Biophysical Approaches for the Multi-System Analysis of Neural Control of Movement And

Biophysical Approaches for the Multi-System Analysis of Neural Control of Movement and

Neurologic Rehabilitation

Dissertation

Presented in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy in the

Graduate School of The Ohio State University

Sarah Hulbert

Graduate Program in Biophysics

The Ohio State University

2018

Dissertation Committee

John Buford, Advisor

Hojjat Adeli, Co-Advisor

Lynne Gauthier, Committee Member

Copyrighted by

Sarah Hulbert

2018

Abstract

The neural control of movement provides a rich testing ground for principles and approaches from the field of biophysics. Multiple centers of the brain take part in motor control. A major focus of this dissertation is on areas crucial for planning and performance of skilled reaching, the primary motor cortex, the supplementary motor area, and the pre-motor area of the cerebral cortex, and the pontomedullary reticular formation (PMRF) in the brain stem.

Neurophysiological studies are based around the electrical signals present in neurons and include techniques to record these signals and to use electricity to stimulate the nervous system to produce responses. One study in this dissertation shows that both simultaneous and offset stimulations of cortical motor areas and the PMRF produced EMG responses in the arms. Some of these patterns were indicative of simple summation of outputs from the cortical and brainstem sites, but there were also responses indicating gating of the effects from one site by the other, and even more complex interactions between the motor outputs. This suggests that the cortex and brainstem utilize a variety of pathways cooperatively for motor control during reaching.

Biophysical approaches can also be applied to the prognosis of both traditional and a gaming version of a motor restorative treatment for human beings recovering from stroke, which is called Constraint-Induced Movement Therapy (CI therapy). A second study in this dissertation shows that, by utilizing machine learning approaches, the prognosis can be determined with high accuracy from pre-therapy scores of motor function, as indicated by the Wolf Motor Function

Test. In the manner that was investigated as part of this dissertation, utilizing machine learning and specifically an Enhanced Probabilistic Neural Network, pre-therapy somatosentation did not increase the accuracy of prognosis. However, a more thorough investigation of specific facets of motor function as measured by the Wolf Motor Function Test found that baseline gross motor ability is a better predictor of therapy outcomes than baseline fine motor ability. Moreover, individuals with poor gross motor ability at pre-therapy baseline demonstrated a more beneficial rehabilitation response in both gaming and traditional CI therapy. This suggests that a person’s baseline gross motor ability may be useful as a supplementary factor in predicting which type of therapy (CI or not) is best for that person.

Finally, this dissertation shows that by exploiting information contained within electrical brain signals (EEG), movement characteristics, specifically quality of movement, can be extracted from features in the frequency domain of EEG data captured during performance of gaming CI therapy by a person with mild hemi-paresis. Unlike previous studies that have exploited features indicative of movement type, this study reveals a more nuanced characteristic of the signal that can be extracted. With the ability to predict the quality of movement, this information could be used for personalized feedback during motor restorative therapy.

After a brief introduction and background (chapters 1 and 2), each of these findings will be presented in the subsequent chapters (3-6). Chapter 7 will conclude the dissertation with a synthesis of these results and future directions

iii

Dedication

To my husband, who selflessly supported me in this endeavor.

To my family who removed all barriers to help me succeed.

To my friends, my Columbus family, who never wavered in their steadfast support.

To my advisors, Dr. Hojjat Adeli and Dr. John Buford, and committee member, Dr. Lynne

Gauthier, for their irreplaceable guidance and mentorship.

Acknowledgments

I would like to express my sincerest gratitude toward my Advisors, Dr. Hojjat Adeli and

Dr. John Buford for their guidance and mentorship throughout this entire process. Dr. Adeli, your encouragement for me to pursue something I am passionate about, and to do so through collaboration, is among the best advice I have received. It has allowed me to create meaningful research that also has a home in my heart. The wisdom and expertise you passed on to me concerning how to write and communicate research and the resources you shared with me concerning the most state-of-the-art computational techniques have been invaluable assets to my success.

Dr. Buford, I am forever thankful for your willingness to see my potential for growth and your patience as I tried and tested different things. I am also very appreciative for your willingness to always meet, each time coming to the table ready to make progress and offer guidance, regardless of how much time had passed. Above all, I am grateful for your support and words of encouragement throughout this process. In this way, you instilled in me a confidence that I did not have when I came to graduate school. Because of this, I am able to approach science with a discerning eye and defend my own work.

Thank you also to my committee member, Dr. Lynne Gauthier, for your guidance and mentorship in the neurorecovery and brain imaging lab. You pushed the limits of my programming skills and my ability to critically think through scientific projects and, in this way,

v you fostered my growth as a scientist. You also gave me the opportunity to work with and for people to find solutions for them. In this way, you helped me realize some of my true passions in life.

So, to my committee, I am truly thankful to have had the opportunity to work closely with not one, but three experts throughout my graduate school career. I am shaped by what you have taught me and am forever grateful.

Thank you also to my fellow Ph.D. students and lab partners, Alexis Ortiz-Rosario,

Ph.D., Alexis Burns, and Mohammad Rafiei, Ph.D.. Your support and guidance were invaluable throughout this process. I enjoyed every day working side-by-side with you.

I would also like to thank, from the motor systems and neurophysiology lab, Tom

Hirschauer, MD, PhD, lab technical staff , Amada Jellick and Rebecca Slattery, as well as previous students in the lab group who help initiate these studies, Wendy Herbert, PhD, PT,

Lynnette Montgomery, PhD, Jacob Banks, MS, and a previous lab tech, Stephanie Moran.

Special thanks to the professionals in ULAR at OSU who took such wonderful care of the subjects. This work was supported in part by NIH R01 NS 037822, and by grants from the provost’s office at Ohio State

Thank you also to the other members of the neurorecovery and brain imaging lab, Dr.

Alexandra Borstad, Troy Richter, Kris Kelly, Carson Herron, and all of the amazing undergraduate students I had the opportunity to work with.

Thank you to the Engineering Education Department for funding part of my stay of the

Ohio State University and giving me the opportunity to explore my love of teaching as well as research.

Finally, thank you to my husband, Jarren. Your support has been incomparable. You are my rock and the purest form of support. To my family who cheered for me and supported me from start to finish and never wavered in their belief that I could do this. And to my Ohio family;

I came to Ohio without knowing anyone and now I leave the closest and dearest, most steadfast friends I have ever had. I could not have done this without you.

vii

Vita

May, 2016 …………………………...... …M.S. in Biophysics, The Ohio State University

April, 2013………………....…………….. B.S. in Physics, Western Michigan University

December, 2015…………….……….Associates in Science, Kellogg Community College

April, 2016 – Present...... Graduate Research Associate; Neuroimaging and

Neurorecovery Lab, The Ohio State University

June, 2013-Present…………………….Graduate Research Associate; Motor Systems and

Neurophysiology Lab, The Ohio State University

August 2015-August, 2016……...…………Graduate Teaching Associate, The Ohio State

University, Department of Engineering Education

August 2016-May, 2018………….….Lead Graduate Teaching Associate, The Ohio State

University, Department of Engineering Education

viii

Publications

George, S. H., Rafiei, M. H., Borstad, A., Adeli, H., & Gauthier, L. V. (2017). Gross motor ability predicts response to upper extremity rehabilitation in chronic stroke. Behavioural Brain

Research, 333, 314-322.

George, S. H., Rafiei, M. H., Gauthier, L., Borstad, A., Buford, J. A., & Adeli, H. (2017).

Computer-aided prediction of extent of motor recovery following constraint-induced movement therapy in chronic stroke. Behavioural Brain Research, 329, 191-199.

Hulbert, S., & Adeli, H. (2015). Spotting psychopaths using technology. Reviews in the neurosciences, 26(6), 721-732.

Hulbert, S., & Adeli, H. (2013). EEG/MEG-and imaging-based diagnosis of Alzheimer’s disease. Reviews in the neurosciences, 24(6), 563-576.

Fields of Study

Major Field: Biophysics

Table of Contents

Abstract ...... ii Dedication ...... iv Acknowledgments ...... v Vita ...... viii List of Tables ...... xiii List of Figures ...... xiv CHAPTER 1. INTRODUCTION ...... 1 1. Problem Description and Significance...... 1 2. Research Objectives and Chapter Outline ...... 3 CHAPTER 2: BACKGROUND AND LITERATURE REVIEW ...... 6 Introduction: ...... 6 Recent Advances in Neurophysiology and the Accompanying Biophysical tools of the trade: 8 Motor Imagery—Decoding imagined movements for interaction with BCI ...... 10 Movement Decoding—Decoding physical movements with implications for rehabilitation .. 11 Movement Intent—Predicting movements: A step toward real-time neurofeedback ...... 13 From Basic Science to the Clinic: Gaming, motion capture, virtual reality and machine learning for Motor Rehab and other Applications ...... 14 Future: Prediction Models for therapy and Decoding Quality “Intent” ...... 15 CHAPTER 3: INTERACTIONS BETWEEN CORTICOSPINAL AND RETICULOSPINAL OUTPUTS DETERMINE MUSCLE RESPONSE IN THE UPPER LIMBS AS REVEALED WITH STIMULUS-TRIGGERED AVERAGING ...... 17 Abstract ...... 17 Introduction ...... 19 Methods...... 21 Subjects and task ...... 21 Stimulation Techniques ...... 22 EMG Integrity Tests ...... 24 x

Data Analysis: ...... 25 Results ...... 27 General Characteristics of post-stimulus effects ...... 27 Paired Pulse ...... 28 Do Brain Region Pairs Influence Interaction Response Type? ...... 29 Discussion ...... 30 Summary and Conclusions ...... 35 CHAPTER 4. COMPUTER-AIDED PREDICTION OF EXTENT OF MOTOR RECOVERY FOLLOWING CONSTRAINT-INDUCED MOVEMENT THERAPY IN CHRONIC STROKE ...... 44 Abstract ...... 44 Introduction ...... 44 Methods ...... 49 Participants...... 49 Intervention ...... 49 Clinical measures ...... 50 Architecture of the prediction models ...... 51 Data collection, description, and preparation ...... 54 Training and testing of the model ...... 56 Sensitivity analysis via EPNN ...... 56 Results ...... 57 Discussion ...... 61 Conclusion ...... 64 Acknowledgements ...... 64 Addition ...... 65 Probabilistic and Enhanced Probabilistic Neural Networks ...... 65 Log-likelihood ratio ...... 66 CHAPTER 5: GROSS MOTOR ABILITY PREDICTS RESPONSE TO UPPER EXTREMITY REHABILITATION IN CHRONIC STROKE...... 68 Abstract ...... 68 Introduction ...... 69 Methodology ...... 71 Participants ...... 71

Intervention ...... 72 Outcome measures ...... 74 Sensitivity analysis and prognosis model ...... 78 Implementation ...... 79 Results ...... 80 Accuracy of the gaming models and rates of selection ...... 80 Robust predictors across intervention type ...... 85 Parsimonious combinations ...... 87 Sensitivity analysis on the parsimonious combinations ...... 87 Discussion ...... 89 Study limitations ...... 92 Conclusion ...... 93 Acknowledgements ...... 93 CHAPTER 6: DECODING NEURAL SIGNATURES ASSOCIATED WITH THE QUALITY OF ARM MOVEMENTS USING HETEROGENEOUS KINETIC AND EEG DATA CONCURRENTLY ...... 95 Abstract ...... 95 Introduction ...... 96 Methods...... 100 Results ...... 107 Discussion ...... 108 Study limitations ...... 112 Conclusion ...... 114 CHAPTER 7: SYNTHESIS AND RECOMMENDATIONS ...... 124 References ...... 129

xii

List of Tables

Table 1 Demographic and clinical characteristics of 35 patients ...... 49 Table 2 Behavioral measures used in the prognostic model ...... 50 Table 3 Patient data used to train and test the EPNN model ...... 55 Table 4 Effectiveness rank of the 18 inputs on the results based on the sensitivity analysis using the EPNN model ...... 58 Table 5 Different combinations of 18 inputs for EPNN resulting in an average accuracy of 100% ...... 59 Table 6: Demographic and clinical characteristics of 19 participants in gaming therapy...... 72 Table 7: Behavioral measures used in the prognostic model ...... 75 Table 8: Participant data used to train and test the EPNN model ...... 77 Table 9: Different combinations of 18 predictors for EPNN resulting in an average accuracy of 94.7% in the current study (1: selected; 0: not-selected) ...... 81 Table 10: Comparison of the rates of selection of predictors for the combinations with accuracy of about 90% and greater using the CI and gaming CI therapies ...... 83 Table 11: Rates of selection of 18 predictors in the prognosis model for the gaming and CI therapies, and the combined approach ...... 84

xiii

List of Figures

Figure 1: Illustration of instructed delay, bilateral reaching set-up and accompanying sketch of a reaching trial. Adapted from Davidson and Buford 2006...... 36 Figure 2: Illustration and description of the stimulation paradigms used to during single-pulse microstimulation of the cortex and PMRF and paired-pulse microstimulation of the two regions...... 36 Figure 3: Process for creating actual averages and simulated averages for paired pulse stimulation...... 37 Figure 4: Cortical stimulation sites ...... 38 Figure 5: Locations of stimulation sites in the pontomedullary reticular formation...... 39 Figure 6: A) Ratio of ipsilateral, bilateral, and contralateral muscle responses during single-pulse microstimulation of M1, PMd, and SMA B) Ratio of ipsilateral, bilateral, and contralateral muscle responses during single-pulse microstimulation of the left and right PMRF...... 40 Figure 7: Representative examples of EMG data recorded during both single (RF Alone and CX Alone) and paired pulse (Latencies 1 -3) microstimulation. A red trace indicates a facilitation response and dark blue represents a suppression response. The light blue shown during paired pulse stimulation represents the simulated summation response. In paired stimulation, Time 0 always indicates the time of the RF stimulus; the time of the cortical stimulus. Panel A shows a typical example of a PMRF driven response. Panel B is an example of an EMG response likely driven by summation of the PMRF and CX signals. Panel C shows an example of a gating interaction where the PMRF seems to gate out the CX contribution during simultaneous paired stimulation...... 41 Figure 8: Number of different types of paired responses based on which cortical area was stimulated (Black bars = M1, striped bars = PMd, and white bars = SMA) and the side of stimulation of the PMRF divided left to right in each box. All of the boxes on the left side of the figure show the number of responses obtained from muscles on the left side of the body and all of the boxes on the right show the number of responses obtained from muscles in the right side of the body...... 42 Figure 9: Simple pathways between the cortex and PMRF lend themselves to interact in a simple, summative way. More complex circuitry between pre- and supplementary motor areas and the PMRF lend themselves to gating and other complex interactions...... 43 Figure 10: This schematic representation of motor learning illustrates how diminished intrinsic somatosensory feedback following stroke could reduce error detection and motor learning...... 47 Figure 11: Architecture of the prognostic EPNN model used to predict the extent of motor recovery after CI therapy...... 52 Figure 12: The histogram of the patients’ change in the natural log of WMFT scores from pre to post CI therapy...... 53 xiv

Figure 13: Depiction of the gaming environment ...... 73 Figure 14: : The histogram of the participants' change in the natural log of WMFT scores from pre to post gaming therapy ...... 76 Figure 15: Flowchart including the sensitivity analysis ...... 78 Figure 16: a) Change in average of natural logarithm of all WMFT gross motor tests from pre therapy to post therapy (R2 = 0.55) versus average of natural logarithm of all WMFT gross motor tests before therapy; b) Change in average of natural logarithm of ...... 85 Figure 17: Change in both fine and gross motor performance as a function of initial motor performance for each task...... 86 Figure 18: Screen shot of Recovery Rapids ...... 115 Figure 19: Flowchart of methodologies ...... 116 Figure 20: Skeleton coordinates captured by Kinect ...... 117 Figure 21: Gesture parsing for a rowing gesture ...... 118 Figure 22: Map of 64 EEG electrodes. Highlighted electrodes were used for analysis in this research ...... 119 Figure 23: Normalcy Residuals for the row, bottle, parachute, and raisearm gestures The red curve indicates the expected performance for the individual. The residuals indicated the actual performance, as compared to "normal”...... 120 Figure 24: Illustration of EEG processing...... 121 Figure 25: Two-layer deep Boltzmann machine for feature extraction consisting of an encoder and decoder ...... 122 Figure 26: Architecture of PNN/EPNN...... 123

CHAPTER 1. INTRODUCTION

1. PROBLEM DESCRIPTION AND SIGNIFICANCE

The brain leaves no shortage of questions to ask as the scientific community attempts to fully discern its functionality and overall role in our lives. This dissertation covers a small section of that functionality as it relates to movement, particularly of the upper extremities. In this area, questions painted with the broadest stroke include, how is it that the brain coordinates and orchestrates any movement at all, let alone complex, coordinated movements? How is it that some brains allow some people to move well and other brains do not? As attempts are made to answer these types of questions, this organ continues to pose a huge challenge to research efforts due to the very complexity that allows such a myriad of questions to be asked.

