A Study of Physicians Interaction with Text-Based and Graphical-Based Electronic Patient

Subdural EEG Classification for Epileptic Seizure Focus Localization

Mercedes Cabrerizo and Malek Adjouadi Center for Advance Technology and Education Department of Electrical & Computer Engineering Florida International University, Miami FL 33174, USA Emails: [email protected], [email protected]

1. Introduction that discriminates in the feature space the two regions of electrodes leading and not leading to an ictal state . Epilepsy surgery is now being increasingly 2. Methods performed in children. Seizures are related to a malformed pathologic substrate that is more extensive 2.1 Participants than the lesion evident on MRI scans [2]. Pre-surgical evaluation is therefore more difficult than in adults and Eight children with medical refractory partial relies heavily on EEG data including interictal spike seizures that underwent pre-surgical evaluation have discharges. The EEG interictal data recorded inside the been analyzed in this study. The subdural EEG data was brain can be processed to define similar patterns evident recorded using XLTEK Neuroworks Ver.3.0.5 in those electrodes that lead to a given seizure to further (equipment manufactured by Excel Tech Ltd. Ontario, facilitate surgical planning [3]. Canada). Sampling frequency of 500Hz with 0.1-70 Hz This study provides a mathematical framework for bandpass filter settings and 12 bits A/D conversion the study of interictal EEG leading or not to an epileptic were used to obtain the digital EEG recordings. seizure. The EEG of epileptic subjects can be divided into two main categories, interictal and ictal. The 2.2 Implementation Steps interictal EEG is the EEG taken when the patient is not having seizures or in between seizures [2]. Interictal To classify those electrodes that lead to an epileptic activity is considered to be abnormal if it can occur in a seizure, a program was developed in order to quantify patient with epilepsy in the absence of an actual seizure. the patterns that are inherent to those electrodes. Input The ictal EEG activity on the other hand is when the data in this study was subdural EEG segments from 20 actual seizure occurs. to 3600 seconds of duration of epileptic patients. The study of this new application elicits how The classification algorithm consists of the different patients react prior to a seizure in view of the following steps: (1) Filtering the input EEG data, (2) collected EEG data such as to better detect such Assessment of the EEG Nonlinear Dynamics, (3) neurological disorders. The main objectives of this Extraction of features from the EEG data, (4) study are as follows: (1) to extract features that best Implementation of regression lines for each electrode characterize those EEG electrodes that lead to an ictal and parameter, (5) Neural network structure for linear activity; (2) to establish mathematical derivations that classification. provide not only quantitative measures, but also In step (1), filtering was performed applying the describes and locates the focus of an ictal activity;(3) to Singular Value Decomposition (SVD), which is based identify and formulate those patterns in EEG recordings on the eigenvalues decomposition. The approximated that are inherent to those electrodes that lead to a matrix, containing less noise was used in the seizure; (4) to correlate the clinical features with the subsequent steps. Step (2) dealt with the dynamics of EEG findings in order to determine whether the patient brain behavior. Since brain dynamics are nonlinear, this has a consistent source of ictal activity, which is study investigated methods such as the calculation of coming from the location concerning the group of correlation dimension integral, mobility and channels that present interictal activity; (5) to classify complexity. Step (3) was associated with extracting and to group those EEG channels that are known in features from the filtered EEG matrix using the advance to lead to seizures in order to extract aforementioned parameters of step 2 in order to similarities in their behavior, so a common behavioral discriminate between the two groups of electrodes pattern could be found; (6) to find a suitable classifier (leading and not leading to an epileptic seizure). All these three parameters were computed for each electrode separately using successive epochs or non- for the remaining electrodes that do not lead to such overlapping windows of 1 second for all the recorded state. Also, using different parameters, characterization subdural EEG data. By computing these parameters, a of the behavior of the interictal EEG over time is behavior for each feature over time was established for possible. The complexity results are the best compared each electrode. In step (4) all the different parameters to the other two parameters implemented. It produces were represented in time, regression lines for all of the most consistent and reliable results across all 8 these parameters were calculated in order to keep a patients included in the study. suitable track of the behavior of each electrode with Making the ANN converges and yielding accurate respect to the computed parameter. This also helps in classification results should be emphasized as well as determining a linear classifier that separates in the that the separability is achieved because of the choices parameter vs. time space two different classes of of the 3 discriminant features of mean, standard electrodes. At this stage (Step 5), a