, but adds that it does not add prognostic

information to what is to be expected using only the classical prognostic variables such

as ventricular arrythmias [67]. Vespa monitored patients after vasospasm following

SAH and observed that in a significant number of patients, a decrease in relative

alpha variability preceded vasospasm by a day or two [148].

In this work, the variability of a signal has been defined simply as its variance.

For a signal such as the heart rate, the entire sample could be evaluated, or regions

of the sample could be evaluated separately. The value of evaluating subregions is

that changes in the variability can be detected. The algorithm implemented in the

NICU monitoring system includes a parameter that specifies the window length. The

window is shifted across the data, and the data points within the window are used

to find the variance. This approach suffers from the filtering effects of rectangular windows - the action is a low-pass filtering action, but the filter is not rectangular

in the frequency domain, it is a sine function. In this research the purpose of the variability analysis is to find long-term changes in variability, and so this problem has been neglected. Figure 5.2 shows the results of a moving window variability analysis of a heart rate signal collected in a stroke patient. Goldberger describes how

186 r«Tw: ThuS«0 iroCKlI 4919E

borqv*a Oft^X-30 00^5 Tt 30-36:52 00-**33 ______Qî-QQ-f* 31 11 SS ______01 2336______01 3&17______01 4a.5fl 01 58:3

Figure 5.2: Results of a variability analysis with a window length of 20 samples performed on the heart rate signal from a stroke patient.

a decrease in heart rate variability accompanies pathologt" and age [421. and describes the loss of complexity associated with illness in other processes such as respiration and the EEG [41|.

In this work, it would be interesting to compare the variability of the heart rate in a normal subject with that of a comatose patient. Furthermore, it would be interesting to monitor the heart rate variability over a length of time in a recovering patient to determine whether the variability increases and thus whether it has prognostic value.

This is why the system allows the analysis of the overall variability of a signal as well as a local windowed variability. It has been shown that the EEG alpha band energ\- also exhibits the property of greater variability under healthy conditions [411. Some results are given in Chapter 7.

5.3 Interpolating a list to form a uniformly sampled vector

The purpose of interpolating parameter lists in the XICU monitoring system is to obtain a uniformly and simultaneously sampled set of data points that can be easily represented in the form of a state space diagram [95[. Since interpolation and

187 Tbtm- U on Orne 7 l O a t S l

9 a £ P ro&uiM i II

NggO-SI !7 1*40 l»5gQ 7______30-A9-S6 ?1 43 45 ______23 37 3* 00-2S t? GT 19C1|

Figure 5.3: BAEP wave I latency data set interpolated by cubic splines. BAEPs were recorded in a normal subject. The jagged trace in each panel represents the original set of latencies and the smooth trace represents the set of interpolated points.

smoothing have the potential to remove important information from the data, the ex­ trapolation and variability analyses are performed on the data before interpolation.

Dimensionality analysis is also performed on the uninterpolated data, but the discus­ sion on dimensionality (Section 5.4.2) is reserved for after the state space diagram description as it is more appropriate to discuss it in this context.

In this research, it is the evoked potential latencies that are typically interpolated, and Figure 5.3 shows an example of such an interpolated data set. The cubic spline algorithm [109| is used for interpolation. This algorithm operates on three samples at a time. A cubic polynomial is fitted to the three data points such that it fits them exactly, and satisfies the requirement that the second derivative be continuous at the boundaries of the spline.

188 5-4 Vector analysis 5.4.1 State space representations and attractors

Molenaar provides a tutorial overview of multivariate nonlinear time-series anal­ ysis in which the central role of state space modeling is emphasized [89]. Bellamy et al. suggest that fuzzy analysis and a state space approach for diagnosis and prognosis is the way to deal with the complexity and uncertainty of mutiparameter diagnos­ tic information [11]. Lehmann estimated the complexity of the mapped brain field distribution in state space and found that it correlated with modes of spontaneous and induced mentation [69]. The NICU monitoring system includes a feature for generating state space diagrams.

Cubic spline interpolation facilitates the construction of a state space diagram.

Every point on the trajectory in a state space diagram represents the value of the state vector at a particular time. This implies that a value must be available for ever}' component of the state vector at that time. Furthermore, if the points are uniformly and closely spaced in time, this makes for a smoother trajectory through state space. Interpolation addresses these two requirements. State space diagrams are used to study system dynamics. In this work, the idea is to demonstrate changes in a patient’s condition as the value of the state vector shifts from one region of state space to another. If regions of state space associated with normality or patholog}" could be identified, the state space diagram would help the clinician to visualize the impending shift in condition, or quickly note the current condition. Of course, the state space diagram shows only three of the many available dimensions, so a lot of information is lost in this representation. With a judicious choice of axes, however, critical information could perhaps be conveyed with only three dimensions. If other

189 information is needed too, a conventional vector plot must be used. WTiile Figure

3.15 shows a state space divided into normal and pathological reasons, it would be especially interesting to observe a shift from one region to another. An attempt is made in Chapter 7.

In the case of a patient for whom illness results in death, it is expected that cortical functioning should be lost first with an associated decrease in the average frequency and energ}' of the cortical EEG energy. Next the SEP is expected to exhibit shifts in latencies of the significant peaks, and finally the BAEP is expected to show latency shifts and loss of amplitude. A definite and predictable trajectory of the state through its space is anticipated. It is of course not possible to represent the entire space graphically, but some subspace could be represented, as was done in Figure 3.15 for the patient with posterior subarachnoid hemorrhage. Nakao et al. observed transitions from REM to slow wave sleep in a space of cardiovascular parameters [97|.

The difference between state space and phase space should be noted at this point.

State space is simply the space containing the totality of possible state vector values.

Phase space is the space of a single variable and its derivatives up to the nth derivative.

While Figure 3.15 is a state space diagram. Figure 5.4 shows two phase space diagrams

(portraits).

When the trajectory of a variable in phase space over all time occupies a finite re­ gion of space, the set of points in the space occupied by the trajectory is the attractor.

Figure 5.4 shows the attractors for the burst and suppression phases of the EEG (left and right panels respectively) and tell something of the EEG dynamics under condi­ tions of pentobarbital-induced coma.The left panel shows the phase portrait during 2

190 Figure 5.4: Phase portraits of EEG burst-suppressiou activity during induced coma. The burst phase is shown on the left and the suppression phase on the right.

seconds of suppression and the right panel shows the phase portrait during 2 seconds of burst activity. The EEG amplitude is represented on the three axes with lags of

0, 0.1 and 0.2 seconds. Whilst activity is minimal during the period of suppression, the portrait from the burst period shows an attractor which suggest low-dimensional periodic activity. This is consistent with the observation that pathological conditions result in a decrease in complexity of the dynamics of functional measures such as heart rate and EEG [41|. Dimensionality is discussed in Section 5.4.2.

5.4.2 Dimensionality of multiparameter data

The dynamics of the EEG display the property that they lose complexity under unhealthy conditions. For example, it has been shown that the complexity of the

EEG is higher in a normal awake person than when the person is asleep, comatose or in a state of epileptic seizure. This is if complexity is quantified in terms of the embedding dimension of the EEG which is normally very high. Arguably it is not

191 meaningful to specify the embedding dimension under these conditions. Under ab­ normal conditions, the embedding dimension has been reported to fall as low as 4 to 6 [120]. Embedding dimension relates to the underljdng number of degrees of freedom of the mechanism which gives rise to the dynamics of the observed variable.

This corresponds to the dimensionality of the phase space needed to fully describe the variable’s dynamics. For a single variable, the vector in phase space consists of that variable and the appropriate number of its derivatives. In early work, the

Grassberger-Procaccia algorithm was used to detect chaos in the EEC, but more re­ cently authors such as Theiler and Pritchard have demonstrated using surrogate data that the assessment of low-dimensionality was due to autocorrelation in the oversam­ pled EEC signal rather than low-dimensional chaos and that correlation dimension is not good for discriminating cognitive activity [lll|[140j. However, Yaylali et al. were able to identify different types of seziures based on the correlation dimension cal­ culated with the Grassberger-Procaccia algorithm [156] and other authors also favor stochastic-complexity measures over traditional amplitude and frequency information for analyzing physiological conditions such as Cheyne-Stokes respiration, anesthesia and sleep [118].

Apart from embedding dimension, phase space portraits and Lyapunov exponents are used to study system dynamics. The Lyapunov exponents describe the rate at which two points which are initially close in phase space diverge in time. Positive

Lyapunov exponents denote an unstable system and negative exponents denote a stable system. Under conditions of visual stimulation, it appears that the alpha activity generated in the occipital cortex becomes entrained and simpler, with an

192 accompanying increase in negativity of the Lyapunov exponents [147|. This suggests

that the system is stabilized by the entraining effect of the visual stimulation.

The preceding discussion of dimensionality has considered only the case of single­ variable data. If the data contains multiple variables, like that obtained in this research, it is far more difficult to estimate the dimensionality of the entire system.

The difficulty lies in the fact that there may be a correlation between observed vari­ able, that is some of the degrees of freedom underlying the observed variables may be common to more than one variable. Therefore a correlation and dimensionality analysis combined would be necessary to estimate the overall dimensionality of the system. This is beyond the scope of this work, but may be interesting for future work.

It would perhaps be useful in the sense that it would provide some means of obtain­ ing the most parsimonious representation of the data and therefore help in automatic analysis of the data. Automatic analysis would imply identifying the current state

(pathological or not, and possibly finer divisions of the state space) and predicting an impending change in the state. This sort of analysis is considered by Abarbanel [Ij.

5.5 Summary

Automatic analysis facilitates interpretation of electrophysiological data by clini­ cians unskilled in neurophysiology. Chapter 4 addressed this need by providing basic methods of analyzing the raw electrophysiological data collected in the NICÜ. This chapter took this further in an attempt to provide a higher level of visualization of the data, and possibly some predictive ability, by supplying a few tools for analyzing the dynamics of the observed system. With state space trajectory representations, it may be possible to have a computer identify normal and abnormal regions of the

193 space and indicate the region of the space that the patient vector occupies. It may be possible by extrapolation to make a prediction of how the vector values may change in future with the purpose of triggering intervention if necessary. By combining the parameters, early subtle changes may be detected that would otherwise go unnoticed if the parameters were observed individually. This work on neurophysiological state dynamics was presented at the Second International Conference on Complex System- s in Nashua, New Hampshire, in October 1998 and subsequently published in the

Inter Journal of Complex Systems [145].

194 CHAPTER 6

Monitoring protocols for experiments

6.1 Introduction

Preliminary experiments, readings and discussions with physicians led to the standard monitoring protocols described in this chapter. Tlu'se protocols are applied as consistently as possible in the two studies of Chapters 7 and S. Some variation is inevitable in practice, since the patients have different disorders, access to the patients varies and unanticipated technical difficulties arise that influence how the protocols are applied. The protocols consist of the acquisition and stimulation [tararneters for the evoked potentials, the montages for the electrophysiological tests, the accpiisition parameters and choice of physiological signals that are recorded and the scheduling of the tests. Peripheral to the actual monitoring are protocols related to setting up the equipment for monitoring - applying and testing electrodes consistently, record­ ing the stimulation parameters that are selected on the equipment front panel and recording any changes to the protocol necessitated by the individual condition of the patient. For example, in some patients the presence of a craniostomy for monitoring intracranial pressure means that ICP can be recorded along with tlu' other physio­ logical parameters, but also has the effect of disturbing the electrode arrangement

195 because the pressure transducer’s access through the skull obscures one or more of the electrode positions defined in the standard arrangement.

6.2 Standard test definitions

There is one standard test definition for each modality, except the somatosensory' evoked potential which has two standard test types. Test types 1 and 2 for the

SEP form a single test definition in a sense, however, because they are symmetrical with respect to the side of the test i.e. SEP test type 1 is to be used for left SEPs and SEP test type 2 is to be used for right SEPs. This facilitates interpretation of the raw waveforms by allowing the more conventional labeling EPi and EPc etc.

(Erb’s point ipsilateral and contralateral to stimulation respectively), as shown in

Figure 6.2, rather than EPl and EP2. Table 6.1 shows the values of the acquisition and stimulation parameters for the evoked potential and electroencephalogram tests excluding the montages. The standard montages for the four electrophysiological modalities are listed in Table 6.2 and shown diagrammatically along with the timing diagrams in Figures 6.1 to 6.4. Table 6.3 lists the standard physiological parameters sampled at the bedside.

6.3 Standard schedules

The monitoring system automatically generates schedules in the standard form used in this research, given the modalities that are to be included in the schedule. The procedure is described in Section 3.2.1. For example for a patient who is obtunded and might feel pain, SEPs are not included in the protocol. In this case, BAEPs,

EEGs and physiological parameters are recorded. In comatose patients, SEPs are

196 Parameter BAEP type 1 SEP type 1 SEP type 2 VEP type 1 EEG type 1 Masking noise on - - - - No. of channels 2 4 4 4 8 Sample rate 50 kHz 10 kHz 10 kHz 2 kHz 250Hz No. of samples 512 512 512 512 2048 No. of cycles 1500 500 500 100 4 Interstim. period 85 ms 369 ms 369 ms 474 ms 15 s Stim. pulse durât. 100 us 100 us 100 us 10 ms - Artifact rejection tr.-reject tr.-reject tr.-reject tr.-reject - Threshold type dynamic dynamic dynamic dynamic - Reject level 2.0 X m.p. 2.0 X m.p. 2.0 X m.p. 2.0 X m.p. - Discard, samples 62 62 62 0 - Notch filter off off off off off Ref. electrode STD STD STD STD STD

Table 6.1: Standard values for the test definition parameters for the evoked potential and electroencephalogram tests.