Therefore, the brain has often only been studied one small piece at a time -- one region at a time, on one scale at a time, with one type of data collection or measurement at a time. What can follow are conclusions, recommendations, and suggestions for how to advance the field of neuroscience, physiology, and every other “brain” science that asks these questions, based on these single piece, single system analyses, when, in fact, these fragments are but small pieces of the hugely complex puzzle that is the brain. Fortunately, technological advances and new fields of study have emerged that allow us to tackle more pieces of the puzzle at a time. In particular, the field of biophysics stands at the juncture of mathematics, physics, and biology and it offers

1 the technical, objective, automated tools and viewpoints to approach the brain using multi- system analysis as it has not been in the past.

Here, I take that multi-system, biophysical approach to studying the brain, specifically as it relates to motor control and rehabilitation thereof. Instead of investigating motor control one piece at a time, I perform a series of multi-system, multi-method experiments to investigate how various parts of the nervous system operate in harmony to produce movement. Each experiment builds on the last and, taken together, they span several levels of investigation. Specifically, at the level of the brain, I investigate how different motor regions, including the motor cortex and pontomedullary reticular formation (PMRF) in the brain stem, interact in motor control. At the level of the individual, I couple EEG data with motion capture data to predict the quality of motor control before the onset of movement and illustrate the findings with a case study. Finally, at the population level, I determine which characteristics of a person’s motor control, including their sensory ability, gross motor ability, and fine motor ability, can predict the extent of their motor function restoration after stroke. Each of these items are broken down further in the chapter outline provided below.

Taken together, the results of these experiments shine a brighter, more complete, light on our understanding of motor control as directed by the brain. This is important because, when we can understand more completely how different regions of the brain interact in motor control, we can approach rehabilitation efforts from a more informed, physiological standpoint. Then, during motor restorative rehabilitation, when we can predict how well a person is going to move before they move, we can give them the associated, real-time, personalized neurofeedback. This may ultimately serve to make their rehabilitation more efficient. Moreover, when we can predict if 2 someone will experience motor restoration or not due to a motor restorative therapy, this can inform clinician’s decisions about which type of therapy their client is most likely to benefit from before that patient even starts therapy. This, in turn, may make the client’s rehabilitation efforts more effective.

2. RESEARCH OBJECTIVES AND CHAPTER OUTLINE

Therefore, this dissertation will encompass each of the above mentioned experiments and be laid out in the following way. In chapter 2, background information providing the foundations upon which this dissertation was built will be presented. I will begin with a review of the recent advances in our understanding of the neurophysiological underpinnings of motor control though the lens of biophysics. Then I will discuss the recent technologies and clinical practices being used to address motor control, specifically in motor impaired individuals.

In chapter 3, I will explore some physiology, and consider that what has been accepted as the main motor control center of the brain, the motor cortex, is not the whole story. We will investigate a more primordial system that is anything but obsolete---the pontomedulary reticular formation (PMRF) in the brain stem. This region is well known as the center for postural control, but it is currently being investigated for its ability to coordinate with the motor cortex to ultimately produce the movements we need to make to live our daily lives (Hirschauer & Buford,

2015; Ortiz-Rosario, Berrios-Torres, Adeli, & Buford, 2014; Zaaimi, Dean, & Baker, 2017).

Specifically, I will demonstrate that stimulus triggered averaging of electromyographic (EMG) signals reveals several types of interactions between the motor cortex and the PMRF. The interactions include simple summation, gating, and more complex interactions. Of course, this

3 has implications for neurorehabilitation and neuroplasticity, in that, should one of these regions be injured, such as by a stroke, the other may, over time and especially with neurorehabilitation, be able to “pick up the slack” and the injured region may begin to make plastic changes. This discussion will take place in the form of a research article titled, “Interactions between corticospinal and reticulospinal outputs determine muscle response in the upper limbs as revealed with stimulus-triggered averaging”.

Chapters four and five will bring us into the clinic as we begin to understand rehabilitation for motor impairments, particularly within the chronic stroke population. We will explore a new type of rehabilitation that is based on game play, rather than the traditional, gold- standard, constraint-induced (CI) motor restorative therapy. Our driving question will be, can we predict how well a person will respond to this type of motor restorative therapy, and do what role do baseline somatosensory ability and motor ability play in this prediction? If we can answer that question, then we can make more informed decisions with and for survivors of motor impairments as to the type of therapy they need before they begin treatment. This investigation will be presented in the form of two papers published in 2017, “Computer-aided prediction of extent of motor recovery following constraint-induced movement therapy in chronic stroke” and

“Gross motor ability predicts response to upper extremity rehabilitation in chronic stroke”.

Next, chapter 6 will take the question of neurorehabilitation fully in hand. We move in the natural progression from an understanding of the motor neurophysiology in chapter 2, to and understanding of current, cutting-edge motor-rehabilitation techniques and exploration of prediction in treatment. Now we step fully into neurorehabilitation by taking those same motor restorative techniques and adding a brain imaging component, EEG, with the ultimate goal to 4 provide real time, personalized, neuro-feedback based on an individual’s brains waves and their rehabilitation performance. We will take the first step toward this by asking, “Can we find characteristic features of the neural signal associated with “good” and “poor” quality movement within an EEG signal?” We will address this question with a combination of EEG recording, motion capture, kinematic analysis and machine learning.

Finally, concluding chapter 7 will synthesize the previous chapters by examining how these experiments have advanced the field in a meaningful way, Overall conclusions are presented along with a look toward future directions for the field of neurorehabilitation.

CHAPTER 2: BACKGROUND AND LITERATURE REVIEW

INTRODUCTION:

Neurophysiology and biophysics are perfect examples of complementary and, indeed, overlapping fields. It may be said that none of neurophysiology can be performed without some knowledge of the biophysical underpinnings. From the basic science of neurophysiological experiments, to the tools we use to conduct those experiments, to the data analytics we use to obtain our results—all of these things benefit from the understanding and use of biophysical approaches. These benefits include more accurate results that rely less on subjective analysis, more efficient data processing, and the ability to handle larger data sets (e.g.: (Rowe, Robinson,

& Rennie, 2004)). Moreover, with the tools employed by, and developed from a biophysical viewpoint, one can naturally navigate between research questions stemming from the basic science of neurophysiology to the clinical realm of neurorecovery and rehabilitation.

Neurophysiology is the practice of understanding the functions of the brain and biophysics can broadly be defined as the application of physics to biological problems. The bridging of these two fields becomes almost unavoidable because the brain and body are, inherently, physical and, in some cases, electrical objects. Therefore, they are subject to the same laws of physics that govern any other animate and inanimate objects. Thus, to truly reveal the full functionality of the brain, as neurophysiology seeks to answer, biophysics is a well poised ally to solve the puzzle.

Over the past couple of decades, the fields of neurophysiology, motor recovery, and neurorehabilitation have seen some impressive strides, in large part thanks to this new alliance and collaboration between these fields and biophysics. This alliance has resulted in expanding the basic science of motor systems and the neurophysiology of movement (Fisher, Zaaimi, &

Baker, 2012; Hirschauer, Adeli, & Buford, 2015a; Ortiz-Rosario et al., 2014; Riddle & Baker,

2010), to employing new advanced technologies and cutting edge computational approaches for motor restorative rehabilitation in the clinic (Bird et al., 2017; L. V. Gauthier, Kane, Borstad,

Strahl, Uswatte, Taub, Morris, Hall, Arakelian, & Mark, 2017a; Laver et al., 2018; Llorens, Noe,

Colomer, & Alcaniz, 2015; Standen et al., 2017).

Therefore, this review will be broken into two sections: 1) Advances in understanding of neurophysiology and motor systems for control of the upper limbs through the lens of biophysics. In this section, I will explore how electrical properties of these systems can be exploited for the direct investigation and functional understanding of the brain through stimulation and recording via devices including micro-electrodes and EEG caps. I will also be particularly concerned with the expansion of investigating multiple motor regions in the brain as many novel advances have been made in this area. 2) Application of advanced technologies and computational approaches in neuro-motor rehabilitation and neurorecovery. This section will include examples of new electroencephalogram (EEG) technologies, brain machine interface and robotics, and gaming and virtual reality technologies for detection and decoding of motor imagery, movement intent, movement, and neurofeedback. The review will conclude by addressing how these technologies and methods are contributing to the advancement of motor

7 rehabilitation, explore some of their strengths and shortcomings, and conclude with recommendations for future directions utilizing this new toolset.

RECENT ADVANCES IN NEUROPHYSIOLOGY AND THE

ACCOMPANYING BIOPHYSICAL TOOLS OF THE TRADE:

Neurophysiology has recently expanded to multi-region investigation and analysis.

Specifically, simultaneous exploration and modeling of multiple brain regions and how they interact has been recently been demonstrated (Baker, Zaaimi, Fisher, Edgley, & Soteropoulos,

2015; Fisher et al., 2012; Ortiz-Rosario et al., 2014). This, of course, lends itself closer to the reality of how the brain functions as a single organ, not as individual areas that previous experiments would investigate. For example, as it relates to motor control, the field has expanded the focus from investigating only the motor cortex, out to the supplementary motor areas, premotor areas, and even further into the more primordial motor regions in the brain stem, specifically, the pontomedullary reticular formation. Recently, Zaaimi et al. (2017) investigated contributions to synergistic movements in the upper limbs from M1, PMRF, and the spinal cord utilizing single-pulse stimulation. Specifically, they report that their results demonstrate that all three regions are important for motor control, but that, “the [PM] RF represents approximate patterns of muscle co-activation, whereas M1 and [spinal cord] permit more nuanced, selective activity, which underlies the fine fractionated movements so important to the primate upper limb.” I will explore the interactions between the motor cortex and the PMRF in detail in chapter 3.

Accompanying this expanded investigation of brain systems, are the biophysical tools that make it possible. For example, the capabilities for electrically stimulating the brain, or simply recording the inherent electrical activity from it, have become further reaching, more precise, and more accurate through the development of more and more advanced microelectrodes and micro-electrode arrays (Gabran et al., 2013). In the early 1980’s, seminal work by

Georgeopoulous (1982) and his group translated the findings of Mountcastle (2013) from posterior parietal cortex that demonstrated directional tuning for perception and discovered that a similar pattern of neuronal encoding was present in the frontal lobe for planning and executing performance. From the original painstaking single electrode, single unit studies, the field has progressed to multi- electrode arrays that can extract the movement code in real time from populations of neurons to control robotic arms and neural prostheses (Ortiz-Rosario & Adeli,

2013). As far as this technology has come, however, the problem of long term implantation of recording electrodes in the human brain remains a challenge. There have been many attempts to discover non-invasive technologies that could extract enough information to provide functional control of a neural prostheses. Among these, EEG is perhaps most directly related to the actual electrical signals of the neurons that have been so carefully decoded in the preceding efforts with microelectrodes in the brain. However, the EEG signal represents a summation of many thousands of cells, and so extracting information in sufficient detail and deducing the original location of the signal can be extremetly challenging. Nonetheless, the information contained in the EEG does come from the neural activity, so this remains a promising signal for further exploration.

MOTOR IMAGERY—DECODING IMAGINED MOVEMENTS FOR

INTERACTION WITH BCI

Motor imagery involves imagining your body making movements and is often employed either as a supplement to motor rehabilitation or in conjunction with a brain computer interface

(BCI). In an early study using invasive micro-electrode arrays in non-human primates, Schwartz and colleagues (Velliste, Perel, Spalding, Whitford, & Schwartz, 2008) focused on the premotor cortex, and originally expected to train animals to control a robotic arm in parallel with their own limb. Much to their surprise, animals quickly and spontaneously elected to leave their own limb at rest and simply move the robotic arm based on pre-motor cortical activity. In retrospect, this made sense in that it gave the monkey fewer degrees of freedom to manage with the population of neurons that were sampled for the electrode array. In essence, it seems that the subjects found it easier to control the robotic arm by imagining through motor imagery that they were moving their arm without actually moving.

In several studies, motor imagery has been validated as a means for supplementary motor rehabilitation, and for providing motor functionality through BCI and robotics in humans (Frolov et al., 2017). For example, this technique has been shown to increase a person’s overall motor ability after stroke (Polli et al., 2017). Because of the utility of motor imagery, researchers are also investigating the associated neural activity. For example, Grant et al. (2012) addressed the feasibility of detecting and classifying neural patterns associated with motor imagery on a single- trial basis within EEG data. Specifically, they utilized Emotiv’s ® EPOC EEG headset to detect mental commands including push, pull, rotate, lift and drop in real-time. This form of motor imagery was used to assess the accuracy of Emotiv’s Cognitive software suite as a real-time 10 classifier. For 57 healthy participants, the author’s report that the accuracy of the classifier was significantly greater than chance (87.5%) and equivalent for all mental commands. Apart from high classification accuracy, this software also provides real-time feedback for the user, thereby increasing the utility of the product. Being able to accurately parse out mental commands in a non-invasive way has far reaching applications in the realms of BCI and other machine computer interface, such as robotic prosthesis. This is particularly true in persons that do not have enough residual movement or motor control to perform tasks for rehabilitation.

MOVEMENT DECODING—DECODING PHYSICAL MOVEMENTS

WITH IMPLICATIONS FOR REHABILITATION

Another one of the great strides in understanding motor control and motor rehabilitation is the ability to use EEG technology on non-stationary subjects and research designs (C.

Kranczioch, Zich, Schierholz, & Sterr, 2014). In the past, movement of a participant undergoing

EEG recording was impracticle because it overwhelmed the signal with movement artifacts and noise. With advances in EEG hardware (e.g.: wireless, dry electrodes) and software (Debener,

Minow, Emkes, Gandras, & Vos, 2012), these barriers to EEG recording are being addressed, making room for neural analyses of movement.

For example, Bradberry et al. (2010) investigated the continuous reconstruction of 3D hand movements from EEG signals obtained from 5 healthy, right handed individuals. They utilized a 64-sensor Electro-Cap to record 58 channels of data (minus 3 frontal channels to reduce the influence of eye movements for a total of 55). Sensors were placed according to the international 10-20 system. After pre-processing, a total of 34 of the channels were used for 11 analysis. Standardized low-resolution tomography (sLORETA) was used to estimate the location of the neural sources responsible for hand-velocity encoding. The author’s reported that the CP3 sensor, which is typically located over the primary sensorimotor cortex, in all cases presented here, contralateral to the reaching hand, provided the greatest contribution. They further reported that EEG recorded at 60 ms time lag encoded the most information. A negative correlation between decoding accuracy and movement variability was also found. Decoding accuracy was quantified by computing the correlation coefficients between measured and reconstructed hand velocities. The peak correlation coefficients for decoding velocities in the x, y and z direction were 0.19, 0.38, and 0.32, respectively.

Similarly, Ofner et al. (2012) attempted to decode arm velocities and positions of arm movements from EEG from 5 healthy, right-handed individuals. They postulated that this information would be useful to control a neuroprosthesis in real time. The experimenters recorded from 49 electrodes over sensorimotor and frontal areas and 3 electrooculography electrodes while subjects performed continuous movement with their dominant arm. The authors did not report use of the international 10-20 system. In order to suppress eye movements, a fixation cross was presented to subjects to focus their gaze while they performed the task.

Furthermore, the first 5 seconds of each trial was removed from analysis to omit movement onset. Position tracking was completed using a Kinect sensor. They reported mean correlation coefficients between the measured and decoded hand velocities in the x, y, and z direction to be

0.70, 0.77, and 0.62 respectively. Note that these values were much higher than the values found by Bradberry et al. (Bradberry et al., 2010) just two years before. The author’s further reported that EEG recorded between 35 ms and 82 ms encoded the most information. This is consistent

12 with Bradberry et al. (Bradberry et al., 2010). They also reported the correlation coefficients for hand positions in the x, y and z direction to be 0.70, 0.78, and 0.62, respectively. Here they reported peak contributions at 12ms and 105ms time lags, with low contributions in between. The authors pointed out that one difference between their work and the work of

Bradberry et al. (Bradberry et al., 2010) was that they did not require subjects to reach for targets. Rather, they simply were asked to move their hand in space. They postulated that this may have contributed to the differences in correlation coefficients along with controlling for movement onset in the first 5 seconds of each trial. Finally, they discussed the issue that smooth, rather than “jerky”, movement (as show in (Lv, Li, & Gu, 2010) can help improve decoding accuracy.

MOVEMENT INTENT—PREDICTING MOVEMENTS: A STEP

TOWARD REAL-TIME NEUROFEEDBACK

From decoding imaginary movements to decoding actual movements, the next question follows, can we predict the next movement? Put another way, can we detect movement intent?