plot of the three deviation, and frequency power. This in itself selected features revealed well defined electrodes constitutes a mayor contribution of this dissertation. clusters. No other features produced class clusters so The three features ( m,s,F ) analyzed have great compact and separated from each other. But potential for classifying electrodes leading to seizure, extrapolation of this mechanism of classification in regardless on what type of classifier used with respect time did not work as anticipated since the time to the 3 parameters. dynamics of the parameters strongly changed from one Key findings can be affirmed as follows: (1) it was recording to the other, despite visible class clustering. found that at any window of time along the EEG signal In a parameter vs. time plot, the separating points (independent of time), acceptable classifiers could be between the two electrode groups changed from one obtained using just the complexity values; (2) A search recording to another. In order to consider this relative for such decision functions across patients is change and yet make real-time classification possible, ineffectual, because experiments reveal that such time independent analysis was performed by computing decision functions are patient dependent; (3) It is for each feature three statistical parameters, namely the extremely important that when one is to search for such mean of the regression line that represents the feature decision functions, electrodes should be analyzed only behavior, the standard deviation of the parameter over if they are localized in different locations and with time and the power of the frequency spectrum of the recorded interictal spikes not happening feature over time. These statistical parameters were simultaneously. then inputted to an artificial neural network (ANN) in order to obtain a linear classifier for each feature. One decision function was created exclusively for each of 4. Conclusions the three parameters (correlation, mobility, and complexity). These specific decision functions would Surgical treatment is being used with increasing find the optimum separating plane between the two frequency for patients with intractable epilepsy. classes of electrodes in a 3D space where the axis are Operative success depends to a large degree on the represented by the statistical parameters used (mean, results of a comprehensive pre-operative patient standard deviation, and frequency power). evaluation, the main purpose of which is to delineate The network configuration used in this research the epileptogenic lesion. The likelihood of the success consist of 3 input neurons that correspond to the mean, of surgery is increased when all test results point to a standard deviation, and frequency power (,,) of the single epileptogenic focus. The unique contribution of parameter analyzed. The output would be 1 or -1, which our study is to understand better the characteristics of indicates if a given channel leads to seizure or not, the different interictal epileptiform activities. In all of respectively. these performance values of the 3 parameters The decision functions consisted of feed- implemented, it can be said that the results obtained forward ANNs trained via backpropagation. These show great promise in delineating electrodes that lead ANNs are structured with 3 input neurons and 1 output to seizure from those that do not. It is fitting to note neuron, with linear activation functions. This type of that when our results failed to discriminate between structure produces a linear classifier. these two sets of electrodes, a clinical analysis revealed that those electrodes were indeed situated in the same region and their interictal spikes were happening 3. Results simultaneously. As this study will involve a higher number of patients as they become available, additional Results of this study indicate that this EEG analysis results will provide more credence to our findings. technique allows defining two regions of electrodes, one for electrodes leading to an ictal state and another References detection using the walsh transform, IEEE Transactions on Biomedical Engineering, Vol. 51, 1. B.Greenstein & A. Greenstein, Color atlas of No. 5, May 2004, 868-873. neuroscience, Neuroanatomy and Neurophysiology 6. M. Ayala, M. Adjouadi, I. Yaylali, & P. Jayakar, (Thieme Stuttgart New York, 2000). An optimization approach to recognition of 2. F. H. Martini, Fundamentals of anatomy & epileptogenic data using neural networks with physiology (Fifth edition, New Jersey 07458, simplified input layers, Biomed. Sciences Prentice Hall, 2001). Instrumentation, Vol. 40, 2004, 181-186. 3. A.S. Gevins & A. Remond, Methods of analysis of 7. A.C.K Soong & Z.J. Koles, Principal-Component brain electrical and magnetic signals. Handbook localization of the sources of the background EEG, of Electroencephalography and Clinical Biomedical Engineering, IEEE Transactions, Nneurophysiology (Elsevier: Amsterdam, 1987). 42(1), 1995, 59-67. 4. M. Adjouadi, M. Cabrerizo, M. Ayala, D. Sanchez, 8. L. Zhukov, D. Weinstein, & C. Johnson, P. Jayakar, I. Yaylali, & A. Barreto, A new Independent component analysis for EEG source approach to the analysis of epileptogenic data using localization, IEEE Engineering in Medicine and statistically independent operators, Journal of Biology Magazine, 19(3), 2000, 87-96. Clinical Neurophysiology, Vol. 22(1), January/February 2005, 53-64. 5. M. Adjouadi, D. Sanchez, M. Cabrerizo, M. Ayala, P. Jayakar, I. Yaylali, & A. Barreto, Interictal spike