For each sttmuius:

. 240 us

- 100 us

10.24 m s {512 samples)

For œmplete test:

JL I n

127 s (1500 slim, cycles)

Figure 6.1: Standard montage and timing diagram for BAEP test type 1.

197 Modality Act. no. Act. name Ref. no. Ref. name Low freq. High freq. Sens. BAEP 23 A1 18 Cz 50 1500 50 (type 1) 24 A2 18 Cz 50 1500 50 SEP 5 CPi 29 EPi 50 3000 50 (type 1) 6 CPc 33 EPc 50 3000 50 6 CPc 5 CPi 50 3000 50 29 EPi 33 EPc 50 3000 50 SEP 6 CPi 33 EPi 50 3000 50 (type 2) 5 CPc 29 EPc 50 3000 50 5 CPc 6 CPi 50 3000 50 33 EPi 29 EPc 50 3000 50 VEP 9 01 17 ME 1 70 50 (type 1) 10 02 17 MF 1 70 50 20 Oz 17 MF 1 70 50 19 Pz 17 MF 1 70 50 EEG 3 F3 18 Cz 1 70 50 (type I) 5 CP3 18 Cz 1 70 50 9 01 18 Cz 1 70 50 20 Oz 18 Cz 1 70 50 17 Fz 18 Cz 1 70 50 4 F4 18 Cz 1 70 50 6 CP4 18 Cz 1 70 50 10 02 18 Cz 1 70 50

Table 6.2: Standard montages for the electrophysiological tests.

Group Param eter EGG heart rate Blood pressure 1 mean systolic diastolic Blood pressure 2 mean S P 02 saturated oxygen pressure Non-invasive blood pressure mean systolic diastolic

Table 6.3: Standard selection of physiological parameters measured at the bedside for physiological test type 1.

198 For each silmulus:

1200 us

100 us

stim.

sam .

51.2 ms (512 samples}

For complete test:

i i J____ L

184 s (500 stim. cydes}

Figure 6.2: Standard montages and timing diagrams for SEP test types 1 and 2.

For each stimulus:

■ 6000 us

- J u — 10000 us - J ï i

256.0 ms (512 samples)

For complete test:

! r I____ L stim. sam.

47 s (too Stim. cycles)

Figure 6.3: Standard montage and timing diagram for VEP test type 1.

199 For each stimulus:

F p l

8192.0 ms (2048 samples)

For complete te s t

J L stim. IT sam. 30 s <3 stim. cycles)

Figure 6.4: Standard montage and timing diagram for EEG test type 1.

included, and the schedule illustrated in Figure 6.5 is automatically generated by the

system. The system interleaves the tests and sides as much as possible to minimize

any fatigue or habituation effects - that is it spreads tests of the same type as evenly

and far apart as possible within the monitoring period, given that: an evoked potential

test must be performed every 5 minutes; EEG tests must be performed in 1-minute

epochs in the gaps between evoked potential tests; physiological tests can overlap

with electrophysiological tests while electrophysiological tests of different modalities

may not overlap.

6.4 Standard forms

Several standard forms were developed for recording pertinent information neces­ sary to interpret the results of the studies in Ghapters 7 and 8. The information on the forms is identical to the information on the web-based interface. Although the

2 0 0 M' ■■ ' r " ~i- ■■ ■' ' ■■ r :

■- , , - ■ - g- M

I I B : I l I I I I I I » ■ I i ■ I I I y B II ■ T r : t i n t i r

Figure 6.5: Standard one-hour schedule incorporating BAEP. SEP. EEG and physi­ ological tests.

system is designed such that paperv\*ork is unnecessary-, the paper forms are some­

times convenient. For example, there is no browser available inside the INR and it is

convenient to note the stimulator settings during the procedure on the paper forms.

The information can be transfered to the patient data files later by means of the web-

based interface. The Patient History Forms shown in Figure 6.6 are relevant for both

the XICU study and the IXR study. The Patient History Forms include information

about preexisting conditions and medications that may give rise to peculiarities in

the EEG or EPs that are not related to the current illness or procedure, as well as

information about the current condition of the patient. Settings on the front panel of

the stimulating equipment are recorded on the Patient History Forms. Any deviations from the normal protocol are also noted on these forms.

It was originally the intention that changes in the monitored parameters in the

XICU be correlated with transcranial doppler (TCD) results, computed tomography

(CT) results and several parameters routinely monitored in the XICU (Glascow coma

2 0 1 CLEVELAND CUNIC FOUNDATION CLEVELAND CUNIC FOUNDATION MEP Monttoring Study: lotttol Infcrm ttoo MEP MonHodnq Study: Inttiai Infom tstioo Padmtt lnfcnrmba i

Rm momhomgdMemd

------i

‘O lher. lAdJtanÊleamwnts

i BAEP wgiului type \ BAEPyuWw* amplitude ■ BA EP timkiWK «mpfatade

Figure 6.6: Patient History Forms.

2 0 2 CLEVELAND CUNIC FOUNDATION MEP Monftortng Study: Row Chart Parameters Patient name | T est date and tim e j ReaMts Neuxio^cat examination Eye opening Verbal Best motor response Total Glascow coma scale Pupil reaction left Pupil reaction right Slectratytei&els Liver function ALT AST billtiubin Renal funaion BUN creatinine Anehal blood gases pH p c a PO. IHCO.I Sc O f sat. Ami* epileptic drug levels phcnytoin pentobarbital Other patameter\a/ues Core temperature | Onset of diabetes insipidus 1 Onset of inappropriate diuretic j hormone secretion ,

Figure 6.7: Flow Chart Form

scale, blood gas levels etc.). For completeness, these forms are included in Figures

6.7, 6.8 and 6.9 respectively. Unfortunately it was beyond the scope of this study to include these data in the research. This was for reasons of limited available assistance and a low number of appropriate patients. Technically the system supports the use of the information. It is hoped that the design presented here may be implemented in a future studv of this sort.

203 CLEVELAND CUNIC FOUNDATION 1 MEP Monitoring Study: CT Results 1 Patient name | | Tea date and time | { R esults Volume of mass I .Mkllinc shift 1 Septum pcHuctdum | mm I 1 Pineal body | mm 1 Uncal herniation j no t yes i Obliteration of basal cisterns ! “o ! yes 1 Hydrocephalus ) no 1 yes I

Figure 6.8: CT Form.

CLEVELAND CLINIC FOUNDATION MEP Monitoring Study: Transcranial Doppler Results Patient name Test date and time Reaàts Artery V dodty Putaatlllty Index Basilar artery U f i ACA

Left PCA Right A C A ! Right M CA I Right PCA

Figure 6.9: TCD Form.

204 6.5 Standard procedures

The following is a general outline of the procedure for monitoring a patient. Fur­ ther details specific to the studies are given in Chapters 7 and 8:

1. Identify appropriate patient from schedule in INR and the recommendations of

the clinicians and fellows in the NICU.

2. Obtain background information on the patient from the hospital medical data

base (Phamis Last Word).

3. Transfer relevant information to Patient History Form and enter by means of

web-interface for permanent storage along with patient monitoring results.

4. For INR patients, notify the patient of suitability for the study, explain the

purpose of the research, and obtain patient consent. The CCF Investigational

Review Board (IRE) application included the approved consent form.

5. Schedule technician to apply electrodes 30 minutes to 1 hour before the INR

procedure is scheduled to start.

6. Set up program of tests to monitor, appropriate to the patient.

7. Apply EEG electrodes, stimulating electrodes and earphones to the patient in

the NICU or in a separate examination room close to the INR suite. Test

the impedance of the electrodes - no electrode impedance should exceed lOkfi.

Record a 1-minute epoch of EEG and check for noisy electrodes. Re-gel elec­

trodes as appropriate. Record a single complete BAEP and SEP to check the

earphones and stimulating electrodes. Reposition as appropriate.

205 8. Start system acquiring data according to the chosen program of tests at the

beginning of the procedure in the INR suite or during the period of possible

changes in patient condition in the NICU. Annotate relevant events. Do not

report results to clinical staff or patient while monitoring.

9. Remove electrodes and stimulating equipment once monitoring is complete.

10. -A.nalyze data (see Section 6.6).

11. Archive data and analysis results to magnetooptical disk.

6.6 Standard analysis

The following is a general outline of the procedure for analyzing the data obtained from a patient. Further details specific to the studies are given in Chapters 7 and 8:

Raw analysis:

1. Calculate derived parameters i.e. find the peaks in the EPs, calculate the spectra

and energy in the bands of the EEG and count EEG bursts if relevant.

2. Check the EP peaks by hand and correct any that may be incorrect.

Vector analysis:

1. Create a parameter list or parameter lists relevant to the study and particular

patient.

2. Create a smooth vector from each parameter list.

3. Plot the parameter list(s)/vector(s).

206 Further analysis:

1. Calculate the variability of the heart rate.

2. Calculate the dimensionality of the heart rate.

3. Calculate the variability of the EEG on various channels.

4. Calculate the dimensionality of the EEG on various channels.

5. If there’s an interesting region, extrapolate from the region before the change

to the region past the change.

6. Create a phase space portrait of the interesting region.

The results are presented and interpreted in Chapters 7 and 8.

6.7 Summary

The standard protocols for conducting the tests for the experimental research part of this dissertation were outlined in this chapter. The procedures for recruiting patients, recording relevant background information, setting up the monitoring e- quipment and conducting the tests were outlined. The general approach to analyzing the data was discussed. More detailed information related to the specific experi­ ments is given in Chapters 7 and 8. Although a uniform protocol was defined for use throughout the experiments, the experiments are preliminary and intend to show proof-of-concept rather than statistically significant findings. For this reason there was quite some variation between patients in terms of medical condition, and conse­ quently also in terms of how uniformly the protocols could be applied. A review of the monitoring system and its intended use in the NICU research study was presented

207 in a poster session at the 1998 Annual Meeting of the Biomedical Engineering Society

in Cleveland, Ohio in October 1998.

208 CHAPTER 7

Neurointensive care unit study

7.1 Introduction

The NICU monitoring system was designed to continuously monitor the central nervous system of the NICU patient at the bedside. The purpose of the NICU study described in this chapter is to verify the correct operation of the system, identify technical and logistical problems and if possible verify the prognostic utility of this type of monitoring. Section 7.2 describes the research protocol designed to this end.

Section 7.3 previews the scope of results obtained and summarizes the circumstances for each patient. Results of specific analyses are discussed by analysis tv-pe in more detail in Section 7.4.

7.2 Research protocol

The patient monitoring protocol has been described in general in Chapter 6. This section describes some of the details pertinent specifically to the NICU study.

7.2.1 Patient selection criteria

Patients in whom changes were anticipated in the electrophysiological parameters over a period of time during their NICU stay were included in the study. In particular,

209 those patients in whom a mass effect or edema was present were chosen because the associated physical shifts in the brain tissue could potentially give rise to pressure on or physical deformation of the brain stem and other structures that are monitored by the NICU monitoring system.

The following categories of patients were included in the study:

1. Pentobarbital coma induced after arteriovenous malformation (AVM) excision.

2. Large hemispheric infarction with edema.

3. Cerebellar infarction with edema.

4. Intracerebral hemorrhage with CT evidence of mass effect.

Patients should be in a decreased state of conciousness - lethargic, obtunded or co­ matose.

7.2.2 Parameters recorded initially

The patient history was recorded on the patient history form (Figure 6.6). This form includes such information as the patient’s date of birth, sex, any preexisting conditions such as deafness that might confound the electrophysiological results and the diagnosis at the time of monitoring.

7.2.3 Monitored parameters

The standard protocol for the NICU study was to monitor the following parame­ ters:

1. BAEP (initially record stimulus and masking noise amplitudes and always use

rarefying click stimulation).

2 1 0 2. SEP (initially record stimulus current on each side).

3. EEG (initially record any changes to electrode placement necessitated by skull

defects and other monitoring equipment).

4. Blood pressure (mean arterial, central venous).

5. Blood oxygen saturation (Sa02).

6. Intracranial pressure (if available).

7. Heart rate.

8. Breathing rate.

9. Annotations describing events that may affect the interpretation of the moni­

toring results.

The circumstances of each patient necessitated exceptions to the protocol that made its uniform application impossible. Every patient was monitored in a different way.

Although this made it impossible to uniformly analyze the results or draw statistical conclusions from the data, it did exercise the various facilities of the monitoring system.

7.2.4 Information recorded after monitoring

The patient outcome was determined and recorded a few weeks after monitoring.

7.3 Overview of patient data

Data were collected from one normal subject (Normal), an initial test patient (PI) and 6 true cases (EP, BL, JT, HV, AH and HK). The range of patients is summarized

2 1 1 in Table 7.1. Although the protocol was designed for an ambitious and comprehensive

study involving a large number of patients, the outcome was modest. In particular,

the incidence of appropriate patients presenting in the NICU was much smaller than

anticipated, and when they did present, it was not necessarily at a time conducive

to convenient monitoring. Given the budgetary limitations of this project, it was

unreasonable to expect a technician to be available at any given time to assist with

placement of the monitoring electrodes. The result is that the data presented in the

rest of this chapter are anecdotal.

Normal

A 49-year old female subject in good health with no known neurological disorders

was recruited on 3/3/98 to obtain representative EP and EEG results. Figure 7.1

shows the normal BAEP for the left and right sides. Figure 7.2 shows normal EEG

results on all eight channels. The philosophy of this approach to monitoring has been

that the patient is his or her own baseline - it is changes that are of more interest

than absolute values. So this normal subject is not intended as a baseline, but rather

as a test to check that the equipment generates normal EPs and EEGs for a normal

subject. Indeed, Figures 7.1 and 7.2 show that this is the case.