Ibanez et al. (2011) demonstrated the feasibility of this idea in both healthy and stroke participants who performed “self-paced” movements. The group recorded from 6 surface EMG locations on the arms and 31 EEG channels according to the International 10-20 system. They reported that their model for detecting movement intent before movement onset was able to accurately detect 74.5% of movements in healthy participants, and 82.2% of the movements in stroke participants with classification happening every 125 ms. Similarly, Lew et. al (2015) investigated movement intent in both healthy and stroke participants performing self-paced 13 movements. One of the major results of this study was that, utilizing the readiness potential, movement intent was detected as early as 500ms, and peaking around 167ms before movement onset. For all except one participant, classification accuracies for movement intent were around

76%, with average classification rates for each participant ranging from 69% to 85%. This variability may point to much needed personalization in neuro-prosthetic and neurorehabilitation efforts.

FROM BASIC SCIENCE TO THE CLINIC: GAMING, MOTION

CAPTURE, VIRTUAL REALITY AND MACHINE LEARNING FOR

MOTOR REHAB AND OTHER APPLICATIONS

As virtual reality and other immersive technologies begin to permeate day to day life, so too does it find a home in the arenas of therapy, recovery and rehabilitation. These uses range from chronic pain management utilizing pain distraction (Gromala, Tong, Choo, Karamnejad, &

Shaw, 2015), to balance training (Llorens et al., 2015) , to exposure-based treatment for post- traumatic stress disorder (Botella, Serrano, Banos, & Garcia-Palacios, 2015), and empowerment interventions for cancer patients, including children (Govender, Bowen, German, Bulaj, &

Bruggers, 2015). One area where this technology really seems to have taken root is in motor rehabilitation (Anderson, Woodbury, Phillips, & Gauthier, 2015; A. L. Borstad et al., 2018; L. V.

Gauthier, Kane, Borstad, Strahl, Uswatte, Taub, Morris, Hall, Arakelian, & Mark, 2017b; Levin,

Weiss, & Keshner, 2015). More and more, motor rehabilitation is being enhanced with technologically advanced measurements and data collection devices, such as EEG and motion

14 capture, and traditional therapy is being supplemented or replaced with virtual reality (VR) and other game-based platforms. Moreover, these therapies are finding their way from the clinic into homes, making them more accessible and easy to use.

For example, it has been shown that game-based therapies are just as effective as traditional versions. An ongoing review published in Stroke (Laver et al., 2018) shows that, for post-stroke participants, there is no significant difference in improvement between someone who underwent therapy and those that underwent a VR or game-based therapy. Moreover, they report that, for those participants who had VR or game based therapy in addition to traditional therapy, resulting in more therapy hours, they showed significant improvement in their upper limb functionality. Similarly, work by Bird et al. (2017) also reported no group differences between stroke participant’s that underwent traditional physical therapy vs. those that played a motion- capture rehabilitation game.

Not only are these game-based therapies on par with traditional therapy, in terms of patient outcomes, but they also tend to be more accessible. For example, a relatively new company, Games That Move You, PBC, has created an in-home gaming platform for motor restoration.(Maung et al., 2014) that has led to promising results for tele-rehabilitation (A. L.

Borstad et al., 2018).

FUTURE: PREDICTION MODELS FOR THERAPY AND DECODING

QUALITY “INTENT”

Since the proof of principle for decoding imaginary movement, actual movement, and movement intent from neural signals has been thoroughly demonstrated, the next question 15 becomes, can we predict more nuanced but specific attributes or characteristics of the upcoming movement? Specifically, can we predict the quality of the upcoming movement? By utilizing a newly developed gaming rehabilitation platform, Recovery Rapids, coupled with state-of-the-art machine learning techniques, I will investigate this question in the form of a case-study.

Taken all together, these recent advances in neurophysiology and neurorehabilitation provide a solid foundation on which biophysical analysis of multiple systems can be built.

Without these foundational and seminal studies that provide a starting point from which to understand how individual areas of the brain participate in muscle recruitment, there would be little of hope of investigating several areas at once. Similarly, the advances in technological capability and the ground work laid to demonstrate the extent of information that can be extracted from EEG signals thus far, allow us now to go deeper; to stretch the capacities of both machine and signal to search for and extract even information that has the opportunity to increase a person’s quality of life, through rehabilitation and neurofeedback.

CHAPTER 3: INTERACTIONS BETWEEN CORTICOSPINAL AND RETICULOSPINAL OUTPUTS DETERMINE MUSCLE RESPONSE IN THE UPPER LIMBS AS REVEALED WITH STIMULUS-TRIGGERED AVERAGING

ABSTRACT

Corticospinal and reticulospinal tracts each contribute to motor control of the upper limbs. Using stimulus-triggered averaging (StimTA) of EMG responses elicited by combined stimulation of the motor cortex and pontomedullary reticular formation (PMRF), we tested the hypothesis that these regions also interact to produce effects in the upper limbs. In three non- human primates, M. fascicularis, three areas of the motor cortex-primary motor cortex (M1), supplementary motor area (SMA), and dorsal premotor area (PMd)-were electrically stimulated at varying time shifts with respect to the electrical stimulation of either the right or left PMRF during a bilateral reaching task. Specifically, three paired pulse paradigms were used: 1) cortical area stimulated before PMRF, 2) both areas stimulated simultaneously, 3) PMRF stimulated before the cortical area. The subjects were implanted with chronic EMG electrodes in 12 bilateral pairs of muscles in the upper limbs. We compared the EMG responses evoked from the individual stimulation of the cortex or the PMRF to EMG responses elicited from the paired pulse paradigms. We identified EMG response patterns that were indicative of dominant contribution from the PMRF or the cortex alone as well as patterns indicative of an interaction between the two brain regions. Results show that 1) outputs from the reticulospinal and corticospinal systems did combine in the upper limbs via cooperation (simple summation), competition (gating), and other complex interactions, 2) interaction type (summation, gating, or complex) was dependent on where in the cortex (M1, SMA, or PMd) and brainstem (ipsi- or 17 contra-lateral) paired stimulation was applied and 3) the majority of responses occurred contralateral to the cortical region being stimulated, but there were several instances of EMG responses in the ipsilateral limb as well. Taken together, these results show that corticospinal and reticulospinal outputs can interact to produce muscle recruitment. Therefore, a systems approach is needed to fully understand the outputs of these motor systems. Furthermore, by advancing our understanding of this system when it is in a healthy, functioning state, we will be better equipped to understand the system when it becomes injured, such as in stroke. This enhanced understanding can lead to more efficient, targeted, and overall improved therapies for the injured system.

Keywords: Pontomedullary reticular formation, motor cortex, reaching, stimulus-triggered averaging

INTRODUCTION

Understanding neural control of movement is foundational for developing effective treatments and therapies for stroke rehabilitation. To date, several brain regions have been identified as playing a role in the neural control of movement, however, few studies have attempted to analyze the interactions among these regions during movement.

The most well-known of these regions is the motor cortex. Much is known about the role of the motor cortex, including the primary motor cortex (M1), dorsal premotor cortex (PMd), and supplementary motor area (SMA), in the recruitment of upper limb muscles for activities such as reaching. Indeed, one can trace the neurological representation of both arms and hands in the motor cortex going all the way back to Penfield and his homunculus in 1937 (Penfield et al.

1937). Traditionally, areas of the motor cortex are known to recruit muscles on the contralateral side of the body (Asanuma, 1973; Kasser & Cheney, 1985). More recently, findings also support ipsilateral and/or bilateral recruitment of muscles from both the supplementary motor area

(SMA) as well as the primary motor cortex (M1) (Brinkman C, 1979; Kermadi, CALCIATI, &

ROUILLER, 1998; Lacroix et al., 2004; Montgomery, Herbert, & Buford, 2013a; Tanji, Okano,

& Sato, 1988).

Another brain region implicated in the control of movement is the pontomedullary reticular formation (PMRF). Various studies have revealed its role in bilateral control of the upper limbs and postural adjustments in the cat (Drew & Rossignol, 1984a; Schepens & Drew,

2006) and the primate (Davidson & Buford, 2006; Hirschauer & Buford, 2015; Soteropoulos,

Williams, & Baker, 2012). Moreover, unlike the mapping of the cortex illustrating both proximal and distal muscle representation, stimulation studies of the brainstem show predominately

19 proximal muscle representation (Davidson & Buford, 2004). There is some evidence for influence over distal movements as well (Soteropoulos et al., 2012) but proximal control seems most prevalent. Therefore, to study the interaction between the motor cortex and brainstem, the present study is focused primarily on EMG responses in proximal muscles.

Even with these distinct brain regions known to participate in motor control of the upper limbs, few studies have been conducted to elucidate any physical circuitry and pathways between the motor cortex and the PMRF and what role those pathways play in motor function

(Matsuyama & Drew, 1997; Ortiz-Rosario et al., 2014). One method that has been used to study these connections is stimulation of cortical neurons during simultaneous intracellular recording in the RF either directly (Canedo & Lamas, 1993; He & Wu, 1985; Jankowska & Edgley, 2006;

Peterson, Anderson, & Filion, 1974) or through the use of transcranial magnetic stimulation

(TMS) (Fisher et al., 2012). While these studies provide evidence of pathways between the cortex and PMRF, the detail available from these approaches is limited. In other studies, there is some evidence to support the hypothesis that cortical and reticular formation interactions play a role in trunk and limb muscle recruitment (Luccarini, Gahery, & Pompeiano, 1990a; Massion,

1992; Schepens & Drew, 2003). In particular, these studies provide compelling, yet indirect, clinical or EMG evidence that the systems work together to create coordinating postural adjustments and limb movement.

Taken together, these studies show that both the motor cortex and PMRF play a role in motor control for the upper limbs and that they must interact to produce muscle recruitment.

What remains un-investigated is how these regions interact. What this suggests is the need for a systems approach to fully define the nature of the interactions that produce cohesive neural

20 control of movement from these two motor regions. In this report we demonstrate that these regions do indeed interact to produce muscle recruitment in the upper limbs. Using stimulus triggered averaging (StimTA) of EMG responses, we reveal several types of interactions between the systems. Specifically, we demonstrate that the interactions can be cooperative, such as a summative response, competitive, such as a gating response, or a more complex type of interaction. Portions of these data were described previously in an abstract (Hulbert et al. 2015).

METHODS

Subjects and task

Three male Macaca facicularis monkeys, H, N, and O, were trained to perform an instructed delay, bilateral reaching task administered by Tempo software (Reflective Computing,

St. Louis, MO, USA) that was previously described (Davidson & Buford, 2006). Briefly, from a sitting position in a primate chair, the subjects held down left and right switches which, after a brief delay, triggered the appearance of target boxes displayed on a touch screen in front of them

(Fig. 1). Based on the color of the box, the subjects reached for and touched the box with either their left or right hand. Subjects were rewarded with banana pellets or other treats for completion of successful trials. This design facilitated bilateral EMG activity in the upper limbs and trunk.

Similar surgical procedures describing implantation of recording chambers and EMG were reported previously in (Montgomery, Herbert, & Buford, 2013b). In aseptic conditions, recording chambers were mounted over craniotomies over the left motor cortex (Horsley Clarke coordinates AP 15 ML -12) and right occipito-parietal region tiled 10° from the sagittal plane, pointed at (AP 0 ML 0 and DV -12) for access to the pontomedullary reticular formation

(PMRF). These coordinates were chosen based on previous experience in the lab and the atlas published by (Szabo & Cowan, 1984). During surgery, subjects were placed under isoflurane anesthesia and vital signs were monitored. They were given analgesics (buprenorphine and

Ibuprophen), antibiotics (chloramphenicol or baytril), and extra enrichment during a recovery period of several days.

Teflon-coated stainless steel EMG electrode wires were implanted subcutaneously during the surgery within 12 pairs of upper limb muscles bilaterally, including: flexor carpi ulnaris, extensor carpi radialis, biceps brachii, triceps brachii, middle deltoid, latissimus dorsi, lumbar paraspinals, supraspinatus, upper trapezius, pectoralis major, sternocleidomastoid, and cervical paraspinals. The connector for the EMG electrodes was included in the cranial implant.

Subject care complied with the NIH Guide for the Care and Use of Laboratory Animals and the institutionally approved animal care protocol for our laboratory.

Stimulation Techniques

Using coordinate grids within the recording chambers, tungsten microelectrodes were guided through stainless steel cannulas and inserted into both the cortex and the PMRF. A single electrode (Fred Haer – UEJB1), driven via a manual hydraulic drive, was used to stimulate and record from neurons within the PMRF. Up to four electrodes (details Alpha Omega, Aplharetta,

GA), placed using an automated electrical drive, were used to stimulate and record from neurons within the motor cortex. Single pulse microstimulation was delivered at each site while the subject performed the bilateral reaching task. Stimulation was delivered using a digital stimulus controller (Master-8, AMPI, Israel) connected to an analog stimulus isolator (Model 2200, AM-

Systems, Carslborg, WA, USA). The current selected for single pulse microstimulation in the

22 motor cortex matched that required to elicit a visible muscle twitch during testing with trains of

36 stimulus pulses delivered at 333 Hz at the same site. Actual currents used ranged from 20µA to 120µA. In no case did single pulse microstimulation at these sites evoke visible muscle twitches (but stimulus trains did). In the PMRF, a standard current of 30 µA has proven effective in other studies with both monkeys and cats and was used here except in cases where these single pulses evoked visible muscle twitches, where it was reduced to as little as 10 µA, as described previously (Davidson & Buford, 2006; Davidson & Buford, 2004; Drew & Rossignol,

1984a; Drew, 1991).

Once the cortical and PMRF electrodes were positioned at sites from which stimulus trains evoked visible muscle twitches and the appropriate currents were established, the electrical stimulation paradigm for this study began. (Fig. 2) A custom written script within the CED

Spike2 software was used to control stimulation. A set of stimulus combinations among electrodes was created to encompass all possible combinations: PMRF site only (RF Only), each cortical electrode alone (CX Only), and paired stimulation between the RF and each cortical electrode at one of three latency combinations: RF before CX, Simultaneous, and CX before RF.

In an example with 2 cortical electrodes, this would produce 9 possibilities: RF Only, CX 1

Only, CX 2 Only, and 2 sets of three paired latencies (1 + 2 + 2x3 = 9). The script then randomly arranged the order of stimulation among these possibilities to create 2000 sets of stimulation. Within each set, the order among the possibilities was random. But over the broader time scale of the entire stimulation session, every possible combination was distributed evenly across time. In this way, we could avoid order effects; it would be confounding if every time CX 1 was stimulated alone, it was preceded by RF only and followed by CX 2 only.

Randomization within each set prevented this. At the same time, allowing full randomization across time could have created the possibility of some combinations being more frequent early in the recording sessions and others concentrated at another time. In the case with 4 cortical electrodes plus the RF electrode, the number of stimuli to be delivered was 34,000 ((1 RF + 4

CX + 4x3 Paired) x 2000 = 17 x 2000 = 34,000). At 10 Hz, this required 3,400 seconds (56 minutes and 40 seconds). Hence, assuring that the sets of stimuli were evenly distributed across time was important to control for the possibility of subtle shifts in electrode position over this time frame having some effect of the effectiveness of stimulation at any given stimulation site during a single recording session. Whatever these effects may have been, they were evenly distributed across all electrode and latency combinations.

EMG Integrity Tests

As reported in (Montgomery et al., 2013a), the EMG leads were tested every two weeks to confirm they were in the correct location. They were tested by passing a current up to 2.0 mA through the implants and looking for overt muscle twitches. If no twitch was observed, the implant was considered compromised and was omitted from analysis. Occasionally, a few EMG connections would malfunction despite intact electrodes. This was determined from the EMG data. If the malfunction appeared for a limited number of sessions, data from those muscles, for those particular days were omitted from analysis. If the EMG electrodes failed to collect data reliably throughout the experiment, muscles implanted with those electrodes were removed from analysis. For this reason, all data from the left and right cervical paraspinal electrodes from all three subjects were omitted from analysis.

A cross-talk analysis was also performed in order to eliminate double counting of events that were recorded from one motor unit in more than one electrode. This procedure was described previously (Davidson & Buford, 2006). As reported in Montgomery et al. (2013b), in subject H, cross talk was detected between right middle deltoid (RMDelt) and right supraspinatus. RMDelt was removed from analysis for this subject. Similarly, data from the electrode in the left lats was omitted from subject O.

Data Analysis:

Stimulus triggered averages of the responses were calculated around the time of the

PMRF electrode stimulation, except when the cortex was stimulated alone. For each muscle, a threshold level was determined that was above baseline in the rectified EMG by at least 3% of the maximal EMG for that muscle. Stimuli applied while the muscle was inactive according to that threshold were not used for construction of the average. This is a necessary procedure for stimulus triggered averaging to provide for the possibility of observing suppression and to prevent stimuli applied while the motor pool is under inhibition from diluting the true effects, which can only be observed as an increase or decrease when the motor pool as a whole is active at some level (McKiernan, Marcario, Karrer, & Cheney, 1998). Other criteria for the detection of a response included a duration of at least 3.5 ms and a minimum of 200 triggers.

To determine whether the response observed during paired stimulation could be best explained by the cortex site alone, the PMRF site alone, by a simple summation of the RF and

CX effects, or none of the above (i.e., some type of gating or other complex effect), a simulation procedure was used. (Fig. 3) To compare the response to paired stimulation with the RF alone, the actual RF Only averaged response to stimulation was compared to the response to paired

25 stimulation at the time of the RF stimulus. To compare the response of each cortical electrode alone, the actual CX Only response was compared to the response to paired stimulation, with the time of the comparison in the response to paired stimulation shifted as needed to match the time of the cortical stimulus in that average.

To compare the responses to paired stimulation to the simulated results of a simple summation of the RF Only and CX Only responses, the following procedure was used. The time of every stimulus used in the actual paired stimulation average was located. Then, the nearest

RF Only stimulus and CX Only stimulus in the recording were located for which the muscle activity level was above threshold. For those RF Only and CX Only EMG records for that muscle, the raw, unrectified EMG was extracted for the time frame around the two stimuli.

These unrectified EMG records were then shifted in time as appropriate to the latency (if any) between the RF and CX stimuli, and the waveforms were added together. This added waveform was then rectified and included in an accumulating average. In this way, a simulated response for the paired stimulus was created from the responses to CX Only and RF Only responses.

The question for each response was, could the response be explained best by CX alone,

RF alone, a simple summation of the individual responses, or something more complicated. A gating response would be noted, for example, if the RF response was blocked by the CX response or some other type of complex interaction. To make the comparisons necessary to address this question, a simple linear regression between the points on one waveform and the points on the other was employed, yielding a goodness of fit (R2) and a significance level (P).

The RF Only, CX Only, and Simulated Paired averages were compared to the actual paired average for each muscle and each latency. The strengths of regressions were then statistically

26 compared against each other to determine, in cases where more than one fit was significant, whether any particular fit was significantly better than the rest. For example, in a case with a small CX response and a moderate RF response, it might be expected that both the RF Only and the Simulated averaged were a significant fit with the actual paired average, and the CX only response would not be a significant fit. In this case, if the RF Only and Simulated fit were not significantly different from each other, this would be considered an example of an “RF Only” response without gating. Further elucidation of the logic employed to categorize responses is provided in the results along with representative examples.

RESULTS

General Characteristics of post-stimulus effects

There were a total of 227 cortical sites and 68 PMRF sites from which stimulation produced significant EMG responses. From the cortex, these sites included 91 from M1, 76 from

PMA, and 60 from SMA (Fig. 4). From the PMRF, these included 30 from the right PMRF and

38 from the left PMRF (Fig. 5). During single-pulse microstimulation, a total of 952 EMG responses were elicited from the CX and 554 EMG responses were elicited from the PMRF.

Investigation of the effectiveness of both cortical and PMRF sites for producing a muscle response reveals that, in general, more responses per site were elicited from the PMRF (8.15) than the cortex (4.19).

Contralateral, ipsilateral, and bilateral responses were detected from each cortical region as well as from the PMRF. Figure 6A shows this distribution from the cortex. As expected, the majority of responses from all cortical motor areas were contralateral (M1=62%, PMA = 54%,

SMA = 49%). SMA and PMA showed more ipsilateral recruitment than M1. All three cortical motor areas demonstrated similar amounts of bilateral recruitment (M1=12%, PMA= 11%, SMA

= 13%). Figure 6B shows the laterality distribution for the PMRF. Both the left and right PMRF activated muscles bilaterally in much larger proportion than the motor cortex. Of the three subjects, N had very different proportions of responses from the left vs. the right PMRF. This subject had the lowest number of sites tested, and the fewest number of responses, so this may be a result of the smaller sample size. Among all individual muscle responses recorded, 24% were ipsilateral to the side of PMRF stimulation, 30% were contralateral and 46% were bilateral. This agrees well with our previous studies (Davidson & Buford, 2006; Davidson & Buford, 2004).

Overall, when considering the PMRF as a whole, 4.5% of sites elicited responses from ipsilateral muscles only, 7.5% percent of sites elicited responses from contralateral muscles only, and the remaining 88% of sites elicited bilateral responses.

Paired Pulse

A representative example of EMG data collected during paired pulse stimulation responses is illustrated in Fig. 7. Here, each row corresponds to a different muscle. Each column corresponds to the type and region of stimulation. Column 1 (RF) corresponds with PMRF stimulation. Column 2 (CX) corresponds with cortical stimulation. Columns 3-5 correspond to each of the latencies for the paired-pulse paradigms, L1, L2, and L3, respectively.

In many ways, these examples typify the types of response combinations detected. For example, one of the most common response patterns is shown in Fig. 7a. Here, no EMG response was detected from cortical stimulation (CX), a significant response (suppression) was detected

28 from PMRF stimulation (RF), and similar significant suppression responses were detected in L1,

L2, and L3. This leads to the conclusion that the PMRF was primarily responsible for the EMG responses of this muscle to the paired pulse stimulation paradigms for this stimulation site.

Figure 7b illustrates a typical example of summation. Here, EMG responses were evoked from stimulation of the cortex alone and from the PMRF alone. Then, during paired pulse stimulation, the latency, shape, and comparison to the simulated waveform of the EMG response are indicative of a summative response that was a simple combination of the cortical and PMRF responses.

Figure 7c illustrates a typical example of gating. Here, an EMG response was evoked during both PMRF and CX stimulation. Then, during paired pulse stimulation, particularly during the 2nd latency, when the cortex and PMRF were stimulated simultaneously, characteristics of the cortical signal are lost in the response, but appear in the simulated summation. This suggests that the effect from the PMRF has blocked or gated out the effect from cortex.

Do Brain Region Pairs Influence Interaction Response Type?

Figure 8 shows the ratio of EMG responses evoked for each pair of brain regions stimulated for each type of response. As shown in the top row of Fig.8, the most common combination of sites for observing a summation response was by stimulating M1 with either side of the PMRF, although this was not statistically significant (χ2 = 2.908, P = 0.245). In the second row of Fig. 8, the best chance to observe a gating response was by stimulating the ipsilateral PMRF and any region of the CX, but again this was not significant (χ2 = 2.222, P =

0.329). In the bottom row of Fig. 8, the best chance to observe a complex response was by stimulation of the left PMRF and PMd, and this was highly significant. (χ2 = 18.459, P = 0.000).

Fisher’s Exact Test was also run to account for the smaller sample size, and the same p-values were obtained.

When these data were further divided by side of the body, Chi square and Fisher’s exact tests revealed that an interaction for response type, in both arms, was dependent on brain region

2 2 pairs (χ left = 12.186, Pleft =0.002, χ right = 11.940, Pright = 0.002), which is in agreement with the findings stated above. Furthermore, as shown in Fig. 8, responses were found to be most prevalent within the right arm. This is not surprising as the right arm is contralateral to the cortical areas that were stimulated. However, there were several ipsilateral examples as well, demonstrating that interactions between the PMRF and the cortical motor areas can occur ipsilateral or contralateral to the cortex.

To determine if the onset latency of an EMG response was indicative of the type of interaction between regions, a one-way-ANOVA test was performed. This test revealed that there was no significant relationship between onset latency and interaction response type (PRF=

0.269, PCX= 0.811).

DISCUSSION

Single-pulse and paired-pulse microsimulation were used in combination to investigate the nature of the interactions between cortical and PMRF regions. Specifically, M1, SMA, and

PMd of the motor cortex and left and right sides of the pontomedullary reticular formation.

Consistent with previous findings, both the motor cortex and the PMRF were found to recruit

30 muscles in the upper limbs with ipsilateral, contralateral, and bilateral outputs evident from all regions. Bilateral outputs were the most common from the PMRF, regardless of side, and contralateral outputs were the most common from the cortex.

While several studies have reported on the neural activity of the motor cortex and fewer reported on the PMRF, we believe this was the first study to investigate the combined output of these motor systems using stimulus triggered averaging. It was determined that the regions do interact to produce muscle recruitment in the upper limbs. These interactions take the form of cooperation (summation), competition (gating) and other, more complex interactions.

For example, while simple summation was the type of response observed least frequently, when it did occur, it was usually in the contralateral (right) arm in response to M1 stimulation.

This is not surprising given what we know about the relatively direct descending pathways from

M1 and the PMRF. For example, it is known that a majority of projections from M1 terminate in the spinal cord (corticospinal) and some, independently, terminate in the PMRF (corticoreticular pathways). However, M1 does not appear to be a source of complex, branching axons that collateralize to both sides to reach the brainstem and spinal cord from individual neurons

(Keizer, K., Kuypers,H.G.J.M., 1989). Similarly, several studies have concluded that reticulospinal neurons also feature relatively direct influence on the motor pools (Jankowska &

Edgley, 2006; Matsuyama & Drew, 1997). Therefore, given that both regions have projections that terminate in the spinal cord, with motor pools potentially being the first point of interaction, it is reasonable to expect that the alpha motor neurons within the motor pool behave as simple summing junctions to integrate the two sources of drive without preferring one over the other.

This is in line with our results that typify summation responses as a result of stimulation of M1 more often than from other regions of the motor cortex.

The most common type of interaction observed was the complex response. These are responses that cannot be easily attributed to simple summation or gating of signals. Here, the ratio of responses evoked from stimulation of pre- and supplementary motor areas are much higher than in the simple summation case. This is not surprising given what we know about the premotor areas. For example, it has been shown that, like M1, premotor areas consist of both corticospinal (Jankowska & Edgley, 2006) and corticoreticular (Matsuyama et al., 2004; Ortiz-

Rosario et al., 2014), cells as with M1 Unlike the primary motor cortex, however, it has also been demonstrated that some of the corticospinal cells deliver collaterals to the reticular formation resulting in more complex, branching connections (Fig. 9). In a retrograde labeling study published by Keizer and Kuypers (1989) SMA neurons were found to be double-labeled from the spinal cord and the brain stem, thus signifying multi-branching, complex neurological pathways between these regions. Similarly, results from (Fregosi & Rouiller, 2017), as well as yet unpublished results from anterograde corticoreticular tracing experiments from our lab, reveal strong coritcotectal and corticoreticular labeling from PMd and SMA neurons, respectively, but sparse, if any, labeling from M1 neurons ((Fregosi & Rouiller, 2017). Due to the difficulty of these experiments and the possibility of technical error, we have been reluctant to interpret negative results from motor cortex injections. However, we seem to have consistent success from the premotor cortical areas in the same lab using the same methods (Montgomery,

Hirschauer, Buford, unpublished results). Moreover, Yeo et al. (2012) report on the corticoreticular pathway in the human originating in the premotor area and terminating in the

PMRF using diffusion tensor tractography. They show that this pathway descends through the corona radiata and the posterior limb of the internal capsule to the cortical spinal tract. It then continues to descend through the mesencephalic tegmentum and the pontine reticular formation and ultimately terminates in the PMRF.

These results are, therefore, consistent with our finding that a higher ratio of complex responses are evoked from stimulation of the pre-motor regions and that simple summation responses are more commonly evoked from stimulation of M1.

Apart from these so-called “expected results”, there were a number of other interesting findings including indications of gating interactions (n=50). Here, we define gating as any competitive interactions exhibited by the regions. In this study, we observed approximately a

2/3rds majority (32:18) of gating responses where the cortex gated out the PMRF. This is congruent with the hierarchical way of thinking about the nervous system. Where the PMRF has predominantly bilateral influence over motor control in the upper limbs, it is often desirable to move only one arm at a time. This is particularly true in primates and especially from the seated position where the arms are not being used for postural support. Therefore, if the cortex and the

PMRF are both active during reaching, we would expect some gating out of the PMRF effects on one side to allow unilateral control of movement from the cortex. However, there is no reason to suppose that the gating mechanism exclusively works in one direction and so therefore the observation that gating occurred from both the PMRF and the cortex is not entirely unexpected.

In cases of rapid postural adjustments driven energetically by subcortical systems (Luccarini,

Gahery, & Pompeiano, 1990b; Schepens & Drew, 2003) for example, one might expect the brainstem to have priority over the cortex.

Moreover, we expect to see a greater representation of gating responses in the left arm than the right arm given that our stimulation paradigm was focused in the left cortex. Therefore, cortical signals to make reaching movements with the right arm should be coupled with gating out the simultaneous ipsilateral portion of PMRF signals that want to move the opposite (left) arm. Indeed, when we compare the left and right arm, we see that gating was indeed overrepresented in the left arm compared to other types of responses.

Aside from evidence of gating between the cortex and the PMRF, another interesting result from this study was that, for summation responses, while response frequency was largely irrespective of the side of PRMF stimulation, stimulation of the PMRF contralateral to the reaching arm did tend to show slightly more summation responses in both arms. This result might be unexpected because typically the ipsilateral (to the arm) brainstem projections are slightly more prevalent (60/40) than the contralateral (Montgomery, 2013). Therefore, we would expect to see the majority of right arm responses after stimulation of the right PMRF. However, given that the ratio of projections on either side are reasonably close to equivalent, it is not improbable to see this imbalance favoring the left PMRF in the right arm. One explanation for this observation that there were slightly more summation responses found during PMRF stimulation while the contralateral arm was reaching (LPMRF/Right Arm and RPMRF/Left

Arm) could be explained by higher levels of extensor activity over flexor activity in that arm, due to the nature of the task, as has been seen in previous studies (Davidson & Buford, 2004;

Drew & Rossignol, 1990). Therefore, the slightly higher number of summation examples we observed from stimulation of the left PMRF may have more to do with the task employed than the prevalence of connections.

Another interesting result was that we observed more responses in the contralateral arm during stimulation of ipsilateral brain regions (eg: left PMRF, left cortex), compared to contralateral brain regions. One explanation for this result is that ipsilateral brain regions (left cortex and left PMRF) are most likely to cooperate in the control of movement of the contralateral arm, due to more or stronger pathways between these regions versus contralateral regions. This explanation is supported by the anterograde tracing study conducted by Matsuyama and Drew (1997), where they showed that projections from area 4 of the motor cortex and areas

6αβ and 6αγ of the premotor cortex had a majority of terminal swellings in the ipsilateral PMRF in the cat. Overall, this pattern of response frequency was seen in all three response types.

SUMMARY AND CONCLUSIONS

Taken together, then, it is clear that both corticospinal and reticulospinal systems need to cooperate for control during skilled reaching. It is not the case that the PMRF is only responsible for control of posture while the CX controls reaching. Both are involved in recruiting arm muscles in complex ways. Here, we have presented some evidence to suggest that there are several types of interactions that can occur, including simple summation, gating, and other more complex interactions, and that some of these interaction types are more prevalent from some brain regions than others. In some cases, we have demonstrated that these interactions fit with what is known about the neural pathways. On the other hand, we have shown that there are many interactions that highlight the complex nature of the system which encompasses corticoreticular, corticospinal, reticulospinal, and spinal networks. Overall, achieving a more complete understanding of how these regions cooperate and compete for the control of movement of the upper limbs can lead to more well-informed rehabilitation efforts.

Figure 1: Illustration of instructed delay, bilateral reaching set-up and accompanying sketch of a reaching trial. Adapted from Davidson and Buford 2006.

Figure 2: Illustration and description of the stimulation paradigms used to during single-pulse microstimulation of the cortex and PMRF and paired-pulse microstimulation of the two regions.

Figure 3: Process for creating actual averages and simulated averages for paired pulse stimulation A raw sample of EMG recorded from right upper trapezius is illustrated  along with the times of single pulse stimulation in the cortex (CX) and brainstem (RF). For the actual paired average response, each individual sweep is  extracted, rectified, and  added into the averaged waveform to create the stimulus triggered average (, black trace). For the simulation of the paired pulse average, the sweep for the nearest stimulus of the  RF only and  CX only were each located and extracted (with the CX data time-shifted as appropriate to mimic the paired pulse stimulus latency). These raw waveforms were then  added together,  rectified, and  added into the averaged waveform to create the simulated paired-pulse response (, grey trace). The post-stimulus waveforms for the actual paired-pulse stimulus triggered average response was then compared to the RF only response, the CX only response, and the simulated paired-pulse response to determine the best fit  according to a linear regression, as explained in methods. In this case, the response to the paired pulse stimulus best matched the response for RF only; the CX stimulus had no effect. For the traces shown at the bottom right, “trigs” indicates the number of triggered sweeps used to create the average (only sweeps where EMG was above baseline were used). The solid bars under the trace for the RF Average and the Paired Pulse response indicate periods of significant post-stimulus suppression

Figure 4: Cortical stimulation sites For each subject (H, N, and O), a surface map of the electrode penetration sites is provided. Horsley-Clarke stereotaxic coordinates are indicated lateral to the midline (towards the left) and rostral of the coronal place. Symbols are color coded by the response evoked with a 36- pulse train of stimulation at 333 Hz, as described in METHODS. The red line indicates the central sulcus (cs). A dotted line rostral to the central sulcus, indicated by (a), shows the point at which the electrode descended into the rostral bank of the central sulcus. These responses are reflected posteriorly, caudal to line (a), with increasing depth represented as a more caudal position. The actual electrode penetrations were all rostral to line (a) in order to avoid the vasculature overlying the central sulcus. Similarly, line (b) indicated a medial boundary past which electrode penetrations were not made in order to avoid the sagittal sinus. Symbols plotted to the right of line (b) represent sites deep in the mesial cortex of the supplementary motor area, reflected medially so that increasing depth is represented as a more medial position. Abbreviations: prcd = pre central dimple, arcs = arcuate sulcus, sp- arcs = spur of the arcuate sulcus. Sites found within the dorsal premotor cortex (PMd) and the supplementary motor area (SMA) are indicated. All other sites were in the primary motor cortex. Sites from which leg, face, or trunk / tail responses were evoked were removed from the analysis.