PI

P i was the first patient, primarily monitored for the purposes of testing the system

in the NICU. A number of small issues had to be resolved, such as obtaining network access at the bedside in the NICU, positioning the equipment at the convenience of other personnel in the small NICU cubicle, organizing for technicians to place the recording and stimulating electrodes, connection to the Tram bedside monitor and

2 1 2 Name DOB* Monitoring Mod.** Condition Electro. Outc. dates results Norm. 6/16/50 3/3/98 BSE Normal Normal - PI 3/26/98: S Basilar artery infarc- 15:20-19:45 t. E P 3/5/42 4/16/98: BSEP Subarachnoid hem­ Lt III&V Rehab. 12:13-13:50 orrhage, basilar tip del. Rt aneurysm clipped N20&P25 4/15/98, rt side. del. BL 10/20/31 8/26/98: BSEP Lt hemiparesis, car­ See dis­ Died 11:00-19:00 diac arrest, rt mid­ cussion. 9/3/98 22:00-0:00 dle and anterior cere­ 8/27/98: bral artery infarc- 0:00-10:00 t. Expected tissue 13:00-16:30 shifts/herniation but 18:30-0:00 no hemicraniectomy 8/28/98: (medical instability). 0:00-2:30 10:00-17:00 JT 2/17/54 9/16/98: SP Large rt middle cere­ SEP Died 17:00-0:00 bral infarct following noisy. 10/7/98 9/17/98: cerebrovascular acci­ 0:00-8:00 dent. 15:20-0:00 9/18/98: 0:00-4:15 HV 5/31/25 12/7/98: BSEP Rt MCA infarct and Lt N20 Nursing 16:38-0:00 secondary ICH. &P25 facility 12/8/98: abs. 0:00-1:30 AH 10/4/14 3/24/99: BEP Rt PICA infarct Rt III&V Acute 16:00-17:30 and hemorrhage, BS abs. rehab. comp. HK 3/24/58 4/11/99: BSEP Hypoxic en­ Lt III Nursing 16:00-18:15 cephalopathy fol­ &V abs. home lowing period of Lt &rt unconsciousness. N20&P25 abs.

Table 7.1: Summary of monitored patients (one normal and seven patholog­ ical). *DOB = date of birth. **Mod. = modality (B=BAEP, S=SEP, E=EEG,P=physiological). Lt = left, rt = right, del. = delayed, abs. = absent.

213 Figure 7.1: Normal left and right BAEPs.

£ £ 0 ;ry>» T tnan««rva* t o lS c y a M « l0 0 7 n r t 5 0 0 »l £ E O (7)p» 1| i w i C f O # t of a eyOM 4: 0 C7 r> TuekUf 3l6 33%6tMB T u , War 3 i« 33 M IMS

.*2 83 -- DO^0 ? 0 4 0 6 OB t o 0 3 Qi 0 6 o a

Figure 7.2: Normal 8-channel EEG.

214 verifying that the system functioned correctly in this environment. The electrical

noise level in the NICU was not found to be excessive, although this is an electrically

hostile environment. Pi had suffered an infarction in the territory of the basilar

artery. After monitoring, some changes were made to the protocol, particularly the

manner in which EEG epochs are recorded (longer epochs). The system was also

secured more strongly to the cart, longer stimulating electrode leads were used and a

set of gender changes were made to allow these electrodes to be used.

EP

EP was a right-handed male of 56 years at the time of monitoring. He presented

at CCF on 4/14/98 with the worst headache of his life. An angiogram revealed a large

right-sided basilar tip aneurysm and CT scan revealed a large mass in the intrape-

duncular cistern. On 4/15/98 he underwent an operation in which the aneurysm was

clipped. Postoperatively, he was deeply unresponsive with left third (oculomotor)

nerve palsy worsening over the next few days. A CT scan showed bilateral thalamic

injuries.

This patient was considered to be a good candidate for monitoring. He was in a

comatose state, his illness was progressing and his thalamic injury might be reflected

in the BAEP and SEP results. This type of patient could potentially benefit from

the monitoring system because subtle changes could trigger intervention if necessary.

Unfortunately it was only possible to monitor the patient for an hour and a half.

Monitoring is sometimes interrupted by the need for another diagnostic test such as a CT scan in a different room. However, this was the first comprehensive test of multimodality monitoring of a real patient - the BAEP, SEP, EEG and physiological

215 Te»t 9ACP ;typ« Il MtasoaeyOMi Tatt BAEPnyp* ur^iiSOOevOMal it 78H4 ) r« n e TH< Aor 14 2 t tz 19M 4 «3 >«8 ran* Tïnâert* 14C3T3190a *n». *.»•<•

KKC.tSOOHz

OOP 1 0 0 z o o 3 0 0 4 00 SOB ftOC 700 « 0 0 3 0 0 lO fl T « » ; l » | i'o O C 1 0 0 3 0 0 3 0 0 4 0 0 VoC 4 0 0 7 00 4 0 0 *130 "COq

Figure 7.3: Left and right BAEP results for the patient "EP". Waves III and V are delayed bilaterally as a consequence of his thalamic injuries. In each case, the results of two tests are superimposed to verify repeatability.

parameters were recorded on both sides. The results showed delayed B.AEP waves III and V on the left and right sides and delayed SEP X20 and P25 peaks in response to stimulation on the right side. The B.A.EP results are shown in Figure 7.3. These results are consistent with the patient's disorder. The abnormal B.A.EP cluster in the state space diagram of Figure 3.15 was generated from the data of this patient.

In the two weeks following monitoring, the patient was operated to evacuate an epidural clot and treated medically for repeated fevers and communicating hydro­ cephalus. After a stay in the neurological StepDown unit, the patient was discharged on 6 22 98 in stable condition, awake and with the ability to weakly move his ex­ tremities. He was transferred to a rehabilitation unit for intensive therapy.

BL

BL was a female of 66 years with a history of cardiomyopathy at the time of monitoring. She fell at home and presented at CCF on 8/24/98 mth left paralysis

216 and facial droop on the right side. En route to CCF she suffered a cardiac arrest and was resuscitated. An initial CT scan revealed a right middle cerebral artery (MCA) infarction. After an emergency laparotomy to remove air from under the diaphragm, a repeat CT scan showed that the MCA infarction had extended and now involved the anterior cerebral artery. She exhibited left V'llth (facial) nerv^e palsy and left hemiparesis on the neurological examination.

This patient was an ideal candidate for monitoring. She had suffered an acute stroke and had the possibility of swelling with tissue shifts and impending herniation.

Neurosurgery evaluated the patient for hemicraniectomy in anticipation of swelling, but determined that she was not a suitable candidate on account of her general medical instability. Monitoring began on 8/26/98 and was intermittent for three days. During this time the patient’s condition worsened despite aggressive medical therapy.

The BAEP, SEP, EEC and physiological parameters were recorded in five sessions.

Figure 7.4 shows the schedule of recorded tests. The sessions lasted 8, 12, 3.5, 7 and 7 hours, giving a total of 37.5 hours of recording. Some physiological tests were skipped due to difficulties with the bedside monitoring interface, and this was resolved later.

The major difficulty experienced with this extended recording was that when the author was not present, the BAEP earphones would dislodge from the patient’s ears and the SEP electrodes would become detached. Nursing personnel were not inclined to rectify the situation without prompting. In later monitoring, smaller earphone inserts and tape were used to keep the earphones in place, and SEP stimulating electrodes with longer leads were used so that they would be less easily detached by

217 M If i| i| 11| i| i| i| i| t| t| i| t| t| i| i| i| h hfth i-hhlnthhph t thhPhhlî 'h itPMMMMhh hhhhhMPMP «rphhhMhr »hhMh

|i |i|i|. M i|'M’11|'11| 1'I'I i|'I <|T’ I 'I 'I ’I ’I ’I ’I 'I 'I '1111 ’I ’I I ri ’I 'I 'I 'I 'I I I ’I ’I ’I ’111 'I 'I 'I 'I r

i ’ I ’

I I II 11 I I4'I’I'MI'I444’I’I44'I'MI'MI44’MM 11'I’I'14 4'I’14’I’I'Ml'I’I’I'I'l'I'I’I’ |i|i|i|i|i|i|i|i|i|i|i|i|i|i|i|i|i|i|i|i|i|i ||i|i|i|i|i|i|i|i|'|i|i|i|i|i|i| 4'Ml'MM

MMI'I'MI'MM'T I'l'l'MI MMI'MMI4 4'14'1'14'I'I'I'M I'l'i'l'l'l'l'l'MMI Ml4 I I I'I'l'l'f I *44444444: 144444444444444444444^?^^ 114444444444444444444?:

Figure 7.4: Schedule of recorded tests for BL. The schedule begins on 8/26/98 and extends over three davs.

218 SEP(iVP«2 ngMiS00 SEPdVMancn(SaacvaM4ll7t H

tsa o a 00 300 aooo 3oo «00 « so o so o o JOO 500 1000 ISOO 3 00 300 1000 3 00 4000 tfX UOO

Figure 7.5: Stacked SEP array obtained in response to stimulationof the right median nerve from BL during the morning of 8/27 98. She herniated at approximately 3 a.m. Left stack shows EP and subcortical P14 and X18 responses. Right stack shows cortical X20 and P25 components. Time scale on the y-axis is 13:15 to 7:19.

pulling on the leads. Xo annotations were recorded during the monitoring of this patient.

Of particular interest in this patient are the events of the first morning after record­ ing began. During this 12 hour recording session, the patient herniated at around

3:00 (on 8/27/98). Unfortunately the author was not present and the left SEP stimu­ lating electrode had become detached and the earphones dislodged. Xevertheless. an interesting stacked SEP record was obtained in response to stimulation on the right, and this is shown in Figure 7.5. The most interesting result would probably have been the response to SEP stimulation on the left. It is in just such a situation that

219 the NICU monitoring system could be of benefit, provided that the electrophysiolog- ical parameters or combination of monitored parameters contain the information to predict the impending herniation. This hypothesis has not been proved in this study.

The volume of data collected during this study emphasized that the review sub­ system was impractically slow with very large volumes of data. This situation was improved later by implementing more of the graphical preprocessing in C rather than

Tcl/Tk and altering the way in which data is recalled from the saved record.

The patient’s condition deteriorated until 9/3/98 at which time her examination was consistent with brain death and ventilatory support was discontinued.

JT

JT was a 44-year old man at the time of monitoring. He had a history of alcohol and intravenous drug abuse with corresponding renal and coronary artery problems.

He was transferred to CCF on 9/14/98 after suffering a stroke. A CT scan showed a right middle cerebral artery infarction.

BAEP changes in this patient were not anticipated and only the left SEP and physiological parameters were recorded. Only the left side was monitored because the patient was awake and stuporous and would perhaps have felt pain on the right side. The left side was paralyzed. Two monitoring sessions of approximately 15 hours and 13 hours occurred on the night of 9/16/98 and 9/17/98 respectively. SEP results were noisy and inconclusive, and the patient tried to pull off the electrodes. This patient provided a lesson in the importance of competent electrode placement. In this case, collodion glue had mistakenly been used instead of conductive gel and as

2 2 0 a result the impedance of the patient-eiectrode contact was far too high to detect meaningful signals.

The patient had Candida endocarditis and on 10/7/98 he become febrile and hy­ potensive with evidence of renal failure. On account of his poor prognosis his family made him a “Do not resuscitate” case. His condition continued to decline until he was pronounced dead several hours later.

HV

HV was a female of 73 years at the time of monitoring. On 11/24/98, she expe­ rienced a right MCA infarct while driving, causing her to run three red trafiBc lights and hit the curb and another car. Left hemiplegia was observed by ambulance per­ sonnel. HV claimed to have remained conscious although she was unaware of what damage was done to her vehicle and claimed no injury upon admission to CCF. .A.n initial CT scan revealed no bleeding or mass in the brain. Later in the evening, her pupils became unreactive and asymmetrical in size. A CT scan revealed a ver}' large intracranial hemorrhage (ICH) in the temporal area with mass effect (midline shift) and she was brought to the NICU.

HV was chosen as a candidate for continuous BAEP, SEP, EEC and physiologi­ cal parameter monitoring due to her comatose state and the likelihood of changing abnormalities in these parameters. She was monitored for a period of 9 hours on

12/7/98 following surgical evacuation of the ICH. .Mthough no interesting changes were observed, the SEP showed absent N20 and P25 peaks in response to stimulation on the left side, consistent with her condition (see Figures 7.6 and 7.7). Figure 5.3 that shows cubic spline interpolation was generated using results from this patient.

2 2 1 Patient: Test: SEP (type 1) left (500 cycles at 2.71 Hz ) Time: Mon Dec 7 20:04:07 1S98 Mon Dec 7 20:24:07 1998

Z83 9.7 ip.6£ 17.1 1 1 1 ; : ;

Ctiannel 1 i/v CPi - EPi / 1 50HZ-3000HZ SOuV (PS) \ / 1 I ^ :

4. N18 \ ^ 1 1 . ^ ; -2.83 2.83 17.15

Channel 2 CPc - EPc 50HZ-3000HZ SOuV (FS)

-Z83 Z83 9.5 23.3

Channel 3 CPc - CPi 50HZ-3000HZ 50uV (FS)

P25 -2.83 2.83

Channel 4 EPi - EPc 50HZ-3000HZ 50uV (FS)

-2.83 L Time (ms) | Q.QO 5.00 lo!oO ls!oO 20.00 2s!oo 3o!oO 35.00 4o!o0 45!oO SO.od

Figure 7.6: Two overlaid SEP tests results obtained from HV in response to stimu­ lation of the left median nerve. The X20 and P25 peaks in the cortical (CPc-CPi) response are absent, consistent with her ICH and left hemiplegia.