AP -2 AP 1 nVII

VII IO spV

IO Py Py

AP -1 AP 2

4th

nVII

IO Py Py

Py = Pyramidal Tract IO = Inferior Olivary Complex SO = Superior Olivary Complex SC nVIII = Vestibulocochlear nerve VI = Abducens Nucleus nVII = Facial Nerve IC = Inferior Colliculus IC SC = Superior Colliculus AP 0 4th = 4th Ventricle AP 3 = Horsley-Clarke DV -12

4th

IO Py

Figure 5: Locations of stimulation sites in the pontomedullary reticular formation.

Locations are plotted on standard brainstem sections adapted from the Szabo and Cowan atlas. Black circles are from subject H, green from N, and purple from O. The inset key identifies selected anatomical structures for reference. In the Horsley-Clarke coordinates, the sections span from approximate -2 to +3 in the coronal plane (AP). In the dorso-ventral direction, the target symbol indicates the midline at DV -12, which is where the center of the recording chamber was aimed for these studies. 39

Figure 6: A) Ratio of ipsilateral, bilateral, and contralateral muscle responses during single-pulse microstimulation of M1, PMd, and SMA B) Ratio of ipsilateral, bilateral, and contralateral muscle responses during single-pulse microstimulation of the left and right PMRF.

A RF RF RF RF CX CX CX CX

-20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60

B RF RF RF RF CX CX CX CX

-20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60

C RF RF RF RF CX CX CX CX

-20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60 -20 0 20 40 60

Figure 7: Representative examples of EMG data recorded during both single (RF Alone and CX Alone) and paired pulse (Latencies 1 -3) microstimulation. A red trace indicates a facilitation response and dark blue represents a suppression response. The light blue shown during paired pulse stimulation represents the simulated summation response. In paired stimulation, Time 0 always indicates the time of the RF stimulus; the time of the cortical stimulus. Panel A shows a typical example of a PMRF driven response. Panel B is an example of an EMG response likely driven by summation of the PMRF and CX signals. Panel C shows an example of a gating interaction where the PMRF seems to gate out the CX contribution during simultaneous paired stimulation.

Figure 8: Number of different types of paired responses based on which cortical area was stimulated (Black bars = M1, striped bars = PMd, and white bars = SMA) and the side of stimulation of the PMRF divided left to right in each box. All of the boxes on the left side of the figure show the number of responses obtained from muscles on the left side of the body and all of the boxes on the right show the number of responses obtained from muscles in the right side of the body.

Midline Midline Supplementary Primary Motor Area Motor Area Cortex

PMRF

Motor Motor Network Network Motor Motor Spinal Cord Network Network

Figure 9: Simple pathways between the cortex and PMRF lend themselves to interact in a simple, summative way. More complex circuitry between pre- and supplementary motor areas and the PMRF lend themselves to gating and other complex interactions.

CHAPTER 4. COMPUTER-AIDED PREDICTION OF EXTENT OF MOTOR RECOVERY FOLLOWING CONSTRAINT-INDUCED MOVEMENT THERAPY IN CHRONIC STROKE

ABSTRACT

Constraint-induced movement therapy (CI therapy) is a well-researched intervention for treatment of upper limb function. Overall, CI therapy yields clinically meaningful improvements in speed of task completion and greatly increases use of the more affected upper extremity for daily activities. However, individual improvements vary widely. It has been suggested that intrinsic feedback from somatosensation may inﬂuence motor recovery from CI therapy. To test this hypothesis, an enhanced probabilistic neural network (EPNN) prognostic computational model was developed to identify which baseline characteristics predict extent of motor recovery, as measured by the Wolf Motor Function Test (WMFT). Individual characteristics examined were: proprioceptive function via the brief kinesthesia test, tactile sensation via the Semmes-

Weinstein touch monofilaments, motor performance captured via the 15 timed items of the Wolf Motor Function Test, stroke affected side. A highly accurate predictive classification was achieved (100% accuracy of EPNN based on available data), but facets of motor functioning alone were sufficient to predict outcome. Somatosensation, as quantified here, did not play a large role in determining the effectiveness of CI therapy.

INTRODUCTION

Constraint-induced (CI) movement therapy is a restorative intervention for motor rehabilitation of upper limb hemiparesis with strong empirical backing. It incorporates two main elements: (1) intensive motor training, and (2) use of behavioral techniques to increase use of the more affected upper extremity for all daily activities throughout the intervention period (Morris, Taub, & Mark, 2006; Uswatte & Taub, 2013). Despite producing consistently clinically meaningful improvements in motor function on a group level [as measured by the Wolf Motor Function Test (WMFT) (Taub, Uswatte,

Mark, & Morris, 2006; Taub et al., 1993; Wolf et al., 2006a), almost a third of participants fail to achieve a meaningful improvement in upper extremity function from

30 h of intensive motor practice (Table 3). For these non-responders, a restorative approach to motor training (i.e., the intensive motor training portion of CI therapy) has a low return on investment and will monopolize valuable treatment time that may be better spent teaching compensatory strategies (e.g., coaching use of adaptive aids) that could enhance use of the weaker upper extremity for daily activities. It is thus critical to identify who may beneﬁt most from such a restorative approach to motor rehabilitation.

Somatosensory function is a largely neglected element of motor learning that may inﬂuence response to CI therapy. The techniques used in CI therapy were derived from basic primate research involving somatosensory deafferentation (Jurewicz et al.,

2018). Following the deafferentation procedure, the monkeys’ brains no longer received sensory input from the deafferented upper extremity. This sensory loss manifested behaviorally as complete disregard for the deafferented limb. However,

spontaneous use of the limb was restored following intensive training and forced use.

The series of deafferentation experiments carried out by Taub and colleagues

(Jurewicz et al., 2018; Taub, 1976; Thibault, Robert T., Lifshitz, Michael,Raz, Amir,

2016) and the successful translation of CI therapy to treat post-stroke motor deﬁcit

(Taub et al., 1993; Taub et al., 2006a) clearly demonstrates the feasibility of treating

individuals who have profound sensory impairment with CI therapy.

However, the literature also provides a theoretical rationale for why somatosensory impairment may adversely affect the potential for motor recovery. Given the extensive intracortical connectedness of primary sensory and primary motor cortices, somatosensory impairment post-stroke may adversely affect use- dependent neural plasticity by reducing intrinsic feedback into the sensorimotor system as shown in Fig.

10. For example, poorer motor function post-stroke has been associated with reduced sensory pathway information transfer (Zotev et al., 2018). Greater integrity of sensory as well as motor cortex has been shown to result in greater motor improvement from

CI therapy (Van Doren et al., 2018), albeit inconsistently (Ortiz-Rosario, Adeli, &

Buford, 2015). One implication of diminished poststroke somatosensation is that it may impede or prevent learning of motor tasks (Liu et al., 2018), which may explain its association with poorer motor recovery after stroke (J. Ibáñez et al., 2017). Taken together, these studies provide the theoretical rationale for how impaired somatosensation may inﬂuence motor recovery. However, existing work on the relationship between sensory impairment and motor recovery lacks the ﬁne-grained analysis necessary to examine the interaction between sensory loss and motor recovery

46 as a function of initial motor impairment.

Figure 10: This schematic representation of motor learning illustrates how diminished intrinsic somatosensory feedback following stroke could reduce error detection and motor learning.

We hypothesized that impaired post-stroke somatosensation may differentially affect response to CI therapy, depending on initial motor deﬁcit. Prior primate work demonstrated that impaired somatosensation is a contributor to learned non-use following stroke (Jurewicz et al., 2018; Taub, 1976; Thibault, Robert T., Lifshitz,

Michael,Raz, Amir, 2016). Thus, for individuals in which motor impairment is minimal and dexterous performance is generally intact, overcoming learned nonuse may be the primary mechanism contributing to recovery. Conversely, if motor impairment is moderate and new motor learning is required, impaired somatosensation may interfere with new learning and result in poorer response to CI therapy. Human studies have not thoroughly examined the role of somatosensory impairment on

47 outcomes from CI therapy (Wright & Jordanov, 2017).

To address this gap, improvements in function of the more affected upper extremity were examined as a function of initial sensory and motor ability using state-of-the-art, objective, computational methodologies in order to achieve the highest prognostic accuracy. In recent years, similar artiﬁcial neural network and machine learning algorithms (Adeli, Hojjat,,Hung, Shih-Lin,, 1995), along with nonlinear science/chaos theory, have been employed to advance diagnosis of a multitude of neurologically- based disorders (Antoniades et al., 2018; Attema, Kosgodagan Acharige, Morales-

Nápoles, & Maljaars, 2017; Hirschauer et al., 2015a; S. Hulbert & Adeli, 2015;

Koziarski & Cyganek, 2017; Y. Lin, Nie, & Ma, 2017; Martín-López et al., 2017a;

Martín-López et al., 2017b; Morabito et al., 2017; Ortega-Zamorano, Jerez, Juárez, &

Franco, 2017; Padillo, Luna, Herrera, & Ventura, 2018; Valenzuela, Jiang, Carrillo, &

Rojas, 2018; S. Y. Zhang, Jiang, & Neild, 2017; Y. Zhang, Wang, Jin, & Wang,

2017). However, these techniques have never before been applied to prediction of outcomes from a rehabilitation intervention, to our knowledge. We present here an enhanced probabilistic neural network (EPNN) prognostic computational model to predict the extent of motor recovery after CI therapy. The resulting work could inform treatment recommendations for up to 89% of stroke survivors affected by somatosensory loss (Daubechies, 1990) .

METHODS

Participants

Participants were 35 individuals who had experienced a stroke of any etiology at least six months prior and provided written informed consent to participate in a CI therapy research protocol approved by the Ohio State University’s Institutional Review Board. All participants met the motor criteria utilized in the EXCITE trial of CI therapy, reﬂective of mild/moderate motor impairment (Wolf et al., 2006b). Unlike in many prior CI therapy trials, participants were enrolled largely irrespective of cognitive or mobility status. Those who lacked capacity to provide informed consent, were unable to comprehend one-step commands in English, or who were currently receiving Botox treatment (confound) were excluded. See Table 1 for participant demographics.

Table 1 Demographic and clinical characteristics of 35 patients

Item Number of case (Average) Minimum Maximum Age (60) 24 84 Sex 24 males N/A N/A Chronicity (years) (2.9) 0.6 19.6 Stroke affected side 14 right N/A N/A Handedness 26 right N/A N/A Affected side was dominant 13 patients N/A N/A Intervention

All participants completed CI therapy over the course of 3 weeks. Treatment consisted of ten 3-hour sessions of intensive motor practice with shaping, restraint of the weaker

arm for a target 90% of waking hours, and a “transfer package” of behavioral techniques

to address learned non-use (approximately 5 h total). More detail on the treatment

protocol can be found in Morris et al. (2006).

Clinical measures

Table 2 summarizes the motor and somatosensory measures, also described brieﬂy

here.

Table 2 Behavioral measures used in the prognostic model Number of Fine or Gross Type of test Function assessed Behavioral measures items motor Forearm to table (side) Gross Forearm to box (side) Gross Extend elbow (side) Gross Hand to table (front) Gross Hand to box (front) Gross Weight to box Gross Reach and retrieve Gross Upper limb functional performance Motor 15 Lift can Fine (timed) Lift pencil Fine Lift paper clip Fine Stack checkers Fine Flip card Fine Turnkey in lock Fine Fold towel Fine Lift Basket Fine Reaching error with visual occlusion Brief kinesthesia test (affected N/A (proprioception) side) Somatosensory 2 Touch test monofilaments Touch perception threshold N/A (affected side)

Wolf Motor Function Test The WMFT assesses upper limb motor ability through 15 timed functional tasks as

summarized in Table 2. The performance time of each item was natural-log-

transformed to reﬂect the relative nonlinearity of potential performance time

improvement (i.e., an improvement from 4 s to 2 s is greater than an improvement from

100 to 102 s). The WMFT summary score reﬂects the mean of the natural-log- transformed item scores, Eq. (1) where Mi is the ith WMFT test score: patients with mild to moderate hemiparesis and distinguishes post- stroke performance from age-, gender- and handedness-matched healthy controls (A. Borstad & Nichols-Larsen, 2016a).

The BKT total score is the sum of the error, in centimeters, for three reaching trials with the more affected limb. The touch monofilament (TM) test identifies the lightest force in grams perceived consistently by an individual on the index finger. For the purpose of this study, participant data was categorized as follows: (1) normal (0.008 to 0.07 g), (2) diminished light touch (0.16–0.4 g), (3) diminished protective sensation (0.6–2.0 g), and (4) loss of protective sensation (4–300 g) [32].

∑𝑖=1 푙푛푀푃표푠푡−푡ℎ푒푟푎푝푦 ∑𝑖=1 푙푛푀푃푟푒−푡ℎ푒푟푎푝푦 퐶 = 15 𝑖 − 15 𝑖 (1) 15 15

A higher score reﬂects poorer performance.

Somatosensory function tests The brief kinesthesia test (BKT) is a measure of error in guided reaching with visual occlusion thought to represent upper limb kinesthetic sense. It has been shown to be feasibly administered to

ARCHITECTURE OF THE PREDICTION MODELS

EPNN was employed to predict extent of motor recovery after CI therapy based on sensorimotor assessments obtained prior to treatment. EPNN can identify how each

51 motor and somatosensory test affects the accuracy of the prognosis. Further, EPNN was used to analyze all combinations of motor and somatosensory predictors to identify the combination that accounted for the most accurate prognosis.

Ahmadlou and Adeli (2010a) developed EPNN as an improvement to Probabilistic

Neural Network (PNN) where the spread of Gaussian function is computed using local decision hyperspheres. A brief summary of the EPNN algorithm is presented in Addition

A. The architecture of the model is presented in Fig. 11.

Figure 11: Architecture of the prognostic EPNN model used to predict the extent of motor recovery after CI therapy. The input layer consists of 18 nodes reﬂecting the 18 assessment items, the predictors.

Inputs consisted of the following: side of motor impairment, each of 15 WMFT natural- log-transformed item times, the BKT score, and the TM score for the affected side. The output of the system was the extent of motor recovery of the more affected upper extremity after CI therapy. For the purpose of this analysis, participants were classiﬁed as

52 non-responders, moderate responders, and best responders to CI therapy based on the difference between their pre-therapy and post-therapy WMFT summary scores. The determination to assign individuals to one of three outcome categories reﬂected a trimodal distribution in WMFT treatment response (Fig. 12). Thresholds for these classiﬁcations are given in Section 2.5.

Figure 12: The histogram of the patients’ change in the natural log of WMFT scores from pre to post CI therapy. The pattern layer in Fig. 11 mathematically computed the motor and somatosensory similarity of different participants within each outcome grouping: (1) non-responders,

(2) moderate responders, or (3) best responders. In this way, one could compute how closely an individual matched with each potential outcome group. In the summation 53 layer, the average likelihoods of the participant matching with each of the three outcome groups were computed using the information obtained from the pattern layer (for an explanation of the concept of likelihood see Addition A). The decision layer assigned each participant to one of the three outcome categories based on the maximum average likelihood from the summation layer: non-responders, moderate-responders, or best- responders to CI therapy.

DATA COLLECTION, DESCRIPTION, AND PREPARATION

Table 3 summarizes the collected participant data used to train and test the EPNN prognostic model. Data were collected within 3 business days before and after CI therapy. Participants were classiﬁed into the aforementioned three categories based on their WMFT treatment change. Non-responders were considered to be those whose

WMFT treatment change was less than the minimal clinically important difference

(0.15, which equates to 16%) [34]. Moderate responders were those with a treatment response that was greater than the minimally clinically important difference, but less than 0.40 (50% change). Finally, best responders were those with changes greater than

0.40. The aforementioned thresholds were based on visual inspection of the 12.

Table 3 Patient data used to train and test the EPNN model Inputs OT P# 10. 11. 12. 15. 17. 18. 1. SA 2. BT 3. CN 4. EE 5. EW 6. FC 7. FB 8. FT 9. HB 13. PL 14. SC 16. TL HT KY LC RR BK TM 1 0 1.26 3.11 1.27 1.25 4.12 1.2 0.72 0.82 0.8 4.01 3.13 2.4 4.79 0.66 3.8 2.97 5.7 1 2 1 1.08 2.01 -0.25 -0.03 3.35 -0.16 0.39 -0.29 -0.39 2.06 4.79 4.79 3.4 -0.25 3.02 1.22 -1.83 1 3 1 4.79 4.79 1.9 2.23 4.79 1.19 1.8 0.85 0.22 4.79 2.33 3.81 4.79 2.38 4.75 3.47 0 2 4 0 3.09 4.79 4.79 4.79 4.24 0.41 0.38 0.41 0.76 4.79 1.99 1.28 4.79 1.21 3.65 2.68 -0.51 3 5 0 1.97 1.45 0.48 0.9 2.8 0.98 0.76 0.87 0.34 2.26 0.89 0.94 2.9 1.45 3.06 2.08 -4.83 2 6 0 0.84 0.74 -0.06 0.2 2.18 0.25 0.12 0.2 -0.27 1.58 0.96 1.84 1.83 0.63 1.81 1.87 -0.92 1 7 1 1.54 1.87 1.99 1.13 3.1 1.23 0.78 0.77 0.72 3.5 2.2 4.79 4.79 0.96 3.49 3.09 -0.51 3 8 0 2.34 1.03 4.79 4.79 2.78 0.36 0.22 0.14 0.06 2.02 3.58 1.52 3.41 0.49 2.62 2.53 -1.83 3 9 1 0.77 -0.06 0.43 0.03 1.26 0.1 -0.06 0.03 -0.16 1.36 0.41 0.31 2.4 0.43 1.58 2.46 -1.83 1 10 1 4.79 4.79 4.79 4.79 4.24 1.82 1.85 1.84 1.37 4.79 2.01 1.98 4.79 2.63 4.79 1.79 5.7 1 11 1 3.41 4.79 4.79 4.79 2.62 1.22 0.32 0.44 0.38 2.29 1.4 1.74 4.79 0.76 3.01 1.55 -1.83 3 12 0 1.14 0.74 0.32 0.74 1.77 0.36 0.43 0.25 0.29 1.5 1.14 0.61 1.39 1.14 1.62 1.74 -1.83 3 13 1 1.81 4.79 1.76 1.15 4.79 -0.53 -0.82 0.43 0.39 4.79 4.79 4.79 4.79 0.48 2.66 2.08 -3.22 1 14 1 1.22 1.59 1.34 1.41 3.03 0.96 0.79 0.64 0.83 2.56 1.28 2.43 4.79 1.28 2.27 2.82 5.7 3 15 1 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 4.79 2.74 -3.22 3 16 1 1.09 1.05 0.22 0.76 2.59 0.81 0.22 0.48 0 2.4 1.69 1.77 2.1 0.52 2.2 1.7 -3.22 3 17 1 4.79 4.79 4.79 1.83 4.79 4.79 4.79 0.72 0.31 4.79 4.79 4.79 4.79 4.79 4.79 3.3 -0.92 3 18 0 2.06 4.79 0.85 1.75 4.79 1.12 0.92 0.36 0.46 4.79 1.53 2.03 4.79 0.01 2.82 2.04 -4.83 2 19 1 4.79 4.79 4.79 4.79 4.79 0.84 0.74 1.28 0.87 4.79 4.79 4.79 4.79 4.79 4.79 2.73 -3.22 3 20 0 4.79 2.26 2.61 2.3 3.68 2.44 1.59 1.52 1.22 3.92 1.77 1.47 3.98 0.34 4.36 3.35 1.79 3 21 1 4.79 2.23 1.04 0.25 4.79 1.56 0.2 0.78 0.61 4.79 4.79 4.79 4.79 0.51 3.28 3.62 5.7 3 22 0 4.79 2.22 0.12 0.43 3.01 1.38 0.21 0.54 0.18 2.66 1.96 1.99 4.79 0.01 2.68 2.9 0 2 23 1 4.79 4.79 4.79 2.43 4.79 0.98 0.78 1.22 0.75 4.79 4.79 4.79 4.79 0.63 4.79 3.94 -1.83 2 24 0 1.44 0.85 0.49 0.6 2.24 0.21 -0.27 -0.15 -0.12 2.14 2.51 0.9 1.93 -0.3 2.28 2.58 2.08 3 25 1 4.79 2.27 4.2 2.29 3.86 4.79 0.45 3.1 1.23 3.88 1.45 1.47 4.79 0.45 2.76 2.58 -0.92 3 26 0 4.79 1.01 0.06 -0.25 2.25 0.76 0.15 -0.13 0.03 2.02 1.1 1.13 1.85 -0.37 1.99 1.79 -3.22 1 27 1 4.79 4.79 4.79 4.79 4.79 4.79 1.06 4.79 1.16 4.79 4.79 4.79 4.79 0.47 4.79 1.97 5.7 3 28 1 4.79 4.79 0.22 0.34 4.79 1.85 0.51 0.28 0.34 4.79 4.79 4.79 4.79 -0.17 4.79 3.41 2.3 1 29 0 1.56 2.46 -0.21 0.01 4.79 -0.02 -0.04 0.08 0.03 2.39 4.79 2.57 2.25 0.31 2.9 4.14 -2.66 3 30 1 1.2 1.1 0.27 4.79 2.32 0.58 0.86 0.22 0.04 2 0.92 0.8 1.63 0.96 1.82 2.13 -1.83 3 31 1 0.82 0.58 -0.2 0.03 2.1 0.8 -0.04 0 -0.03 2.01 1.13 0.78 2.13 0.11 1.73 2.79 0.69 1 32 1 4.79 4.21 4.79 4.79 4.79 0.43 0.32 4.79 0.34 4.79 4.79 4.79 4.79 0.82 4.03 2.16 -1.83 3 33 0 1.11 0.47 -0.29 0.48 1.53 0.15 0.39 -0.51 -0.33 0.88 0.45 0.22 1.5 -0.06 1.9 1.46 -0.92 1 34 1 4.79 1.98 0.2 1.2 4.79 0.99 0.46 1.25 0.9 4.79 3.89 4.79 4.79 0.46 2.96 4.79 -3.91 3 35 0 1.89 1.56 -0.09 -0.25 2.37 1.27 0.92 1.27 0.56 1.82 2.38 2.63 4.79 -0.39 2.56 2.51 -3.91 3 EPNN: Enhanced probabilistic neural networks; P#: Patient number; SA: Stroke affected side; BT: WMFT basket; CN: WMFT can; EE: WFMT extend elbow; EW: WMFT extend elbow weight; FC: WMFT flip cards; FB: WMFT forearm to box; FT: WMFT forearm to table; HB: WMFT hand to box; HT: WMFT hand to table; KY: WMFT key; LC: WMFT lift paper clip; PL: WMFT pencil; SC: WMFT stack checkers; RR: WMFT reach retrieve; TL: WMFT towel; BK: BKT; WMFT: Wolf motor function test; TM: Touch monofilament; OT: Output

TRAINING AND TESTING OF THE MODEL

The EPNN model presented in this paper was implemented in MATLAB version

9.1.0.441655 (R2016b). The conventional method for training and testing a neural network model is to divide the data into training and testing sets randomly using different ratios of testing to training (RTT) data such as 10% and 20%. Because the data available for training a sophisticated neural network classiﬁcation model were limited in this research, the model was trained and tested 35 times (equal to the number of poststroke participants), each time using the data for a different participant for testing and the remaining data for training. This results in an RTT of about 3.0%. The accuracy values reported in this research were the average of testing accuracies of the 35 trials.

SENSITIVITY ANALYSIS VIA EPNN

A sensitivity analysis was performed with two objectives: (1) to study the inﬂuence of each of the 18 individual predictors on the accuracy of the prediction of performance change with CI therapy (EPNN) and (2) to ﬁnd the combination of predictors that resulted in the most accurate prediction of performance change with CI therapy. To achieve the

ﬁrst objective, the model was run 18 times, each time removing one of the 18 inputs. If removal of an input resulted in an accuracy of less than 74.3%, the accuracy obtained for

EPNN using all 18 inputs, then that input was considered necessary for the model. On the other hand, if removal of an input increased the accuracy to a value greater than 74.3%

196

then that input was considered redundant for the model. Among inputs whose removals yielded the same accuracy, the most inﬂuential input was the one with the minimum of the maximum average likelihood, represented by a log-likelihood ratio normalized to the range of 0–1 using Eq. (7) explained in Addition A.

To achieve the second objective, all possible combinations of 18 inputs, a total of

262,125 combinations, were generated and K-Nearest Neighbor (KNN), PNN, and

EPNN were applied to each combination. This required signiﬁcant computing resources and was done using 15 CPUs, each with 12 cores, of the Oakley cluster at the Ohio

Supercomputer Center (OSC, 1987) . A second set of sensitivity analysis was performed to maximize the accuracy of each classiﬁcation model.

RESULTS

Using all 18 inputs shown in the input layer of Fig. 2 and described earlier KNN, PNN and EPNN yielded average classiﬁcation accuracies of 34.3%, 71.4%, and 74.3%, respectively. This indicates that when examining individual predictors in isolation without accounting for the complex interactions between them, we can only predict motor outcome with a maximum of about 74.3% accuracy. Table 4 shows the effectiveness rank of the 18 inputs on the prognosis accuracy based on the aforementioned sensitivity analysis using the EPNN model. The most necessary inputs were WMFT lift paper clip (ﬁne motor) following by WFMT extend elbow (gross motor). The most redundant input was WMFT forearm to box (gross motor) followed by TM

(somatosensory). The somatosensory inputs, BKT and TM, ranked 12 and 17, 57 respectively. They were among the inputs with negative or impartial effects on the accuracy of prognosis.

Table 4 Effectiveness rank of the 18 inputs on the results based on the sensitivity analysis using the EPNN model Inputs Accuracy % Changes in accuracy % Log-likelihood ratio Rank 7.WMFT forearm to box (gross motor) 80.0 +5.7 N/A 18 18.TM (somatosensory) 77.1 +2.9 0.0000 17 9.WMFT hand to box (gross motor) 77.1 +2.9 0.1654 16 14.WMFT stack checkers (fine motor) 77.1 +2.9 1.0000 15 10.WMFT hand to table (gross motor) 74.3 0.0 0.0000 14 8.WMFT forearm to table (gross motor) 74.3 0.0 0.2936 13 17.BKT (somatosensory) 74.3 0.0 0.6168 12 11.WMFT key (fine motor) 74.3 0.0 0.6760 11 5.WMFT extend elbow weight (gross motor) 74.3 0.0 0.9813 10 6.WMFT flip cards (fine motor) 74.3 0.0 0.9881 9 2.WMFT basket (fine motor) 74.3 0.0 1.0000 8 15.WMFT reach retrieve (gross motor) 71.4 -2.9 0.0000 7 3.WMFT can (fine motor) 71.4 -2.9 1.0000 6 16.WMFT towel (fine motor) 68.6 -5.7 0.0000 5 13.WMFT pencil (fine motor) 68.6 -5.7 0.4277 4 1.Stroke affected side 68.6 -5.7 1.0000 3 12.WMFT lift paper clip (fine motor) 62.9 -11.4 0.0000 2 4.WFMT extend elbow (gross motor) 62.9 -11.4 1.0000 1 WMFT: Wolf motor function test; BKT: Brief kinesthesia test; TM: Touch monofilament; EPNN: Enhanced probabilistic neural networks

Next, when using all possible combinations of the 18 inputs, KNN, PNN and EPNN yielded maximum classiﬁcation accuracies of 71.4% (3 out of 262,125 combinations),

85.7% (8 out of 262,125 combinations), and 100% (52 out of 262,125 combinations), respectively. This sensitivity analysis took about 5 h on the aforementioned supercomputer. Since EPNN outperforms the other two classiﬁcation algorithms, results were presented in Table 5 for EPNN only for 52 different combinations of inputs with average accuracies of 100%. The most frequently selected inputs among different combinations were WMFT towel (49 times), WMFT pencil (41 times), and WMFT lift paper clip (39 times), all of which include ﬁne motor movement. The two somatosensory inputs, BKT and TM, appear 37 and 19 times in Table 5. These are among the most frequently chosen items. 58

Table 5 Different combinations of 18 inputs for EPNN resulting in an average accuracy of 100% Inputs LR # 1. SA 2. BT 3. CN 4. EE 5. EW 6. FC 7. FB 8. FT 9. HB 10. HT 11. KY 12. LC 13. PL 14. SC 15. RR 16. TL 17. BK 18. TM 1 0 0 0 0 1 1 1 0 0 0 0 1 1 0 1 1 1 0 0.000 2 0 0 1 1 1 0 1 0 1 0 0 1 1 0 1 1 0 0 0.118 3 0 0 0 0 0 1 1 0 1 0 0 1 0 0 0 0 1 1 0.174 4 1 0 1 0 1 0 0 1 0 1 0 0 1 0 1 0 0 1 0.178 5 0 0 1 1 1 0 1 0 1 1 0 1 1 0 1 1 0 0 0.218 6 0 0 0 0 1 0 1 1 0 0 1 1 1 0 0 1 1 0 0.228 7 0 0 0 0 1 0 1 1 0 0 0 1 1 0 0 1 1 0 0.233 8 0 0 0 0 1 0 1 1 0 0 0 1 1 0 1 1 0 0 0.238 9 0 0 0 0 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0.252 10 0 0 0 0 0 0 1 0 1 0 0 1 0 0 0 1 1 1 0.256 11 0 0 1 0 1 0 1 1 0 0 0 1 1 0 0 1 1 0 0.263 12 0 0 0 0 1 0 1 0 1 0 0 1 1 0 1 1 0 0 0.276 13 0 1 0 0 0 0 1 0 0 0 0 1 1 0 1 1 1 1 0.29 14 1 0 1 0 1 0 1 1 0 0 1 0 1 0 0 0 0 0 0.293 15 1 1 0 0 0 0 0 0 1 1 0 0 0 0 1 1 0 1 0.302 16 0 0 0 0 1 0 1 1 1 0 1 1 1 0 0 1 1 0 0.314 17 0 0 0 0 1 0 0 1 1 1 0 1 1 0 0 1 1 0 0.334 18 0 0 0 0 1 0 1 1 1 0 0 1 1 0 0 1 0 0 0.339 19 0 0 1 0 1 1 1 1 0 1 1 0 1 0 0 1 0 0 0.342 20 1 0 0 0 0 1 0 1 1 1 0 0 1 0 1 1 1 1 0.344 21 1 0 0 0 0 0 0 1 1 0 0 0 1 0 1 1 1 1 0.346 22 0 0 0 0 1 0 1 1 0 1 1 1 1 0 0 1 1 0 0.351 23 1 1 0 0 0 0 0 0 1 1 1 0 1 0 1 1 0 1 0.352 24 0 0 1 0 1 0 1 1 1 0 0 1 1 0 0 1 1 0 0.356 25 0 0 0 0 1 0 1 1 0 0 0 1 1 0 1 1 1 0 0.358 26 0 0 0 0 1 0 1 0 0 1 0 1 1 0 1 1 1 0 0.358 27 0 0 0 0 1 0 1 1 0 1 0 1 1 0 0 1 1 0 0.358 28 0 0 0 0 1 0 0 1 1 0 0 1 1 0 1 1 1 0 0.36 29 0 0 0 0 1 0 1 1 0 1 0 1 1 0 1 1 0 0 0.36 30 0 0 0 0 1 0 1 1 1 0 0 1 1 0 0 1 1 0 0.361 31 0 0 1 0 1 0 1 1 0 1 0 1 1 0 0 1 1 0 0.362 32 0 1 0 0 0 0 1 0 0 1 0 1 1 0 1 1 1 1 0.384 33 0 0 0 0 1 0 1 0 1 1 0 1 1 0 1 1 0 0 0.399 34 0 0 0 0 0 0 0 0 1 1 0 1 0 1 1 1 1 1 0.404 35 0 0 0 0 0 0 0 1 1 0 0 1 0 1 1 1 1 1 0.407 36 0 1 0 0 0 0 1 1 0 0 0 1 1 0 1 1 1 1 0.436 37 0 0 0 0 0 0 1 0 0 1 1 1 0 1 1 1 1 1 0.442 38 0 0 1 0 1 1 1 1 0 0 1 0 1 0 0 1 0 0 0.456 39 0 0 0 1 0 1 1 0 1 0 0 1 1 0 1 1 1 0 0.469 40 0 1 0 0 0 0 0 0 1 1 0 1 0 0 1 1 1 1 0.47 41 0 0 0 0 1 0 0 1 1 1 0 1 1 0 1 1 1 0 0.47 59

Tale 5 (Cont.) # 1. SA 2. BT 3.CN 4. EE 5.EW 6. FC 7. FB 8. FT 9.HB 10.HT 11.KY 12. LC 13. PL 14. SC 15. RR 16. TL 17. BK 18. TM LR 42 0 0 0 0 1 0 1 1 0 1 0 1 1 0 1 1 1 0 0.479 43 0 0 0 1 0 0 1 0 1 0 1 1 1 0 1 1 1 0 0.484 44 1 1 0 0 0 0 0 1 1 1 0 0 0 0 1 1 0 1 0.49 45 0 1 0 0 0 0 1 1 0 1 0 1 1 0 1 1 1 1 0.515 46 0 1 0 0 0 0 1 0 1 1 0 0 0 0 1 1 1 1 0.553 47 0 0 0 0 0 0 0 1 1 1 0 1 0 1 1 1 1 1 0.559 48 0 0 0 1 0 1 1 0 1 1 0 1 1 0 1 1 1 0 0.568 49 0 0 0 1 0 1 1 0 1 0 1 0 1 1 1 1 1 0 0.623 50 0 0 0 1 0 0 1 0 1 1 1 1 1 0 1 1 1 0 0.634 51 0 0 0 1 0 1 1 0 1 1 1 0 1 1 1 1 1 0 0.736 52 1 1 0 0 0 0 1 1 0 1 0 0 0 0 0 1 0 1 1.000 TS 8 10 9 8 28 9 38 30 30 27 12 39 41 6 34 49 37 19 N/A EPNN: Enhanced probabilistic neural networks; C#: Combination number; SA: Stroke affected side; BT: WMFT basket; CN: WMFT can; EE: WFMT extend elbow; EW: WMFT extend elbow weight; FC: WMFT flip cards; FB: WMFT forearm to box; FT: WMFT forearm to table; HB: WMFT hand to box; HT: WMFT hand to table; KY: WMFT key; LC: WMFT lift paper clip; PL: WMFT pencil; SC: WMFT stack checkers; RR: WMFT reach retrieve; TL: WMFT towel; BK: BKT; WMFT: Wolf motor function test; TM: Touch monofilament; OT: Output; TS: Times selected; LR: Likelihood ratio.