2 2 2 Patient T est SEP (type 2) ngtit (5(M cycles at 2.71 Hz ) Time: Mon Dec 7 20:34:07 1998 Mon Dec 7 20:54:06 1998

3.83 10.0S ^3.9‘ 17.2j

Channel 1 CPi - EPi 50HZ-3000HZ SOuV (FS)

13.9: 17.2

Channel 2 CPc - EPc 50HZ-3000HZ SOuV (FS)

-3.83

Channel 3 CPc - CPi 50HZ-3000HZ SOuV (FS)

Channel 4 EPi - EPc S0HZ-3000HZ SOuV (FS)

-3.83

Time (ms) I 0.00 S.QO 10.00 1s!oO 20.00 25!o0 3o!oO 3s!oO 4o!o0 4s!oO SO.Qtj

Figure 7.7: Two overlaid SEP tests results obtained from HV in response to stimula­ tion on the right. All components are present and none delayed.

223 HV's condition improved somewhat until she was discharged on 12/15 in a con­ scious but neurologically impaired state to a nursing facility for further therapy.

AH

AH was a right-handed male of 84 years with a history of coronary artery disease at the time of monitoring. He was admitted to CCF on the afternoon of 3/21/99.

That morning he had experienced nausea and vomiting and had collapsed in church.

After eliminating myocardial infarction as a possible diagnosis, he was examined neu­ rologically. This initial examination did not reveal any significant abnormalities. By

3/22/99, however, an MRI revealed a right PICA and right superior cerebellar artery distribution stroke and right cerebellar tonsil herniation. On 3/23/99 his conditioned deteriorated drastically and a CT showed that the cerebellar herniation had worsened.

The patient was intubated and two procedures were undertaken in the operating room

- a right ventriculostomy and a suboccipital craniectomy. The results were good and he improved significantly during the following day. It was on this day (3/24/99) that he was monitored for a period of an hour and a half to see whether the BAEPs correlated with his condition.

Figure 7.8 shows the series of BAEP results for the left and right sides respectively.

Figure 7.9 shows the physiological parameter list for this patient. The results were pleasing in that the right side showed absent waves HI and V consistent with the condition of the patient. Also the annotations subsystem (separate keyboard) was exercised and worked appropriately, for example the annotation in Figure 7.9 referring to patient restlessness matches perturbations in the results.

224 Ta« t)ManaOOe«cW4l II 7«MX] Tm t SA£Pffir9*i>Mi

r« # Q«EP-Tvoa T tncniTSC O ooH M n SA£P I] r^O U 0M a»4l M n H

:

_J

Figure 7.8: Left and right BAEP results for AH. The right waves III and V are absent, but the system automatically marks spurious peaks in an attempt to find the closest match. The absence of waves III and V is consistent with right cerebellar stroke and herniation.

225 1989

v «jc

BA£P ipMuani

1«R 8A £P ip iiau fat

5CG -nftvtrat* pA‘r'^A/'vV 'V

AMonoen - Rtp.

SPQ2 - Sat 33 c

Figure 7.9; A physiological parameter list plot for .AH. The left BAEP components are undelayed, while the right components are absent, but marked spuriously on the plot. The cardiac parameters are fairly constant throughout the recording, except at 17:15 when patient restlessness gives rise to a disturbance in the measurements.

226 AH was extubated the following day and his ventriculostomy (ICP monitor) dis­ continued the next. Several days later, he was awake and alert. He was discharged on 3/31/99.

HK

HK was a female of 41 years at the time of monitoring. She had suffered a cardiac arrest on 3/19/99 and was found after an unknown period of time and resuscitated by emergency personnel. She presented at CCF wdth dilated pupils and other evidence of hypoxic encephalopathy. An initial EEC showed diffuse slowing, and this is reflected in the EEC spectral information shown in Figure 7.13.

The patient was monitored for 2.25 hours on 4/11/99. At this time she was sitting upright, obtunded and did not follow commands. The BAEP, SEP, EEC and physiological parameters were monitored. Figure 7.10 shows the BAEP results on the left and right, and Figures 7.11 and 7.12 show the left and right SEPs. Figure 7.14 shows the physiological parameter list plot for the brainstem, and Figure 7.15 shows the same for the somatosensory tests. Note the absence of waves III and V on the left and bilateral absence of N20 and P25, consistent with hypoxic encephalopathy.

HK was discharged in a stable but obtunded state to a nursing home on 4/13/99.

7.4 Additional analysis results

Due to the fact that none of the patients underwent significant changes during the monitoring period, applying the parameter list analysis techniques is not partic­ ularly appropriate. Neverthless, this section shows how these features of the NICU monitoring system are applied to the data.

227 Tm i BA£PitvM n w a t u o c n a » * » t< 7«rui SAS>nvpai)M i(tnoeKM 4(ti 7*wn

7«M aA£PV30CTOa»4> 1< Tt M. T t M a * E ? nfÇm 1> r^ft ilSOO iT>OM W II 7* M

» aoo »O0 JQO 103 400 ÎOO 400 700

Figure 7.10: Left and right BAEP stacks for HK. Left waves III and V are absent. This patient had hypoxic encephalopathy.

228 SEP (NM n M iSM cycm » ZTi h a SCPttyM t)Ml(SOOevOM4(£7< HA

2SOO %0O SCO 4000 «00 sooo TSoo a o o a o o m o o s «

T«ac SEP live» li Wfl i5COciOm*l2 7t MO 7m - SEPlTfOa n MTISOOCVOM 4 (2 71

MH2-J000»U

SOO 1000 1SÛ0 a o o a o o MOO a o o 40 00 «5C0 s o »

Figure 7.11: Left SEP stacks for HK. The cortical (CPc-CPi) X20 and P25 compo­ nents are absent, consistent with severe hypoxic encephalopathy.

229 T«t sCPcivp»a"grciSMev(M«r2.7iMs S£PffiteB2>n9«(SOOcvaM4C17l KR

JOOO a 30 3000

SEP,Tv»»a"gM<*0cW»«lZT1 MR SEP (fw a nqM rOQ cv^*» *: Z 7: Mg

2000 aoo 3000 3500 4000 4S00 WOO ■SCO 9 0 0 « 00 30 00 S 00

Figure 7.12: Right SEP stacks for HK. The cortical (CPc-CPi) N20 and P25 compo­ nents are absent, consistent with severe hypoxic encephalopathy.

230 n«jAOr T T&C703 1999

2 z « n «7 ( y «1 Oaa

CP3 «rwgyn

CP4«n«ryyin

CP4 «nwgy *1 ucra X

- z #n#i?y 'rt ;

P:«n#fgY4i

!@ 31 -27______16*139 16 55 51 ______1 7-Q em ______17-gai5 *7 33 27 7 6661 iaC9i=)|

Figure 7.13: Parameter list showing the EEC energ}' in the delta, alpha and beta bands at the midline occipital, left and right parietal and midline frontal electrodes. The energ}' in the alpha and beta bands is severely attenuated at all electrodes, with only the beta activity containing significant energ}^ This low frequency EEG is consistent with diffuse slowing.

231 TN iA pr ^ IdT loaO

Left BAEP > cK *ie«i

Left SAEP tpariaMm

Left BAEP «W W M I

«^gnBAEP

Figure 7.14; Parameter list showing brainstem related parameters in HK. The left waves III and V are absent. This patient has hypoxic encephalopathy.

232 THjAp» t 19»)

Lam UMncy

^ EP-^tfl UMncr

LafTEP-fCCiMmrey

_«n £P-j>2S uuncy

R«grt £ P - P 1 * « » n c y

EP-Nia lAHncv

Piçn tatancv

^QTt t?-P 25 UMncv

Figure 7.15: Parameter list showing somatosensor\' related parameters in HK. The X20 and P25 components were absent bilaterally, consistent with severe hypoxic en­ cephalopathy.

233 Tkn#: WMMw 24 16:38:21 1909

■7 ,' Jl TV A •'- ~xr- - ¥ - ^ 7 ''

P63821 164414 ______16:5 0 07 ______165600 17^1 53 ______1707 46 17-13:38______1 7 1 » 3 2 ______17-2S-2S i 7 3 l 18 T7Q1 3

Figure 7.16: Heart rate, respiration rate and mean arterial blood pressure plots for AH.

7.4.1 Variability analysis

A variability analysis with a window length of 10 minutes (60 samples spaced

10 seconds apart) was applied to the heart rate, respiration rate and mean arterial

blood pressure recorded on AH. The original traces are shown in Figure 7.16 and

the variability results are shown in Figure 7.17. These results show that the moving average variance is rather sensitive to outlying points in the recorded data and that

the window is perhaps too short to eliminate short-term changes due to inactivity or

restlessness of the patient. It is unfortunate that a patient with significant neurological changes that may have been reflected in the variability of the physiological parameters

was not available for this studv.

234 *#dW ^2*16.4Z21 1999

Itnarem et naan sa oc

ha-taai 1* 4*15 17 27 2 7 17-32-;3

Figure 7.17: Heart rate variability, respiration rate variability and mean arterial blood pressure variability plots for AH. The variability of physiological parameters may be expected to decrease in critically-ill patients [20|.

7.4.2 Extrapolation analysis

A second order polynomial fit to the BAEP latency traces shown in Figure 7.18 was performed and the polynomial extrapolated for an additional 2 hours. The ex­ trapolated polynomial is shown in Figure 7.19. The BAEPs are those recorded in

HK, who was in a stable condition at the time of the recording and the BAEPs were not expected to change after monitoring. The extrapolation predicts that some of the latencies should increase and others decrease. The extent and direction of the changes make their likelihood of actually occuring very low. A first order fit may be more appropriate and only very short-term forecasts may be reasonable. Without data reflecting true changes in neurological condition, it is difficult to determine the value

235 rm : ThuAor 1

ipaAuam

La(t3A£pio«ut

i^SACP ip iÉ ifm i I uiianev

hamso T622-19

Figure 7.18: Left and right BAEP wave I. Ill and \' latency plots for HK.

of this type of extrapolation for predictive purposes, the appropriate extrapolation parameters and the extent of the the prediction that is possible.

7.4.3 State space diagrams

Figure 3.15 earlier in this dissertation showed the state space of normal and ab­ normal BAEP components. At that point, it was explained that the state space could be divided into regions of normality and abnormality. In this section an attempt is made to show how the state may make a transition from one region to another. The attempt is shown in Figure 7.20. This diagram was constructed from the data of BL.

236 T>»iA©r 1 l& I M O V

Einrapeuaon ef L«R BAEP ipAiMrai I UMncy

E m o o u o o n of PigPt BAEP iQ»faf. ill laMncy

Excaoouoon of Piqpt BAEP VUioncv

Figure 7.19: Two hour extrapolation of left and right BAEP wave I. Ill and \' latency plots for HK. Polynomial is fit to data between 16:10 and 18:05. Extrapolation is from 18:05 to 19:57.

237 in whom a meaningful recording of a change was not actually obtained, hence the inconclusive diagram.

The purpose of such an image is threefold. The first is instructive. This diagram helps to visualize the concept of physiological state, the dynamics of the physiological state vector, the idea of regions of state space and the concept of prediction arising from an understanding of the state space dynamics. The second purpose is as a possibly useful and simple display for nursing personnel in the NICU. The diagram is easy to interpret at a glance, in the sense of the region occupied as well as the direction that the vector is going. Thirdly, these same concepts are easily translated into a computer algorithm for automatically determining the current state of the patient and possibly making a prediction as to the patient’s future state. This could be used to assist the clinician in making a prognosis, or sounding an alarm to catalyze early intervention if necessary.

7.5 Summary

The NICU monitoring system was tested on 8 subjects. During testing, it un­ derwent several refinements. The result is a robust monitoring system that monitors multimodality electrophysiological and physiological parameters in the hostile envi­ ronment of the NICU. The monitored signals represent a range of functions of the brain and can reflect several types of abnormalities, as shown by the data in this chapter. Many of the features of the system were exercised and shown to be func­ tional and perhaps diagnostic, if not prognostic. The predictive value of the system remains to be proved. Technical problems were mostly related to electrodes becoming

238 Figure 7.20: State space diagram of BAEP components in which an attempt is made to show a transition from one region of state space to another. The data are from BL before, during and after the period of herniation. The lighter colored region corresponds to the period after herniation.

239 detached and stimulators dislodging, and several administrative problems were en­

countered. The original mission of this research - to design and test a multimodality

monitoring system for use in the NICU - has been accomplished. .A. comprehensive

protocol was designed for the NICU study, and to take the study to completion given

the rate at which appropriate patients are available would either take a long time or would require additional patients (such as head trauma patients) to be included in the protocol. The groundwork has been laid for such a study, but its complete execution requires time and funding beyond the resources of this dissertation project.

.-At this point, for several reasons, mostly the scarcity of appropriate patients, it was decided to continue the medical investigation of the system in the interventional neuroradiolog}' (INR) suite. Here patients are scheduled in advance, they undergo short procedures during which known and controlled changes to the CNS occur and these may be reflected in the electrophysiological signals. .Although the NICT system was not designed specifically for this environment, the INR provides a good setting to test the hpothesis that appropriate analysis of the multimodality data yields pre­ dictive power early in the course of changes in the patient's condition. Despite the scarcity of experimental subjects, it should be noted that the system potentially has wide application. The constraints in this study were the protocol and experimental methods, not the applicability of the system.

240 CHAPTER 8

Interventional neuroradiology study

8.1 Introduction

In Chapter 7, the NICU system was tested on a number of patients and a nor­ mal volunteer. It was shown that it could reliably acquire and analyze the BAEP,

SEP and EEG along with the physiological variables routinely collected in the NICU.

However, it was not shown that the system could detect early subtle changes in these patients and thus trigger intervention at an early reversible stage. This would be a truly valuable property of such a monitoring system. Due to the scarcity of appro­ priate NICU patients, it was decided to monitor patients in the INR suite. Young et al. [159] state that EEG and EP monitoring may be helpful during INR procedures for cerebrovascular abnormalities. Patients typically undergo an INR procedure to correct a cerebrovascular abnormality. For example, during balloon angiography a balloon is inflated in an artery such as the vertebral artery that may have become occluded due to cerebrovascular disease. The balloon widens the artery so that the brain tissue in its territory may once again be properly supplied with oxygenated blood. The interest in applying the NICU monitor in this situation is that the pre­ viously ischemic tissue should be reperfused after the balloon has been removed and this should be reflected in the electrophysiology of that region of the brain. All of

241 this occurs in a short period of time, and the exact timing is known. In contrast, patients in the NICU undergo changes unpredictably and over a longer period of time.

INR procedures are done under local or general anesthetic and last one hour to a few hours. Because the patient has elected to undergo the INR procedure and no other study of this sort was being done in neuroradiology, it was necessary to prepare a doc­ ument for review by the Investigational Review Board (IRB) of CCF. This included a patient consent form. Obtaining IRB approval was a somewhat time-consuming undertaking, and approval had been granted only a short time prior to the prepara­ tion of this dissertation. For this reason, only the results of a single patient along with a control study are presented. The patient underwent a test balloon occlusion.

Quinonez states that EEG during endovascular procedures allows early identification of impending neurological deficits before irreversible damage occurs [113].

8.2 Research protocol

Apart from the general aspects of the protocol that are discussed in Chapter 6, some specific steps in the INR study are listed below:

1. Identify appropriate patient. Possible patients are identified from the schedule

of patients by the INR scheduler and the investigator (the author) informed of

their names and patient numbers.

2. Appropriate patients are chosen in consultation with Dr. Burgess and Dr.

Krieger of Neurology.

3. Patients are from the categories: balloon angiogram and balloon angiogram

with stent placement. These are the most appropriate, but patients undergoing

242 carotid, vertebral or basilar artery procedures in which significant changes in

the cerebral or brainstem blood fiow are expected would also be of interest.

4. For the chosen patient, select modalities to monitor - if the patient is awake

with local anesthetic only, monitor EEGs and possibly BAEPs. If the patient

is anesthetized generally, include SEPs also. The main considerations here are

patient comfort and interference with the procedure.

5. Counsel the patient on the research and obtain consent if the patient wishes to

participate.

6. Arrange for the patient to come early for electrode placement on the day of the

INR procedure.

7. Monitor intensively and annotate liberally, with annotation such as “balloon

inflated/deflated'’, “neuro. exam”, “movement”, “BP value” and “occlusion”. The

test schedule may be tailored to the specific patient.

8. Frequently, TCD is used to monitor blood flow in the vessel of interest during

the procedure. TCD reflects blood flow velocity with good time resolution, but

does not reflect neurological function. The NICU monitoring system pro\ddes

this function, and the two types of monitoring are therefore complementary.

9. Remove electrodes just after suturing and before anesthesiology removes their

monitoring equipment, making sure that the patient’s head is supported to

minimize bruising.

243 Name DOB* Mon. dates Mod. Condition Electro. Outc. results

Norm. 6/2/70 7/19/99: BSVEP Normal Normal - 15:45 - 17:45 WE 10/4/39 7/14/99: E Nasopharyngeal Muscle Disch. 14:00 - 16:00 tum or artifact, EEG spectral changes.

Table 8.1: Summary of monitored patients (one normal and one pathological). *DOB = date of birth. **Mod. = modality (B=BAEP, S=SEP, E=EEG,P=physiological).

8.3 Overview of patient results

Data were collected from one normal subject (Normal), and one true INR patient

(WE). The range of patients is summarized in Table 8.1.

Normal

A 29-year old male subject with no known neurological abnormalities was recruited to obtain baseline stacked BAEP, SEP, VEP and EEG data. Baseline arterial blood pressure and heart rate monitoring was not possible because of the invasive nature of the probes. Figures 8.1 and 8.2 show the SEP responses to stimulation on the left and right respectively. These are unremarkable. Spectral analyses were performed on the EEG samples and these are shown for the frontal EEG (left and right) in Figure

8.3, the parietal EEG (left and right) in Figure 8.4 and the occipital EEG (midline electrode only) in Figure 8.5. The information contained in these stacks is better conveyed in terms of the energy in the significant bands. Figure 8.6 shows a plot of the parietal EEG energy and Figure 8.7 shows the occipital and frontal EEG energy.

244 It is interesting to note how the energies are distributed across the head. For example there is more alpha band energy in the occiptal region, as expected. Also, as the test progresses, the subject becomes more restful and sleeps toward the end (for the last

20 minutes, and for a short period 10 minutes before this). At this time, the relative level of delta activity increases across all of the electrodes.

WE

WE was a 59-year old male at the time of monitoring. W^E had an occluded right internal carotid artery and minimal (<30%) stenosis of the left internal carotid artery.

His reason for having the test balloon ooclusion (TBO) was the presence of a recurrent solitarj' fibrous tumor in the left eustachian tube. The purpose of the TBO was to occlude the left carotid artery for an extended period to test the effect of sacrificing the left carotid arten; if the tumor was resected. A balloon was inserted by means of a catheter into the left carotid artery. This was inflated for an extended period of time, during which the patient underwent neurological examinations. Simultaneously, blood flow in the carotid artery was monitored by transcranial doppler (TCD), and the flow velocity had decreased significantly. Since the right carotid artery was occluded, this was expected to decrease the blood supply to the brain significantly. Interestingly, the neurological examination showed very subtle, if any, changes in the neurological condition of the patient, indicating that the vertebral blood supply to the brain was sufficient to keep it functioning normally. Later in the procedure, the anesthetists medically decreased the patient’s blood pressure from around 100 mmHg to just over half of this. With the balloon still inflated, the patient continued to perform almost as well on the neurological examination. The balloon was deflated after this

245 r«« SCP(T|>i»niiA(900efOW4lZ7iH SEPavo«n>iia(9oocvow«£nKA

t«m t£P>rvo* tM«.»ooKMBji2r< SEPTfO# t| «wt%Qc*om«*ZTi

r« r» i(m i 0QÛ500 ?0 00 ‘500 20 00 2S00 WOO 36 00 «300 COO WOO

Figure 8.1: Stacked SEP response to left median nerve stimulation in the normal volunteer.

246 r«M sEP«vp*a'^(SQ0cv

SEP (tytM a ngn isoo n a m ^ 2 7t n a SEP fTW# 21 rv« '5C0 OfdM^ZTTha

I

Figure 8.2: Stacked SEP response to right median nerve stimulation in the normal volunteer.

247 Figure 8.3: Frontal EEG spectral stacks recorded for the normal volunteer. The low frequency activity increases in the last 20 minutes when the subject sleeps.

3.00 *0 00 «300 MOO a 00 3000 UOO «O0O 4000 MOO o o e to o o a 00 3000 3000 4000 4100 MOO

Figure 8.4: Parietal EEG spectral stacks recorded for the normal volunteer.

248 T«SC

Figure 8.5: Occipital EEG spectral stacks recorded for the normal volunteer. The alpha activity that predominates during the first half of the recording is replaced in the last 20 minutes by delta activity as the subject sleeps.

249 Mon Jl4 IS :637ae ISSS

CP3 «rw^y ■'Mfa

It6372i 712-3^

Figure 8.6: Pariétal EEG energy in the alpha, beta and delta bands on the left and right sides for the normal volunteer. The beta activity is minimal throughout most of the recording as the subject relaxes with closed eyes. Delta energ}' increases significantly in the last 20 minutes as the subject falls asleep.

250 Tima: Mtan Jki 19 l& S rZ S 199S

Oz «nwyv m aaiu

OxarwgyinMona

1000X0

i-r «rwgy >n xçf-*

-£*n«9yini

i-aszW

Figure 8.7; Occipital and frontal EEG energ}' in the alpha, beta and delta bands along the midline for the normal volunteer. Low frequency activity increases toward the end of the recording, when the subject sleeps.

251 and the procedure ended. Figure 8.8 is the angiogram showing blood flow in the left carotid artery before the balloon was inflated and Figure 8.9 is the angiogram showing the blood supply to the brain via the vertebral artery while the balloon was inflated. The latter also explains how the brain continued to function well despite the occluded carotid arteries - the vertebral arteries alone provided adequate cerebral circulation. Both figures illustrate the extent to which the electrodes interfere with visibility during the INR procedure. Although the interference appears to be minimal in these films, it is the choice of the surgeon whether to tolerate electrodes during the procedure.

During the procedure, the NICU monitoring system monitored a total of 65 e- pochs of EEG lasting 60 seconds each and spaced apart by a gap of 12 seconds. Since the system was originally designed for the ICU, it acquires data in the ADC mem­ ory, and then transfers it to the computer for storage - it cannot acquire EEG data continuously. This is the reason that the 12 second gap between epochs of EEG is necessary. The EEG did not appear to change much during the balloon inflation. This is consistent with the fact that EEG reflects brain function, and as the neurological examination showed, function was minimally impaired. TCD measures flow velocity in a particular vessel, and this had necessarily decreased. The EEG was unfortunately rather corrupted by muscle artifact. The patient had a lot of muscle tension, perhaps associated with his smoking. Since muscle artifact occupies the higher frequencies and the range of interest in this study is the lower frequencies, a spectral analysis and band energ}' analysis was done to see whether this yielded significant results.

The results are shown in Figures 8.10 to 8.13. Note the increase in delta activity

252 Figure 8.8: Angiogram showing the left carotid artery circulation in WE before bal­ loon inflation. Note the outlines of seven of the electrodes and their leads belonging to the NICU monitoring system.

253 Figure 8.9: Angiogram showing the vertebral artery circulation in WE during inflation of the balloon. The outline of the electrodes is present along with the outline of the transcranial doppler probe. The patient’s almost unaffected performance on the neurological examination and the limited elect rophysiological changes during inflation of the balloon may be attributed to the good vertebral circulation shown in this figure.

254 Ijpvsl.: «0aTft49î999 «awi ^8^ oce.

2400.00 >200.00

Eta* 0

Figure 8.10: Energ\- in the alpha, beta and delta bands of the left and right frontal EEG in WE. Delta activity increases throughout the procedure and beta activity is attenuated slightly at the time of occlusion.

during the procedure and attenuation of frontal beta approximately coincident with the balloon occlusion on several of the channels.

8.4 Summary

As in Chapter 7, the results of this chapter show that the NICU system operates as designed. A series of normal results were obtained in a healthy volunteer and this baseline serves to indicate what can be expected of the EEG signal during a record­ ing in an awake, restful subject drifting to sleep at times. .A. series of BAEPs and

SEPs were also recorded in the normal subject. A single INR patient was recruited

255 iirr" VMC JU i4 C O ;:a :4 S i9 9 S J IT -&W'^1:* #

J i y - ...-...... - - - ...... /

C P 3 «n«fvy n « e r a

CP3«ra*yymMtt 1200 00

A / V / V/ ' \; ■ A . ^ \ V 3 0 0

* 0 X 0 0 w y i n n

J400 0C CM

CP* «ratyy *n ooa

Figure 8.11: Energ}- in the alpha, beta and delta bands of the left and right parietal EEG in WE. Delta activity increases throughout the procedure.

256 4 00:>a;49l9S9

C2 «n«<9 y

C2 « SOU

bo-184* 00:27*3

Figure 8.12: Energ}' in the alpha, beta and delta bands of the left and right occipital EEG in WE. Delta activity increases throughout the procedure.

251 »'1 rw>

02 «rargy «lavta

1400 00 / s . / — - — ^ ooo

32 «n«Qy m a«n«

A y \ _ - r - \ / . " ' - 1-0000

700COO seocoQ «2COOC

DO-18: *9 00-27 43

Figure 8.13: Energy* in the alpha, beta and delta bands at the frontal and occipital midline EEG electrodes in WE. Delta activity increases at both electrodes as the procedure progresses, while higher frequency activity attenuates.

258 to participate in the study. This patient demonstrated changes in the spectral distri­ bution of the EEG in response to the conditions of the INR procedure - inflation of the balloon in the carotid artery and lowering of the blood pressure. These changes are consistent with what was expected. Given the extent of contamination of the

EEG by muscle artifact and the resilience of the patient’s cerebral circulation when challenged by bilateral carotid artery occlusion and a medically induced decrease in blood pressure, the changes in the EEG spectra were expected to be subtle. Never­ theless, an examination of the band energy plots shows that there is a shift toward the lower frequencies in response to the interference. This response could perhaps also be expected in NICU patients.

2 5 9 CHAPTER 9

Conclusion

9.1 Contributions

Originally, the proposal for this dissertation specified that a continuous neurolog­ ical monitoring system for use in the NICU would be designed, built and tested. A system of hardware and software would be developed to continuously and automati­ cally monitor the progression of the amplitudes and latencies of specific peaks of the averaged evoked potentials in NICU patients along with other routinely monitored physiological parameters. A visual display summarizing the history of the parameter values would be provided for the physician to review the patient’s progress. A facility for extrapolation would be provided to identify trends in the data and to warn of possible deterioration in the patient’s condition. The system would be evaluated in

NICU patients.

Table 9.1 summarizes how these objectives were met. The left column lists the aspect of the system and subsequent columns list the figures in the dissertation that show corresponding normal or abnormal results. A reference to an external source that verifies that the observed behavior is expected is listed in the right column in each case. The real contribution of this system lies in the combination of tests rather

260 Test Norm. Abn. Abnormality Support, ref. fig- fig- BAEP single 7.1 7.3 Thai. inj. III,V [23) del. BAEP stack 3.9 7.8 Cerebel. hem. [23] Ill, V abs. SEP single 1.3 7.6 ICH N20,P25 [86| abs. SEP stack 8.1 7.5 ICH hern. [86] VEP single 1.5 - - [86]

VEP stack -- - [86] EEG single 7.2 - - [23] EEG stack 8.1 - - [1491

EEG spectrum - -- [149) EEG spectral s- 8.3 8.13 Ind. hypot., oc- [149) tack cl. Phys. parame­ 8.6 7.9 Cerebel. hern. [83) ters Extrapolât. - 7.19 Hypox. enc. analysis

Variability anal­ - 7.17 Cerebel. hern. [41j ysis State space repr. 3.15 7.20 Hera. ICH [20] Remote review 3.16 - --

Table 9.1: Summary of system features verified in the results. Missing parameters do not imply that the features are untested, only that no explanatory figure is included in this dissertation or that the feature is irrelevant (for example abnormal and normal review are identical).