DISCUSSION

The present work demonstrated that EPNN can be effective in predicting the response to CI therapy in a relatively small group of individuals with chronic stroke. In this study, the individual items that most strongly predicted improvement on the WMFT following

CI therapy were lifting a paper clip (ﬁne motor) and extending the elbow (gross motor).

The most redundant items, those that had negative effect on the prediction power of the model, were WMFT forearm to box (gross motor), TM (somatosensory), WMFT hand to box (gross motor), and WMFT stack checkers (ﬁne motor).

Somatosensory items in this predictive model were the BKT, thought to represent kinesthetic performance, and TM, the touch perception threshold. Touch sensation was not predictive of improved motor function following CI therapy to a meaningful extent.

Though present in some models, TM was much less prevalent amongst the most parsimonious models (those with the fewest predictors) and was a poor contributor to the overall accuracy of these models.

The BKT somatosensory measure, on the other hand, contributed equally to outcomes prediction as the individual motor items within each model. Contrary to our hypothesis, those participants with poor proprioceptive sense at baseline tended to show larger motor improvement than those with relatively good proprioception at baseline.

This is likely due to motor contamination within the measure itself. Individuals with poorer motor performance tended to score poorly on the BKT, which reﬂects that the 61

BKT measures motor precision as well as proprioceptive awareness. 100% accuracy can also be achieved absent the BKT, arguing against its relevance within a predictive sensorimotor testing battery.

The second sensitivity analysis revealed the best combination of inputs to achieve an accuracy of 100% for prediction of the extent of motor recovery. The most parsimonious combination required a maximum of 6 predictors were required for accuracies of 100%. A signiﬁcant conclusion of this research is that a high predictive accuracy can be achieved using only a small sampling of motor tasks that can be very quickly administered. After replicating on a larger sample, open-sourcing this computational model will afford very cost-efﬁcient ways of estimating rehabilitation prognosis and treatment planning.

The most inﬂuential inputs (predictors) obtained from the ﬁrst sensitivity analysis

(Table 4) may or may not be seen in the best combinations (Table 5) because, although they are effective independently, they may not help increase the accuracy of a combination of predictors in a high dimensional Euclidean space. Due to the high nonlinearity of the classiﬁcation problem, the direction of effect of each input cannot be easily recognized by either clinicians or conventional statistical tools. However, some trends did emerge when graphically examining the raw data. Overall, those with the poorest motor function at baseline, particularly on ﬁne motor function tasks (paper clip and towel) tended to show the largest motor improvements, whereas those with higher motor ability at baseline tended to show more limited motor improvement. It is important to note, however, that following these generalities cannot enable perfect

62 prediction of treatment response. The EPNN model presented in this paper is capable of recognizing the highly complex interactions of 18 input variables (Table 5) in a multidimensional Euclidean space to determine the likely response of an individual participant to the therapy with nearly perfect accuracy.

Finally, it is important to note that this paper only examines the impact of somatosensation on motor restoration. Though CI therapy is known to aid motor restoration, its effect on everyday motor behavior is typically even more profound

(Taub et al., 2006; Taub et al., 1993; Taub et al., 2006b; Taub et al., 2013; Uswatte &

Taub, 2013) . For example, non-responders on the WMFT still showed clinically meaningful improvements in arm use for daily activities, as measured by the Motor

Activity Log. In this dataset, there was no significant relationship (P = .78) between treatment-induced improvement in arm use and improvement in amount/quality of arm use for daily activities. Given that motor function and arm use appear to be independent constructs, CI therapy may still be an appropriate approach for many potential participants for whom the model indicates that a restorative approach is not indicated. In these cases, however, an approach that more heavily emphasizes the carry- over (transfer package) components of CI therapy and less heavily emphasizes motor practice may be more beneficial. Additional research is required to determine the minimal “dose” of motor training that is required to promote carry-over to daily activities amongst individuals unlikely to benefit from CI therapy in terms of motor restoration. Future work will also determine the extent to which the findings presented here generalize to a larger sample of participants and to other restorative motor

63 treatment paradigms.

CONCLUSION

EPNN was able to predict extent of motor recovery following CI therapy with 100% accuracy utilizing only 6 items from the WMFT. The somatosensory measures used here did not enhance the accuracy of predicting extent of motor restoration over motor testing alone.

ACKNOWLEDGEMENTS

Financial support for data analysis was obtained through The Ohio State University

Ofﬁce of the Provost Chronic Brain Injury Discovery Theme initiative. Data collection was supported by American Heart Association 12SDG12200013. Additional support was obtained from Grant UL1TR001070 from the National Center For Advancing Translational

Sciences. The content is solely the responsibility of the authors and does not necessarily represent the ofﬁcial views of the funding sources. Kala Phillips and Alli Hall are acknowledged for their contribution toward data collection. The sensitivity analysis computations were performed on the supercomputers at the Ohio Supercomputer Center.

ADDITION

Probabilistic and Enhanced Probabilistic Neural Networks

In probabilistic neural networks (PNN), the conditional probability of an I-dimensional data point 푿 belonging to the class or category of another I-dimensional data point 풀 can be represented in the form of a Gaussian probability density function (PDF):

1 ‖푿 − 풀‖ퟐ 훷(푿) = × exp (− ) (3) 퐼 2휎2 (2휋)2휎퐼 where 휎 is the spread of the Gaussian function (휎 ∈ [0 1]). Ahmadlou and Adeli (RW.ERROR -

Unable to find reference:309) present EPNN by improving Eq. 3 through the introduction of a modified spread, 휎푟 = 훼 × 휎, where 훼 (0 < 훼 ≤ 1) is the ratio of the number of datapoints with the class of 풀 in a local 퐼-dimensional decision hyper-spheres with the radius 푟 (푟 ∈ [0 1]) and center 풀:

1 ‖푿 − 풀‖ퟐ 훷(푿) = 퐼 × exp (− 2 ) (4) 퐼 2휎푟 (2휋)2휎푟

A higher value of 훼 (close to 1) indicates a higher probability of the embedded feature vectors in the hypersphere belonging to class of 풀 compared to other classes. If 풀 belongs to class 푐 ∈

{1, 2, … , 푀} where 푀 is the total number of classes, the average likelihood of 푿 belonging to the same class 푐 is:

푁푚 1 푃푚(푿) = ∑ 훷푛(푿) (5) 푁푚 푛=1 where 푁푚 is the number of data points in the available training set in class 푚. The estimated class of 푿 is the one with the maximum average likelihood: 65

퐶푙푎푠푠 표푓 푿 = arg max {푃푚(푿)} (6) 푚

Log-likelihood ratio

For a testing data point in a data set, 푀 average likelihoods are computed using Eq. 5, each representing the probability of that testing data point being classified into one of the 푀 classes.

Eq. 6 provides a rigid classification rule where a slight variation in the magnitude of average likelihood results in a class change without considering any uncertainty in the data sample..

In this case, the certainty of classification is not high although it might have led to correct classification. The larger the difference between the magnitude of the average likelihood corresponding to a class compared to those of other classes, the higher is the certainty of the testing data being in that class. This concept is used to evaluate the certainty of EPNN in assigning a class to a testing data point. In addition, the classification of a correctly-classified testing data point can be compared to other correctly-classified testing data points in the same data set through an index, called log-likelihood ratio, explained in the following paragraphs.

Suppose there are 퐿 different data sets all yielding the same accuracy for a particular RTT using repeated random sampling (for example, in Table 4, L=10 datasets yield the same accuracy of 72.7%). Consider the 푙푡ℎ data set, where 푙 ∈ {1, 2, … , 퐿}, is divided into training and testing subsets randomly 푁 times (equal to 35 for the stroke patients database used in this research with one used for the testing). EPNN is trained and tested 푁 times. If 푁푙 testing data points are classified correctly, there are 푁푙 values of max {푃푚(푿)} (Eqs. 5 and 6), for all training and 푚

푡ℎ testing subsets. If 푃푙,푛 is the 푛 value of max {푃푚(푿)} where 푛 ∈ {1, 2, … , 푁푙}, the following 푚

66 index, called log-likelihood ratio, is defined as follows as an overall measure of the certainty of the classification of correctly-classified data points:

푁푙 푁푙 ∑푛=1(푃푙,푛) ∑푛=1(푃푙,푛) log10 [ ] − min {log10 [ ]} 푁푙 푙 푁푙 푅 = (7) 푙 푁푙 푁푙 ∑푛=1(푃푙,푛) ∑푛=1(푃푙,푛) max {log10 [ ]} − min {log10 [ ]} 푙 푁푙 푙 푁푙

The value of 푅푙 varies between 0 and 1. A lower value means a lower certainty in classification.

CHAPTER 5: GROSS MOTOR ABILITY PREDICTS RESPONSE TO UPPER EXTREMITY REHABILITATION IN CHRONIC STROKE.

ABSTRACT

The majority of rehabilitation research focuses on the comparative effectiveness of different interventions in groups of patients, while much less is currently known regarding individual factors that predict response to rehabilitation. In a recent article, the authors presented a prognostic model to identify the sensorimotor characteristics predictive of the extent of motor recovery after Constraint-Induced Movement (CI) therapy amongst individuals with chronic mild-to-moderate motor deficit using the enhanced probabilistic neural network (EPNN). This follow-up paper examines which participant characteristics are robust predictors of rehabilitation response irrespective of the training modality. To accomplish this, EPNN was first applied to predict treatment response amongst individuals who received a virtual-reality gaming intervention (utilizing the same enrollment criteria as the prior study). The combinations of predictors that yield high predictive validity for both therapies, using their respective datasets, were then identified. High predictive classification accuracy was achieved for both the gaming

(94.7%) and combined datasets (94.5%). Though CI therapy employed primarily ﬁne-motor training tasks and the gaming intervention emphasized gross-motor practice, larger improvements in gross motor function were observed within both datasets. Poorer gross motor ability at pre-treatment predicted better rehabilitation response in both the gaming and combined

68 datasets. The conclusion of this research is that for individuals with chronic mild-to-moderate upper extremity hemiparesis, residual deﬁcits in gross motor function are highly responsive to motor restorative interventions, irrespective of the modality of training.

INTRODUCTION

Motor restorative therapies aim to restore motor function by emphasizing practice with the more affected upper extremity while minimizing compensatory movement by the less affected upper extremity. Recently, the authors showed that this therapeutic approach may not be appropriate for all individuals who have sufficient motor ability to participate (George et al.,

2017a). George et al. (2017a) presented a novel prognostic computational model to identify which baseline sensorimotor characteristics predicted the extent of motor recovery during

Constraint- Induced Movement (CI) therapy, an established motor restorative intervention(Taub et al., 2006; Taub et al., 1993; Taub et al., 2006b; Wolf et al., 2006b), employing the enhanced probabilistic neural network (EPNN) model of Ahmadlou and Adeli (Ahmadlou & Adeli,

2010a). They found that the extent of motor restoration, as measured by the Wolf Motor Function

Test (WMFT) (Taub et al., 2006; Taub et al., 1993; Taub et al., 2006b; Wolf et al., 2006b), varied markedly among individuals and was generally poor amongst those with higher baseline ability.

The purpose of this follow-up research is to determine robust predictors of motor restoration irrespective of the type of motor training. This is accomplished by applying the aforementioned machine learning-based model to a very different treatment modality: motor training delivered at home via Recovery Rapids, a Kinect-based video game (Maung et al., 2013). Like CI therapy, this motor restorative video game-based intervention involves high repetition practice with the more affected upper extremity for several hours per day over two weeks, progressive shaping of 69 motor tasks, and an emphasis on carry-over of motor gains to daily activities (L. V. Gauthier et al., 2017b; Morris et al., 2006). There are several important differences between the two types of therapies, however. Recovery Rapids harnesses the benefits of a virtual world (i.e., no task set-up time) to dramatically increase task variability. As such, the client switches rapidly between different types of motor movements. In contrast, CI therapy utilizes blocked practice, in which the same task is practiced repeatedly for a period of about 10–20 min. Game-based therapy through Recovery Rapids also involves substantially more repetitions per time (> 1000 per hour on average), is largely delivered at home without direct therapist supervision, incorporates limited tactile feedback (participants do not touch objects), and distal (fine-motor) training comprises a smaller percentage of tasks (∼30% versus > 90%). The authors hypothesized that there are likely to be some commonalities in individual sensorimotor presentation that would make that individual a better candidate overall for motor restorative therapies, irrespective of therapeutic modality. Additionally, they expected that training-related factors would interact with individual characteristics to produce different patterns of poor versus good responders for the two different interventions. Specifically, they hypothesized that those with poorer function on the domain being trained would benefit more from the intervention.

Consistent with this hypothesis, George et al. (2017a) found that those with relatively greater

fine-motor ability at baseline benefited less from CI therapy, an intervention that targets fine motor tasks. In keeping with this hypothesis and the authors’ prior findings, we hypothesized that those with poorer gross-motor performance at baseline would be better candidates for gaming therapy, as this approach does not provide as many fine-motor training opportunities. To test this hypothesis, the best combination of predictors from the prior paper will be compared

70 with an identical analysis for the Recovery Rapids gaming therapy. To determine which elements of sensorimotor presentation predict more favorable outcome irrespective of therapeutic modality, the most predictive combinations of baseline sensorimotor ability for both gaming therapy and CI therapy will be identiﬁed.

This research aims to identify those individual characteristics at baseline that can predict response to two diﬀerent motor restorative therapies. Improved predictions of treatment response based on a person’s individual characteristics at baseline will enable therapists to devise cost- eﬃcient personalized care plans with the goal of balancing restorative versus compensatory intervention approaches to maximize the motor functions of their patients.

METHODOLOGY

Participants

Participants were 19 individuals with chronic (> 6 months) mild to moderate upper extremity hemiparesis who had experienced a stroke of any etiology. All participants met the motor inclusion criteria utilized in the EXCITE trial of CI therapy (Wolf et al., 2006a) but were enrolled largely irrespective of cognitive or mobility status. The sample utilized in this analysis is thus more inclusive than in prior CI therapy trials. Those who were unable to provide informed consent, or who had received Botox treatment in the past 12 weeks were excluded. Inclusion criteria and recruitment approaches for this study were the same as those used in the earlier study by the authors (George et al., 2017a). See Table 6 for participant demographics.

Table 6: Demographic and clinical characteristics of 19 participants in gaming therapy.