261 than the individual tests that it provides. In addition, the data are automatically an­ alyzed and presented in a way that no other system has done. It is this combination of integration and automatic processing for which the time is now right in biomedical engineering. Computing facilities have improved to the point that very sophisticated analysis of data is possible at the bedside. Simple and complex monitors that moni­ tor singular aspects of the patient's physiology abound. It is now time to apply the available computing power to combining all this information into a meaningful overall assessment of the patient’s state rather than viewing each parameter and each organ system as an entity separable from the whole. With the current state of internetwork­ ing, it is also the right time to integrate computing systems and patient monitors in the hospital with the network for convenient access at the patient bedside, in the physician’s office and possibly at the home of the physician or by expert personnel outside the hospital.

This research attempts to express these various philosophies. The research began with the requirement for a continuous monitor of neurological function at the bedside.

This is an important area that has been neglected for a long time - while cardiac, renal and other monitors are routinely used in intensive care, a critically ill stroke patient entering the hospital has no monitor continuously monitoring the state of his or her brain. Part of the reason for this neglect is that brain function can only be measured noninvasively at the bedside by electrophysiological means. Electrophysio- logical signals are not easy to obtain, and require sophisticated training to interpret correctly. NICU nursing personnel cannot be expected to have this expertise. The

NICU system attempts to alleviate this problem by interpreting the signals itself, while at the same time keeping a record of the raw (averaged) data for review by the

262 neurologist should this be necessary. Since neurologists have duties elsewhere and are seldom to be found in the NICU, the networking feature of the hospital is used to bring the data to the neurologist rather than expecting the neurologist to appear in the NICU at the patient’s bedside. In the NICU system, this feature was extended to a web-based review system so that the data could potentially be verified by ex­ perts at other institutions should there not be a suitably qualified person available on location.

Another reason that bedside monitoring of the function of the central nervous system (CNS) has perhaps been neglected is that its function is intimately connected with the functioning of all the other organ systems, so that deciding just what to measure and what to do with the information is not trivial. This is the reason that from the start, information such as heart rate and respiration rate were included along with the other data collected directly from the brain. The dynamics of all of these signals affect one another. Subtle changes in the condition of the patient would presumably be reflected in the combination of parameters rather than as a large change in a single parameter. The NICU monitoring system lays the groundwork for further study in analyzing this combination of CNS-related parameters by providing a mechanism for acquiring and storing all of the information in an organized way along with its timing information. It also provides review mechanisms for different levels of user - casual browsing by internet or serious analysis using the local review workstation and the several analysis tools that have been developed.

263 9.2 Limitations and future study

The main question that remains is whether the NICU system gives an early indi­ cation of changes before they can be observed by other methods in the NICU. Several patients were monitored in the NICU and a short study is ongoing with the INR department. The author undertook these studies in some isolation and with minimal funding. Within these restrictions it was possible only to verify the correct operation of the system under various conditions, that the system may be used to correctly diagnose normal and abnormal conditions, and to provide some anecdotal evidence that the system is able to detect changes as they occur in the patient.

The protocols for two studies have been described in some detail in this dissertation and a serious effort to implement them has been made. If sufficient funding could be procured and the cooperation of technicians and physicians enlisted, these protocols could provide the basis for a meaningful set of clinical trials that would make a definite contribution to the medical literature on the subject of neurological intensive care monitoring and quite possibly to the state of patient care.

If the trials were undertaken, a useful spin-off would be that a large amount of valuable multimodality data would be generated. This could be used to develop further algorithms for biosignal interpretation. With enough data, such interesting approaches as neural networks, meaningful state vector analysis and a thorough study of the dynamics of the CNS could be undertaken and this would contribute not only to the state of NICU monitoring and patient care, but also to our knowledge of how the brain functions.

264 BIBLIOGRAPHY

[1] H. D. I. Abarbanel, Analysis of Observed Chaotic Data, New York: Springer Verlag, 1996.

[2] J. J. Ackmann, S. J. Larson, A. Sances Jr., R. E. Barr, Non-invasive monitoring techniques in neurosurgical intensive care. Journal of Clinical Engineering, vol. 4, no. 4, pp. 329-337, Oct.-Dec. 1979.

[3] J. J. Ackmann, Acquisition and signal processing of evoked potentials in neurolog­ ical intensive care, Proceedings of the 1987 International Conference on Industrial Electronics, Control, and Instrumentation, Cambridge, MA, pp. 1066-1071, 1987.

[4] J. J. Ackmann, Integrated system for the study of serial evoked potentials in head injury. Proceedings of the Ninth Annual Conference of the IEEE Engineering in Medicine and Biology Conference, Boston, MA, pp. 593-594, 1987.

[5| R. Agarvval, J. Gotman, D. Flanagan, B. Rosenblatt, Automatic EEG analysis during long-term monitoring in the ICU, Electroencephalography and Clinical Neu­ rophysiology, vol. 107, no. 1, pp. 44-58, July 1998.

[6j M. Akay, Detection and Estimation Methods for Biomedical Signals, London: A- cademic Press, 1996.

[7] M. Akay, J. A. Daubenspeck, Investigating the contamination of electroencephalo­ grams by facial muscle electromyographic activity using matching pursuit. Brain and Language, vol. 66, pp. 184-200, 1999.

[8] C. H. Allen, J. D. Ward, An evidence-based approach to management of increased intracranial pressure. Critical Care Clinics, vol. 14, no. 3, pp. 485-495, 1998.

[9] D. C. Anderson, S. Bundlie, G. L. Rocks wold. Multimodality evoked potentials in closed head trauma. Archives of Neurology, vol. 41, pp. 369-374, April 1984.

[10] T. G. Buchman, Physiologic stability and physiologic state. The Journal of Trau­ ma, Injury, Infection, and Critical Care, vol. 41, no. 4, pp. 599-605, 1996.

265 [11] J. E. Bellamy, Medical diagnosis, diagnostic spaces, and fuzzy systems. Journal of the American Veterinary Medical Association, vol. 210, no. 3, pp. 390-396, Feb. 1, 1997.

[12] A. Bezerianos, N. Laskaris, S. Fotopoulos, P. Papathanasopoulos, Data depen­ dent weighted averages for recording of evoked potential signals.Electroencephalog­ raphy and Clinical Neurophysiology, vol. 96, no. 5, pp. 468-471, Sep. 1995.

[13] A. Bianchi, S. Cerutti, G. Comi, S. Natali, M. G. Sora, Power spectrum analysis of heart rate variability signal in neurological central and autonomic pathologies. Proceedings of the 12th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, part 2, Philadelphia, PA, pp. 753-754, Nov. 1990.

[14] L.M. Bidaut, R. Pascual-Marqui, J. Delavelle, A. Naimi, M. Seeck, C. Michel, D. Slosman, O. Ratib, D. Ruefenacht, T. Landis, N. de Tribolet, J. Scherrer, F. Terrier, Three- to five-dimensional biomedical multisensor imaging for the assessment of neurological (dys)function. Journal of Digital Imaging, vol. 9, no. 4. pp. 185-198, Nov. 1996.

[15] G. E. Birch, P.D. LawTence, J. C. Lind, R. D. Hare, Application of prewhitening to AR spectral estimation of EEG, IEEE Transactions on Biomedical Engineering, vol. 35, no. 8, pp. 640-645, Aug. 1988.

[16] K. J. Blinowska, P. J. Durka, Application of wavelet transform and matching pursuit to the time-varying EEG signals, Proceedings of the Artificial Neural Net­ works in Engineering Conference (ANNIE’94), vol. 4, St. Louis, MO, pp. 535-540, Nov. 1994.

[17] M. J. Bloom, Anesthesia for the patient with neurologic disease: EEG monitor­ ing: Intraoperative application. Anesthesiology Clinics of North America, vol. 15, no. 3, Sep. 1997.

[18] P. O. Borjesson, C. W. Sundberg, Simple appproximations of the error function Q(x) for communications applications, IEEE Transactions on Communications, COM-27, pp. 639-643, Mar. 1979.

[19] J. R. Boston, Automated monitoring of evoked potentials. Proceedings of the Eighth Annual Conference of the IEEE Engineering in Medicine and Biology Soci­ ety, Fort Worth, TX, pp. 1311-1314, 1986.

[20] T. G. Buchman, Physiologic stability and physiologic state. The Journal of Trau­ ma: Injury, Infection, and Critical Care, vol. 41, no. 4, pp.599-605, 1996.

[21] J. Buckheit, S. Chen, D. Donoho, I. .lohnstone, WaveLab Reference Manual, Stanford University, Palo Alto, December 1995.

266 [22j C. E. Buzea, Understanding computerized EEG monitoring in the intensive care unit, Journal of Neuroscience Nursing, vol. 27, no. 5, pp. 292-297, Oct. 1995.

[23] Department of Neurology, Cleveland Clinical Foundation, Cleveland Clinic Com­ prehensive Clinical Neurophysiology, vols. 1 and Cleveland, 2, 1998.

[24] F. H. Y. Chan, F. K. Lam, P. W. F. Poon, W. Qiu, Detection of brainstem auditory evoked potential by adaptive filtering. Medical and Biological Engineering and Computing, vol. 33, pp. 69-75, 1995.

[25] F. H. Y. Chan, W. Qiu, F. K. Lam, P. W. F. Poon, M. K. Lam, Evoked potential estimation using modified time-sequenced adaptive filter. Medical and Biological Engineering and Computing, vol. 36, pp. 407-414, 1998.

[26] K. H. Chiappa, R. A. Hill, Evaluation and prognostication in coma, Electroen­ cephalography and Clinical Neurophysiology, vol. 106, no. 2, pp. 149-155, Feb. 1998.

[27] R. R. Clancy, Differential diagnosis and contribution of EEC. In: L. L. Wright, C. B. Merenstein, D. Hirtz (editors), Report of the Workshop on Acute Perinatal Asphyxia in Term Infants. NIK, US Dept, of HHS, Rockville, pp. 115-121, 1993.

[28] P. M. Clarkson, H. Stark, Signal Processing Methods for .A.udio, Images and Telecommunications, Academic Press, London, 1995.

[29] P. J. M. Cluitmans, J. W. Jansen, J. E. W. Beneken, Artifact detection and re­ moval during auditory evoked potential monitoring. Journal of Clinical Monitoring, vol. 9, no. 2., pp. 112-120, Apr. 1993.

[30] Z. Cvetkovic and M. Vetterli, Discrete-time extrema representation: design and consistent reconstruction, IEEE Transactions on Signal Processing, vol. 43, no. 3, March 1995.

[31] C. E. Davila, R. Srebro, I. A. Chaleb, Optimal detection of visual evoked poten­ tials, IEEE Transactions on Biomedical Engineering, vol. 45, no. 6, June 1998.

[32] C. M. de Ciorgio, .A.. L. Rabinowicz, P. S. Cott, Predictive value of P300 event- related potentials compared with EEC and somatosensor\" evoked potentials in non-traumatic coma, Acta Neurologica Scandinavica, vol. 87, pp. 423-427, 1993.

[33] T. Demiralp, A. Ademoglu, M. Schurmann, C. Basar-Eroglu, E. Basar, Detection of P300 waves in single trials by the wavelet transform (WT), Brain and Language, vol. 66, pp. 108-128, 1999.

[34] T. Demiralp, J. Yordanova, V. Kolev, A. Ademoglu, M. Devrim, V. J. Samar, Time-frequency analysis of single-sweep event-related potentials by means of fast wavelet transform. Brain and Language, vol. 66, pp. 129-145, 1999.

267 [351 R. A. Dobie, Objective response detection, Ear and Hearing, vol. 14, no. 1, 1993.

[36) P. J. Durka, K. J. Blinowska, Analysis of EEG transients by means of matching pursuit. Annals of Biomedical Engineering, vol. 23, no. 5, pp. 608-611, Sep.-Oct. 1995.

[37) J. I. Frank, Large hemispheric infarction, deterioration, and intracranial pressure. Neurology, vol. 45, pp. 1286-1290, 1995.

[38) P. J. Godin, T. G. Buchman, Uncoupling of biological oscillators: a complemen­ tary hypothesis concerning the pathogenesis of multiple ogran dysfunction syn­ drome, Critical Care Medicine, vol. 24, no. 7, pp. 1107-1116, 1996.

[39) P. J. Godin, L. A. Fleisher, A. Eidsath, R. W. Vandivier, H. L. Preas, S. M. Banks, T. G. Buchman, A. F. Suffredini, Experimental human endotoxemia in­ creases cardiac regularity: results from a prospective, randomized, crossover trial. Critical Care Medicine, vol. 24, no. 7, pp. 1117-1124, 1996.

[40) V. Goel, A. M. Brambrink, A. Baykal, R. C. Koehler, D. F. Hanley, N. V. Thakor, Dominant frequency analysis of EEG reveals brains response during injury and recovery, IEEE Transactions on Biomedical Engineering, vol. 43 no. 11, pp. 1083-1092, 1996.

[41) A. L. Goldberger, Fractal variability versus pathologic periodicity: complexity loss and stereotypy in disease. Perspectives in Biology and Medicine, vol. 40, no. 4, pp. 543-561, Summer 1997.

[42) A. L. Goldberger, Non-linear dynamics for clinicians: chaos theor}^ fractals, and complexity at the bedside. The Lancet, vol. 347, pp. 1312-1314, 1996.

[43) J. Gotman, L. Y. Wang, State-dependent spike detection: concepts and pre­ liminary results. Electroencephalography and Clinical Neurophysiology, vol. 79, pp. 11-19, 1991.

[44) S. Goto, M. Nakamura, K. Uosaki, On-line spectral estimation of nonstationary time series based on AR model parameter estimation and order selection with a forgetting factor, IEEE Transactions on Signal Processing, vol. 43, no. 6, pp. 1519- 1522, June 1995.

[45) B. A. Green, L. F. Marshall, T. J. Gallagher, Intensive Care for Neurological Trauma and Disease, Academic Press, New York, 1982.

[46) T. Gronfors, M. Juhola, R. Johansson, Evaluation of some nonrecursive digital filters for signals of auditory evoked responses. Biological Cybernetics, vol. 55, pp. 533-536, 1992.

268 [47| T. Gronfors, Peak identification of auditory brainstem responses with multifilters and attributed automaton, Computer Methods and Programs in Biomedicine, vol. 40, pp. 83-87, 1993.

[48] P. Omen, C. Liu, Contemporary issues in trauma: Current trends in the man­ agement of head injury. Emergency Medicine Clinics of North America, vol. 16, no. 1, Feb. 1998.

[49] J.P. Gruen, M. Weiss, Management of complicated neurologic injuries. Surgical Clinics of North America, vol. 76, no. 4, pp. 905-922, 1996.

[50] R. A. Hill, K. H. Chiappa, Electrophysiologic monitoring in the intensive care unit, Canadian Journal of Neurological Sciences, vol. 21 suppl. 1, pp. S12-16, 1994.

[51] M. J. Hilz, G. Litscher, M. Weis, D. Claus, K. F. Druschky, G. P furtscheller, B. Neundorfer, Continuous multivariable monitoring in neurological intensive care patients - preliminary reports on four cases. Intensive Care Medicine, vol. 17, pp. 87-93, 1991.

[52] N. Hock, Neuroprotective and thrombolytic agents: advances in stroke treatmen- t. Journal of Neuroscience Nursing, vol. 30, no. 3, pp. 175-184, pp. 175-184, June 1998.

[53] D. Hoyer, K. Conrad, H. Wagner, R. Bauer, U. Zwiener, Method for the adaptive design of power spectral parameters using artificial neural networks and its applica­ tion in the EEG-classification during brain ischemia, Proceedings of the 1995 IEEE Engineering in Medicine and Biology 17th Annual Conference and 21st Canadian Medical and Biological Engineering Conference, vol. 17, no. 1, Montreal, Canada, pp. 787-788, Sep. 1995.

[54] S. A. Hyman, P. H. King, J. Xu, Automated evoked potential monitoring, In­ ternational Journal of Clinical Monitoring and Computing, vol. 10, pp. 201-214, 1993.

[55] H. Jager, C. Schelenz, L. Leistritz, H. Witte, M. Specht, K. Reinhart, Burstde- tektion in EEG-Verlaufsuntersuchungen bei intensivpflichtigen Patienten, Biomedi- zinische Technik, vol. 42, suppl. pp. 415-416, 1997.

[56] W. Jiang, W. R. Hathaway, S. McNulty, R. L. Larsen, K. L. Hansley, Y. Zhang, C. M. O’Connor, Ability of heart rate variability to predict prognosis in patients with advanced congestive heart failure, American Journal of Cardiology, vol. 80, no. 6, pp. 808-811, Sep. 15, 1997.

[57] K. G. Jordan, Neurophysiologic monitoring in the neuroscience intensive care unit. Neurologic Clinics, vol. 13, no. 3, pp. 579-626, Aug. 1995.

269 [o8j K. G. Jordan, Continuous EEG monitoring in the neuroscience intensive care unit and emergency department. Journal of Clinical Neurophysiology, vol. 16, no. I, pp. 14-39, 1999.

[59] P. A. Karjalainen, J. P. Kaipio, A. S. Koistinen, M. Vauhkonen, Subspace regu­ larization method for the single-trial estimation of evoked potentials, IEEE Trans­ actions on Biomedical Engineering, vol. 46, no. 7, July 1999.

[60] S. M. Kay, Modem Spectral Estimation, Prentice Hall, Englewood Cliffs, 1988.

[61] J. Kay, Continuous EEC monitoring in the intensive care unit, Canadian Journal of Neurological Sciences, vol. 25, no. 1, pp. 812-815, Feb. 1998.

[62] H. L. Kennedy, Heart rate variability-a potential, noninvasive prognostic index in the critically ill patient. Critical Care Medicine, vol. 26, no. 2, pp. 213-214, Feb. 1998.

[63] I. Kondakor, D. Brandeis, J. Wackermann, K. Kochi, T. Koenig, E. Frei, R.D. Pascual-Marqui, T. Yagyu, D. Lehmann, Multichannel EEC fields during and with­ out visual input: frequency domain model source: locations and dimensional com­ plexities, Neuroscience Letters, vol. 226, no. 1, pp. 49-52, Apr. 18, 1997.

[64] X. Kong, T. Qiu, Adaptive estimation of latency change in evoked potentials by direct least mean p-norm time-delay estimation, IEEE Transactions on Biomedical Engineering, vol. 46, no. 8, pp. 994-1003, 1999.

[65] D. Krieger, Personal communication, Cleveland, 1999.

[66] 8. C. Lage, M.A. Gutierrez, M. A. L. Nicolelis, 8. 8. Furuie, Bedside computerized system for monitoring and processing of biological signals in intensive care units. Computers in Cardiology Conference, Louvain-La-Neuve, Belgium, pp. 561-564, 1987.

[67] C. A. Lanza, V. Guido, M. M. Caleazzi, M. Mustilli, R. Natali, C. lerardi, C. Milici, F. Burzotta, V. Pasceri, F. Tomassini, A. Lupi, A. Maseri, Prognostic role of heart rate variability in patients with a recent acute myocardial infarction, American Journal of Cardiology, vol. 82, no. 11, pp. 1323-1328, Dec. 1, 1998.

[68] H. 8. Lee, 8. H. Park, E. J. Woo, Remote patient monitoring service through World-Wide Web, Proceedings of the 1997 19th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 2, Chicago, IL, pp. 928-931, Oct.-Nov. 1997.

[69] D. Lehmann, T. Koenig, 8patio-temporal dynamics of alpha brain electric fields, and cognitive modes. International Journal of Psychophysiology, vol. 26, no. 1-3, pp. 99-112, June 1997.

270 [70) D. Liberati, Evoked potentials in the brain: physiological and abstract optimal models for single sweeps processing, Proceedings of the Fifteenth Annual Northeast Bioengineering Conference, Boston, MA, pp. 95-96, 1989.

[71) D. Libes, Writing CGI scripts in Tel, Proceedings of the Fourth Annual Tcl/Tk Workshop ’96, Monterey, July 1996.

[72) C. Madl et al.. Detection of nontraumatic comatose patients with no benefit of intensive care treatment by recording of sensory evoked potentials. Archives of Neurology, vol. 53, pp. 512-516, June 1996.

[73) S. Mallat, Zero-crossings of a wavelet transform, IEEE Transactions on Infor­ mation Theory, vol. 37, no. 4, July 1990.

[74) S. Mallat, Z. Zhang, Matching pursuit with time-firequency dictionaries, IEEE Transactions on Signal Processing, December 1993.

[75) S. Mallat and S. Zhong, Characterization of signals from multiscale edges,IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 14, no. 7, July 1992.

[76) E.M. Manno, Transcranial doppler ultrasonography in the neurocritical care unit, Critical Care Clinics, vol. 13, no. 1, pp. 79-104, Jan. 1997.

[77) B. Mat ta, D. Menon, Severe head injury in the United Kingdom and Ireland: a survey of practice and implications for management. Critical Care Medicine, vol. 24, no. 10, pp. 1743-1748, 1996.

[78) S. Naik-Tolani, J. M. Oropello,E. Benjamin, Intensive care unit complications: Neurologic complications in the intensive care unit. Clinics in Chest Medicine, vol. 20, no. 2, June 1999.

[79) J. H. Mathews, Numerical Methods, International Edition, Prentice Hall Inter­ national, London, pp. 232-233, 1987.

[80) L. McLean, R. N. Scott, P. .A.. Parker, Stimulus artifact reduction in somatosenso­ ry evoked potential acquisition: A conceptual model. Proceedings of the 1995 IEEE Engineering in Medicine and Biology 17th Annual Conference and 21st Canadian Medical and Biological Engineering Conference, vol. 17, no. 2, Montreal, Canada, pp. 909-910, Sep. 1995.

[81) C. D. McGillem, J. I. Aunon, D. G. Childers, Signal processing in evoked po­ tential research: Applications of filtering and pattern recognition, CRC Critical Reviews in Bioengineering, vol. 6, no. 3, pp 225-265, 1981.

271 [82| F. B. Meyer, Occlusive Cerebrovascular Disease, 2nd Edition, W. B. Saunders Company, Philadelphia, 1994.

[83] R. E. Minahan, A. Bhardwaj, M. A. Williams, Critical care monitoring for cere­ brovascular disease. New Horizons, vol. 5, no. 4, pp. 406-421, 1997.

[84] E- Miskiel, O. Ozdamar, E. A. Peterson, Microcomputer-based automatic ABR monitoring in the ICU, Proceedings of the 37th Annual Conference on Engineering in Medicine and Biology, vol. 26, Los Angeles, CA, pg. 168, 1984.

[85] E. Miskiel, O. Ozdamar, Computer monitoring of auditory brain stem reponses, Computers in Biology and Medicine, vol. 17, no. 3, pp. 185-192, 1987.

[86] K. E. Misulis, Spehlmann’s Evoked Potential Primer (2nd edition), Butterw^orth- Heinemann, Newton, 1994.

[87] R. Moberg, C. Woodruff, E. lida, S. Mesibov, EEG analysis for long term mon­ itoring in the ICU, Proceedings of the Tenth IEEE Engineering in Medicine and Biology Society Annual Conference, part 4, pp. 1755-1756, 1988.

[88] D. W. Molavd, The Washington School of Medicine Neuroscience Tutorial, http:// thalamus, wustl. edu/course, 1997.

[89] P. C. Molenaar, Time series analysis and its relationship with longitudinal anal­ ysis, International Journal of Sports Medicine, vol. 18, suppl. 3, pp. S232-237, July 1997.

[90] E. J. Moore, Bases of Auditory Brain-stem Evoked Responses, G rune and Strat­ ton, New York, pp. 157-195, 1983.

[91] R. Muhler, H. von Specht, J. Pethe, Vergleich von Verfahren zur Reduktion der Reststorung bei uberschwelligen FAEP-Registrierungen, Laryngo- Rhino- Otologie, vol. 77, no. 7, pp. 382-387, 1998.

[92] R.S. Murchison, Get patient information anytime, anywhere, Nursing Manage­ ment, vol. 30, no. 5, pp. 19-20, 1999.

[93] J. Muthuswamy, N. V. Thakor, Spectral analysis methods for neurological sig­ nals, Journal of Neuroscience Methods, vol. 83, no. 1, pp. 1-14, -A.ug. 1998.

[94] J. Muthuswamy, D.L. Sherman, N.V. Thakor, Higher-order spectral analysis of burst patterns in EEG, IEEE Transactions on Biomedical Engineering, vol. 46, no. 1, pp. 92-99, Jan. 1999.

[95] J. H. Nagel, K. Han, B. E. Hurwitz, N. Schneiderman, Assessment and diagnostic applications of heart rate variability. Biomedical Engineering Applications Basis Communications, vol. 5, no. 2, pp. 147-158, Apr. 1993.

272 [96] M. Nakamura, H. Shibasaki, S. Nishida, Method for recording short latency e- voked potentials using an EKG artifact elimination procedure, Journal of Biomed­ ical Engineering, vol. 12, no. 1, pp. 51-56, Jan 1990.

[97] M. Nakao, K. Hirokawa, N. Kawachi, M. Yamamoto, M. Munakata, Y. Imai, K. Abe, Trajectory analyses of time-varying multivariate cardiovascular dynamics based on a distance measure. Proceedings of the 1995 IEEE Engineering in Medicine and Biology 17th Annual Conference and 21st Canadian Medical and Biological Engineering Conference, vol. 17, no. 1, Montreal, Canada, pp. 117-118, Sep. 1995.

[98] H. E. Nan, J. Rimpel, Multimodality evoked potentials and electroencephalog­ raphy in severe coma cases. Intensive Care Medicine, vol. 13, pp. 249-255, 1987.

[99] V. Nenov, J. Klopp, Remote access to neurosurgical ICU physiological data using the World Wide Web, Studies in Health Technology and Informatics, vol. 29, pp. 242-249, 1996.

[100] V. Nenov, J. Klopp, Remote analysis of physiological data from neurosurgical ICU patients. Journal of the American Medical Informatics Association, vol. 3, no. 5, pp. 318-327, Sep.-Oct. 1996.

[101] S. Nishida, M. Nakamura, K. Shindo, M. Kanda, H. Shibasaki, A morphologi­ cal filter for extracting waveform characteristics of single-sweep evoked potentials, Automatica, vol. 35, pp. 937-943, 1999.

[102] J. B. Nitschke, G. A. Miller, E. W. Cook III, Digital filtering in EEG/ERP analysis: Some technical and empirical comparisons. Behavior Research Methods, Instruments, and Computers, vol. 30, no. 1, pp. 54-67, 1998.

[103] M.R. Nuwer, Fundamentals of evoked potentials and common clinical applica­ tions today. Electroencephalography and Clinical Neurophysiology, vol. 106, no. 2, pp. 142-148, 1998.

[104] N. Papoulis, Probability, Random Variables and Stochastic Processes, 3rd Edi­ tion, McGraw-Hill, New York, 1991.

[105] J. Pardey, S. Roberts, L. Tarassenko, A review of parametric modeling tech­ niques for EEC analysis. Medical Engineering and Physics, vol. 18, no. 1, pp. 2-11, Jan. 1996.

[106] V. Parsa, P. Parker, R. Scott, Nonlinear stimulus artifact cancellation in so­ matosensory evoked potential studies. Proceedings of the 1995 IEEE 21st Annual Northeast Bioengineering Conference, Bar Harbor, ME, pp. 61-62, May, 1995.

273 [107] Y. Pei, J. Lin, Z. VVu, A practicable adaptive averaging method for evoked potential research. Proceedings of the 12th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, part 2, Philadelphia, PA, pp. 893-894, Nov. 1990.

[108] G. Pfurtscheller, G. Schwarz, O. Schroettner, G. Litscher, H. Maresch, L. Auer, W. List, Continuous and simultaneous monitoring of EEG spectra and brainstem auditory and somatosensorj" evoked potentials in the intensive care unit and the operating room. Journal of Clinical Neurophysiology, vol. 4, no. 4, pp. 393-398, 1987.

[109] W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery, Numerical Recipes in C: The Art of Scientific Computing, 2nd edition, Cambridge University Press, Cambridge, 1992.

[110] P. Prior, Rationale and utility of neurophysiological investigations in clinical monitoring for brain and spinal cord ischaemia during surgery and intensive care. Computer Methods and Programs in Biomedicine, vol. 51, no. 1-2, pp. 13-27, Oct. 1996.

[111] W.S. Pritchard, D.W. Duke, K.K. Krieble, Dimensional analysis of resting hu­ man EEG. II: Surrogate-data testing indicates nonlinearity but not low-dimensional chaos. Psychophysiology, vol. 32, no. 5, pp. 486-491, Sep. 1995.

[112] W. S. Pritchard, K. K. Krieble, D. W. Duke, On the validity of estimating EEG correlation dimension from a spatial embedding. Psychophysiology, vol. 33, no. 4, pp. 362-368, .July 1996.

[113] D. Quinonez, Common applications of electrophysiology (EEG) in the past and today: the technologist’s view. Electroencephalography and Clinical Neurophysiol­ ogy, vol. 106, no. 2, Feb. 1998.

[114] M. Y. Rady, E. P. Rivers, R. M. Nowak, Resuscitation of the critically ill in the ED: responses of blood pressure, heart rate, shock index, central venous oxygen saturation, and lactate, American Journal of Emergency Medicine, vol. 14, no. 2, pp. 218-225, Mar. 1996.

[115] I. J. Rampil, J. S. Kim, R. Lenhardt, C. Negishi, D. I. Sessler, Bispectral EEG index during nitrous oxide administration. Anesthesiology, vol. 89 no. 3, pp. 671- 677, 1998.

[116] J. Raz, L. Dickerson, B. Turetsky, A wavelet packet model of evoked potentials. Brain and Language, vol. 66, pp. 61-88, 1999.

[117] D. Regan, Human Brain Electrophysiology, Elsevier, New York, 1989.

274 [118] L A. Rezek, S. J. Roberts, Stochastic complexity measures for physiological signal analysis, IEEE Transactions on Biomedical Engineering, vol. 45, no. 9, pp. 1186-1191, Sep. 1998.

[119] K. Rieke, S. Schwab, D. Krieger, R. von Kummer, A. Aschoff, V. Schuchardt, W. Hacke, Decompressive surgery in space-occupying hemispheric infarction: results of an open, prospective trial. Critical Care Medicine, vol. 23, no. 9, 1995.

[120] J. Roschke, J. Aldenhoff, A non-linear approach to brain-function - determin­ istic chaos and sleep eeg". Sleep, vol. 15, no. 2, pp. 95-101, 1992.

[121] B. Rosenblatt, Monitoring the central nervous system in children with con­ genital heart defects: clinical neurophysiological techniques. Seminars in Pediatric Neurology, vol. 6, no. 1, pp. 27-31, Mar. 1999.

[122] M. T. Rosenstein, J. J. Collins, Visualizing the effects of filtering chaotic signals. Chaos and Graphics, vol. 18, no. 4, pp. 587-592, 1994.

[123] V. J. Samar, A. Bopardikar, R. Rao, K. Swartz, Wavelet analysis of neuro­ electric waveforms: conceptual tutorial. Brain and Language, vol. 66, pp. 7-60, 1999.

[124] F. Sattar, B. Lovstrom, B. Mandersson, G. Salomonsson, Estimation of evoked potentials by using a bandpass filter bank. Proceedings of the 13th Annual Interna­ tional Conference of the IEEE Engineering in Medicine and Biology Society, vol. 13, part 1, pp. 425-426, Orlando, FL, Oct.-Nov. 1991.

[125] S. Schwab, A. -\schoff, M. Spranger, F. Albert, W. Hacke, The value of in­ tracranial pressure monitoring in acute hemispheric stroke, Neurology, vol. 47, pp. 393-398, 1996.

[126] S. Schwarz, S. Schwab, .A.. .Aschoff, W. Hacke, Favorable recovery from bilateral loss of somatosensory evoked potentials. Critical Care Medicine, vol. 27, no. 1, pp. 182-187, 1999.

[127] J. A. Sgro, R. G. Emerson, T. A. Pedley, Real time reconstruction of evoked potentials using a new two-dimensional filter method. Electroencephalography and Clinical Neurophysiology, vol. 62, no. 5, pp. 372-380, Sep., 1985.

[128] D. L. Sherman, A. M. Brambrink, D. Walterspacher, V. K. Dasika, R. Ichord, N. V. Thakor, Detecting EEG bursts after hypoxic-ischemic injury using energy- operators. Proceedings of the 1997 19th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, vol. 3, Chicago, IL, pp. 1188- 1190, Oct.-Nov. 1997.

275 [129j D. L. Sherman, A. M. Brambrink, R. N. Ichord, V. K. Dasika, R. C. Koehler, R. J. Traystman, D. F. Hanley, N. V. Thakor, Quantitative EEG during early re­ covery from hypoxic-ischemic injury in immature piglets. Unpublished manuscript obtained from the author at Johns Hopkins University, Baltimore, 1998.

[130] T. Shibata, K. Kaneko, Heterogeneity-induced order in globally coupled chaotic systems, Europhysics Letters, vol. 38, no. 6, pp. 417-422, 1997.

[131] Y. Si, J. Gotman, A. Pasupathy, D. Flanagan, B. Rosenblatt, R. Gottesman, An expert system for EEG monitoring in the pediatric intensive care unit, Elec­ troencephalography and Clinical Neurophysiology, vol. 106, no. 6, pp. 488-500, June 1998.

[132] T. B. Sloan, Electrophysiologic monitoring in head injury. New Horizons, pp. 431-438, August 1995.

[133] M.S. Smith, C.F. Feied, The next-generation emergency department, Annals of Emergency Medicine, vol. 32, no. 1, pp. 65-74, 1998.

[134] M. P. Spencer (editor), Ultrasonic Diagnosis of Cerebrovascular Disease, Mar- tinus Nijhoff Publishers, Dordrecht, 1987.

[135] R. L. Strub, F. W. Black, The Mental Status Examination in Neurology, 3rd Edition, F. A. Davis Company, Philadelphia, 1993.

[136] A. Suzuki, N. Mori, H. Hadeishi, K. Yoshioka, N. Yasui, Computerized mon­ itoring system in neurosurgical intensive care, Journal of Neuroscience Methods, vol. 26, pp. 133-139, pp. 445-475, 1993. 1988.

[137] O. Svensson, Tracking of changes in latency and amplitude of the evoked poten­ tial by using adaptive LMS filters and exponential averagers, IEEE Transactions on Biomedical Engineering, vol. 40, no. 10, pp. 1074-1079, Oct. 1993.

[138| M. Swiercz, Z. Mariak, J. Lewko, K. Chojnacki, A. Kozlowski, P. Piekarski, Neural network technique for detecting emergency states in neurosurgical patients. Medical and Biological Engineering and Computing, vol. 36, no. 6, pp. 717-722, Nov. 1998.

[139] N. V. Thakor, Evoked potential processing in neuro-critical monitoring. Annals of Biomedical Engineering, vol. 19, no. 5, pp. 622-623, 1991.

[140] J. Theiler, P. E. Rapp, Re-examination of the evidence for low-dimensional, nonlinear structure in the human electroencephalogram. Electroencephalography and Clinical Neurophysiology, vol. 98, no. 3, pp. 213-222, Mar. 1996.

276 [141] D. Hal Unwin, C. A. Ciller, T. A. Kopitnik, Central nervous sys tem monitoring: What helps, what does not. Surgical Critical Care, vol. 71, no. 4, August 1991.

[142] A. J. W. van der Kouwe, Signal Processing in a VEP-based Computer Interface for the Severely Physically Handicapped, M. Eng. Thesis, University of Pretoria, Pretoria, September 1995.

[143] A. J. W. van der Kouwe, R. C. Burgess, Peak detection in auditory and so­ matosensory evoked potentials by means of the zero-crossings wavelet transform. Proceedings of theSfth Rocky Mountain Bioengineering Symposium, Dayton, Ohio, U.S.A., pp. 71-76, 1996. [144] A. J. W. van der Kouwe, R. C. Burgess, Long-term evoked potential and EEC monitoring system for the neurointensive care unit. Journal of Clinical Neurophys­ iology, vol. 16, no. 2, 1999. [145] A. J. W. van der Kouwe, R. C. Burgess, Examining the dynamics of a vector representing neurophysiological state during intensive care, InterJoumal of Com­ plex Systems, submission no. 257, 11 December, 1998.

[146] A. J. W. van der Kouwe, R. C. Burgess, A continuous electrophysiological mon­ itoring system with web-based review for the neurological intensive care unit. Pro­ ceedings of the 3rd International Workshop on Biosignal Interpretation, Chicago, IL, June, 1999. [147] A.. J. W. van der Kouwe, P. J. Cilliers, A comparison of linear and nonlinear techniques for distinguishing phase-dissimilar VEP’s, Proceedings of the 1995 IEEE Conference on Nonlinear Signal and Image Processing, Halkidiki, 1995.

[148] P. M. Vespa, M. R. Nuwer,C. Juhasz, M. .Alexander, V. Nenov, N. Martin, D. P. Becker, Early detection of vasospasm after acute subarachnoid hemorrhage using continuous EEC ICU monitoring. Electroencephalography and Clinical Neurophys­ iology, vol. 103, no. 6, pp. 607-615, Dec. 1997.

[149] P. M. Vespa, V. Nenov, M. R. Nuwer, Continuous EEC monitoring in the intensive care unit: Early findings and clinical eflBcacy, Journal of Clinical Neuro­ physiology, vol. 16, no. 1, pp. 1-13, 1999.

[150] C. M. Viswanathan, C. -K. Peng, H. Eugene Stanley, A. L. Coldberger, Devia­ tions from uniform power law scaling in nonstationary time series. Physical Review E, vol. 55, no. 1, pp. 845-849, January 1997.

[151] Z. S. Wang, Y. S. Xia, W. H. Li, Z. Y. He, J. D. Z. Chen, Neural network solver for basis pursuit and its applications to time-frequency analysis of biomedical sig­ nals, Proceedings of the 1997 IEEE International Conference on Neural Networks, part 4, Houston, TX, pp. 2057-2060, June 1997.

277 [152] L. V. Williams, C. Eswaran, Automatic detection of intraoperative neurological injury, IEEE Transactions on Biomedical Engineering, vol. 46, no. 4, pp. 481-483, 1999.

[153] R. J. Winchell, D. B. Hoyt, Analysis of heart-rate variability: a noninvasive predictor of death and poor outcome in patients with severe head injury. Journal of Trauma, vol. 43, no. 6, pp. 927-933, Dec. 1997.

[154] T. Yagyu, J. Wackermann, M. Shigeta, V. Jelic, T. Kinoshita, K. Kochi, P. Julin, O- Almkvist, L.O. Wahlund, I. Kondakor, D. Lehmann, Global dimension­ al complexity of multichannel EEG in mild Alzheimer's disease and age-matched cohorts. Dementia and Geriatric Cognitive Disorders, vol. 8, no. 6, pp. 343-347, Nov.-Dee. 1997.

[155] S.-S. Yang, N.-S. Chu, C.-H. Wu, Role of somatosensory evoked potentials on hepatic encephalopathy. Biomedical Engineering Applications, Basis Communica­ tions, vol. 9, no. 3, pp. 154-157, June, 1997.

[156] I. Yaylali, H. Kocak, P. Jayakar, Detection of seizures from small samples using nonlinear dynamic system theory, IEEE. Transactions on Biomedical Engineering, vol. 43, no. 7, pp. 743-751, July 1996.

[157] Z. Ying, U. D. Schmid, J. Schmid, C. W. Hess, Motor and somatosensory evoked potentials in coma: analysis and relation to clinical status and outcome. Journal of Neurology, Neurosurgery, and Psychiatry, vol. 55, pp. 470-474, 1992.

[158] G. B. Young, K. G. Jordan, G. S. Doig, An assessment of nonconvulsive seizures in the intensive care unit using continuous EEG monitoring: an investigation of variables associated with mortality. Neurology, vol. 47, no. 1, pp. 83-89, July 1996.

[159] W. L. Young, J. Pile-Spellman, L. Hacein-Bey, S. Joshi, Invasive neuroradiologic procedures for cerebrovascular abnormalities: anesthetic considerations, Anesthe­ siology Clinics of North America, vol. 15, no. 3, pp. 631-653, Sep. 1997.

278