Item Number of case (Average) Minimum Maximum

Age 47.5 14.1 69.6

Sex 8 females 14.1 69.6

Chronicity (years) N/A 0.55 5.34

Stroke affected side 8 left N/A N/A

Handedness At least 6 right N/A N/A

Affected side was dominant At least 4 participants N/A N/A

Sex 8 females N/A N/A Intervention

The gaming therapy intervention was designed such that physical/ occupational therapists manage patients in a consultative role with the majority of the motor practice occurring through

Recovery Rapids, an in-home gaming rehabilitation system (Maung et al., 2013). The gaming system utilizes the KinectOne™ sensor to capture particular therapeutic movements (gestures), each of which is tied to a game objective. Gestures include elbow ﬂexion/extension, shoulder

ﬂexion with elbow extension, shoulder abduction, shoulder adduction, overhead reaching, forearm supination, grasp release, and wrist extension. The CI therapy principal of shaping

(progressively increasing task difficulty as a person improves) is incorporated. In just one example, the user attempts to capture parachutes as they fall from above. An introductory difficulty level for this gesture may require only 30 ° of shoulder flexion. As a user demonstrates the capacity to perform more difficult movements, the software requires greater shoulder flexion, then increased concurrent elbow extension and forearm supination to accomplish the same game 72 objective. See Fig. 13 for a depiction of the gaming environment (http:// gamesthatmoveyou.com/). Carry-over of motor improvements to daily life is promoted through an interactive Motor Activity Log problem- solving module that occurs after each 15–20 min of the game play.

Figure 13: Depiction of the gaming environment Five therapist/patient contact hours occurred over 4 home visits. The ﬁrst session (2h) involved instruction in game play, customizing the game to the participant, establishing the treatment contract, and establishing the home program (target functional activities to accomplish daily). Thereafter, sessions focused on review of progress with the home program, modifying game customization as needed, and on “transfer package” elements that could not be readily addressed through the game (Morris et al., 2006; O’Sullivan et al., 2017). “Transfer package” elements include re- viewing the treatment contract, daily self-assessment of arm use, guided problem-solving to increase the use of the weaker upper extremity for activities of daily living,

73 and collaboratively establishing a home program focused on functional task practice.

Participants agreed to play Recovery Rapids for 30 h over a two-week period.

Outcome measures

Three outcome measures were utilized: the WMFT, the Brief Kinesthesia Test (BKT), and

Touch Test Monoﬁlaments (TM). The WMFT was utilized to assess the motor function of the upper limbs (Taub et al., 2006; Taub et al., 1993; Taub et al., 2006b; Wolf et al., 2006b). As in

George et al. (2017a), the WMFT scores, recorded in seconds, were natural-log-transformed to account for the non-uniform interpretation of performance time improvement (i.e., an improvement from 5 s to 3 s is greater than an improvement from 105 to 103s). The BKT is a measure of error in guided reaching with visual occlusion considered to represent upper limb kinesthetic sense (A. Borstad & Nichols-Larsen, 2016b). TM is sensitive to tactile impairment; it identiﬁes the lightest "force" in grams perceived consistently by an individual on the index

ﬁnger (Callahan, Hunter, & Mackin, 1995). These same sensorimotor measures were used in the authors’ earlier research (George et al., 2017a) and are summarized in Table 7.

Table 7: Behavioral measures used in the prognostic model

Number Fine or Type of test Functions assessed Behavioral Measures of items Gross motor Upper limb functional performance Motor 15 Forearm to table (side) Gross (timed) Forearm to box (side) Gross Extend elbow (side) Gross Hand to table (front) Gross Hand to box (front) Gross Extend elbow weight Gross Reach and retrieve Gross Lift can Fine Lift Pencil Fine Lift paper clip Fine Stack Checkers Fine Flip card Fine Turn key in lock Fine Fold towel Fine Lift basket Fine Reaching error with visual occlusion Brief kinesthesia test N/A Somatosensory 2 (proprioception) (affected side) Touch test Touch perception threshold monofilaments N/A (affected side)

Table 8 summarizes the collected patient data used for the prognostic computational EPNN model. From these data, there were 2 missing values corresponding to the somatosensory measures of only one participant. These were replaced using a simple regression analysis. Each participant was categorized based on their natural-log-transformed WMFT treatment change score as either a non-responder (> −0.15; class 1), moderate-responder (−0.15:−0.40; class 2), or best responder (< −0.40; class 3). Classiﬁcation thresholds are consistent with the earlier research (George et al., 2017a). These categories are represented in the last column of Table 8.

A histogram of WMFT change is shown in Fig. 14.

Figure 14: : The histogram of the participants' change in the natural log of WMFT scores from pre to post gaming therapy

Table 8: Participant data used to train and test the EPNN model

Predictors P# OT 1. SA 2. BT 3. CN 4. EE 5. EW 6. FC 7. FB 8. FT 9. HB 10. HT 11. KY 12. LC 13. PL 14. SC 15. RR 16. TL 17. BK 18. TM 1 0 1.16 1.48 0.00 0.10 2.62 0.26 0.34 0.41 0.34 2.17 1.81 1.44 1.74 0.34 2.12 1.84 -0.92 2 2 1 1.90 1.28 0.10 -0.22 3.31 0.64 0.11 0.47 0.53 2.37 3.87 2.13 3.48 0.18 3.99 1.39 -3.22 2 3 0 3.48 2.14 1.60 1.39 3.66 1.03 0.71 1.15 0.94 3.03 1.40 3.12 4.79 0.99 3.04 3.22 1.39 1 4 0 1.86 1.41 0.59 0.41 2.47 1.06 0.26 0.47 0.26 1.46 1.03 1.06 1.82 0.59 2.56 1.90 0.69 3 5 1 0.99 0.73 0.46 0.77 2.41 0.83 -0.21 0.19 0.02 1.57 1.05 0.48 1.81 0.39 2.16 2.08 -0.92 2 6 0 1.83 3.33 1.00 1.48 4.79 0.58 -0.29 -0.08 -0.60 4.79 4.79 4.79 4.79 0.12 3.04 2.63 -0.92 2 7 1 3.68 4.79 0.34 0.00 4.79 1.41 0.34 0.00 0.18 4.79 4.79 3.47 4.79 0.88 4.79 3.37 0.10 1 8 1 3.00 4.79 0.39 1.90 4.79 0.21 0.38 0.41 0.36 4.79 4.79 4.79 4.79 4.79 3.70 2.32 -1.83 3 9 0 4.79 1.77 1.27 1.28 2.31 1.75 -0.40 -0.67 -0.31 2.10 2.25 1.11 1.95 0.63 2.92 2.22 -1.83 1 10 0 2.58 4.79 1.35 0.00 4.79 1.33 -0.08 1.04 0.41 4.79 4.79 4.79 4.79 0.00 4.75 2.80 -1.83 3 11 1 0.99 0.96 -0.11 0.18 1.97 -0.11 0.18 -0.11 -0.11 1.03 1.03 0.92 1.65 -0.11 2.35 2.05 -0.92 2 12 0 1.77 1.11 0.61 1.17 2.63 0.59 -1.24 0.68 0.52 1.54 1.09 0.97 1.91 0.19 2.40 2.00 -1.83 2 13 0 3.68 4.79 1.99 1.57 4.52 1.25 0.18 0.59 0.41 2.80 4.79 4.79 4.79 0.26 3.43 2.42 -3.22 1 14 0 1.56 2.35 1.88 1.83 3.14 0.56 0.13 0.50 0.69 3.33 1.95 2.26 4.33 1.84 3.58 2.46 -2.66 1 15 1 4.79 4.79 4.79 4.79 4.79 4.79 1.35 4.79 1.75 4.79 4.79 4.79 4.79 4.79 4.79 1.76 5.70 3 16 1 1.26 2.20 -0.07 -0.07 2.83 0.98 0.26 0.54 0.18 2.65 4.79 4.79 4.79 -0.53 3.54 3.05 1.39 1 17 0 0.88 0.25 -0.43 -0.25 1.59 0.00 0.10 -0.51 -0.46 1.50 0.10 0.41 1.43 -0.53 1.05 1.70 -1.83 1 18 1 1.18 1.14 0.03 4.79 4.79 0.11 0.39 0.03 0.00 2.12 4.79 4.79 1.91 0.47 2.43 1.69 -0.92 3 19 0 4.79 4.79 4.79 4.79 4.79 0.96 0.50 4.79 0.63 4.79 4.79 4.79 4.79 -0.42 4.79 1.39 -0.92 1 A N/A 2.43 2.57 1.08 1.36 3.52 0.96 0.16 0.77 0.30 2.97 3.09 2.93 3.43 0.78 3.23 2.23 -0.76 N/A S N/A 0.49 1.35 1.64 1.44 1.64 1.13 1.03 0.50 1.45 0.52 1.35 1.76 1.75 1.44 1.48 1.04 0.57 N/A EPNN: Enhanced probabilistic neural networks; P#: Participant number; SA: Stroke affected side; BT: WMFT basket; CN: WMFT can; EE: WFMT extend elbow; EW: WMFT extend elbow weight; FC: WMFT flip cards; FB: WMFT forearm to box; FT: WMFT forearm to table; HB: WMFT hand to box; HT: WMFT hand to table; KY: WMFT key; LC: WMFT lift paper clip; PL: WMFT pencil; SC: WMFT stack checkers; RR: WMFT reach retrieve; TL: WMFT towel; BK: BKT; WMFT: Wolf motor function test; TM: Touch monofilament; OT: Output; A: Average; S: CI therapy deviation.

Sensitivity analysis and prognosis model

In order to identify the best method of classiﬁcation, the author previously used three diﬀerent algorithms to classify participants: k- nearest neighbors (Siddique & Adeli, 2013) , the probabilistic neural network, and the enhanced probabilistic neural network (EPNN)(Ahmadlou

& Adeli, 2010b) and

Figure 15: Flowchart including the sensitivity analysis

78 reported the most accurate results by EPNN (George et al., 2017a). Therefore, EPNN was also utilized in this follow-up research. Using the available data a total of 262,125 combinations of 18 motor, somatosensory, and stroke-aﬀected side predictors exist.

Sensitivity analyses were performed to identify the prediction accuracy of each predictive model. The ﬂowchart of the sensitivity analysis is presented in Fig. 15. It consists of 8 steps and

3 decision diamonds. In step 1, one out the 262,125 combinations is selected. In step 2, each participant’s motor and somatosensory data are arranged based on the selected combination. In step 3, the data corresponding to a single participant is input to the EPNN (predictor layer in

EPNN). In step 4, the motor and somatosensory data of the single participant is compared statistically to the same predictor data of all other patients in the pattern layer of EPNN. In step 5, the summation layer determines how similar the single participant’s predictor data are to the average values in each of the three categories: non-, moderate-, and best-responders. In step 6, the predictor data are assigned the category with the maximum average similarity (decision layer in

EPNN). In step 7, if the participant was classiﬁed correctly, an accurate prognosis is counted.

Steps 3–7 are repeated for every patient. Once the data for all participants have been analyzed in this manner, the accuracy percentage of the prognosis

Corresponding to the selected combination is computed in step 8 as follows: (number of accurate prognosis/total number of participants)×100. This process is repeated for every combination

(the outer loop in Fig. 15).

Implementation

All possible combinations of motor and somatosensory predictors (Table 27) are

79 considered to identify the combinations with the most accurate prognosis. Because the stroke patient data available for training a sophisticated neural network classiﬁcation model (EPNN) were limited in this research, the model was trained and tested 19 times (equal to the number of stroke participants), each time using the data for a diﬀerent patient for testing and the remaining

18 sets of data for training (steps 3–6 in Fig. 15). This results in an RTT (rate of testing to training) of about 5.0%. The accuracy values reported in this research are the average of testing accuracies of the 19 trials (step 8 in Fig. 15).

RESULTS

Accuracy of the gaming models and rates of selection

Within the gaming therapy cohort, EPNN yielded maximum classiﬁcation accuracies of

94.7% for 8 out of the 262,125 combinations. The highest accuracy obtained was 89.5% for 80 out of 262,125 combinations. The most frequently selected predictor in the 8 combinations with the highest accuracy was WMFT forearm to table (gross motor), which was selected in all eight combinations. The next most frequently selected predictors comprised mainly gross motor predictors: WMFT extend elbow weight (gross motor), and WMFT hand to table (gross motor),

WMFT basket (gross motor), and stroke-aﬀected side, which were selected 6 times.

Somatosensory predictors, BKT and TM, were selected in 1 and 3 combinations, respectively.

Table 9 presents the 8 diﬀerent combinations of predictors with average accuracies of 94.7%.

Table 9: Different combinations of 18 predictors for EPNN resulting in an average accuracy of 94.7% in the current study (1: selected; 0: not-selected)

Predictors C# LR 1. SA 2. BT 3. CN 4. EE 5. EW 6. FC 7. FB 8. FT 9. HB 10. HT 11. KY 12. LC 13. PL 14. SC 15. RR 16. TL 17. BK 18. TM 1 1 1 0 0 1 0 0 1 1 1 0 0 0 0 0 0 0 1 0.00 2 1 1 0 0 1 0 0 1 0 1 0 0 0 0 0 0 0 1 0.08 3 1 1 0 0 1 0 1 1 0 1 0 0 0 0 0 0 0 1 0.12 4 1 1 0 0 1 0 0 1 0 1 0 0 0 0 0 0 0 0 0.59 5 0 0 1 0 0 0 0 1 1 0 0 0 0 0 1 1 0 0 0.61 6 1 1 0 0 1 0 1 1 0 1 0 0 0 0 0 0 0 0 0.62 7 0 0 0 0 0 0 0 1 0 0 0 0 0 0 1 1 1 0 0.76 8 1 1 0 0 1 0 0 1 1 1 0 0 0 0 0 0 0 0 1.00 TS 6 6 1 0 6 0 2 8 3 6 0 0 0 0 2 2 1 3 N/A EPNN: Enhanced probabilistic neural networks; C#: Combination number; SA: Stroke affected side; BT: WMFT basket; CN: WMFT can; EE: WFMT extend elbow; EW: WMFT extend elbow weight; FC: WMFT flip cards; FB: WMFT forearm to box; FT: WMFT forearm to table; HB: WMFT hand to box; HT: WMFT hand to table; KY: WMFT key; LC: WMFT lift paper clip; PL: WMFT pencil; SC: WMFT stack checkers; RR: WMFT reach retrieve; TL: WMFT towel; BK: BKT; WMFT: Wolf motor function test; TM: Touch monofilament; OT: Output; TS: Times selected; LR: Likelihood ratio.

Table 10 presents a comparison of the rates of selection of predictors for the combinations with accuracy of about 90% and greater using the CI and gaming therapies. For the gaming therapy, the most frequently selected predictors consist of a combination of gross, fine, and somatosensory predictors: WMFT forearm to table (gross motor), WMFT basket (fine and gross motor), WMFT towel (fine motor), stroke affected side, and TM (somatosensory), with rates of selection of 80.7%,

71.6%, 63.6%, 60.2%, and 55.7%, respectively. The least selected predictors consist of only fine motor tasks: WMFT stack checkers (fine motor), WMFT key (fine motor), WMFT pencil (fine motor), WMFT can, and WMFT lift paper clip (fine motor) with rates of selection of 0.0%, 0.0%,

3.4%, 12.5%, and 21.6%, respectively.

Table 10: Comparison of the rates of selection of predictors for the combinations with accuracy of about 90% and greater using the CI and gaming CI therapies Rate of selection (%) CI therapy Difference Predictors Current study (Hulbert at al. (Gaming –CI therapy) (Gaming therapy) 2017) 8.WMFT forearm to table (gross motor) 80.7 46.9 33.8 2.WMFT basket (fine motor) 71.6 57.4 14.2 16.WMFT towel (fine motor) 63.6 71.5 -7.9 1.Stroke Affected Side 60.2 29.6 30.6 18.TM (somatosensory) 55.7 35.3 20.4 10.WMFT hand to table (gross motor) 51.1 50.2 0.9 5.WMFT extend elbow weight (gross motor) 48.9 39.4 9.5 17.BKT (somatosensory) 43.2 59.8 -16.6 4.WFMT extend elbow (gross motor) 40.9 37.2 3.7 9.WMFT hand to box (gross motor) 36.4 56.9 -20.5 7.WMFT forearm to box (gross motor) 34.1 51.6 -17.5 6.WMFT flip cards (fine motor) 26.1 35.8 -9.7 15.WMFT reach retrieve (gross motor) 22.7 63.0 -40.3 12.WMFT lift paper clip (fine motor) 21.6 56.4 -34.8 3.WMFT can (fine motor) 12.5 52.4 -39.9 13.WMFT pencil (fine motor) 3.4 57.5 -54.1 11.WMFT key (fine motor) 0.0 43.6 -43.6 14.WMFT stack checkers (fine motor) 0.0 32.0 -32.0 WMFT: Wolf motor function test; BKT: Brief kinesthesia test; TM: Touch monofilament

As a point of comparison, in the earlier research for CI therapy, EPNN yielded maximum

classiﬁcation accuracies of 100% (52 combinations), with ﬁne motor tasks comprising the most

frequently selected predictors (George et al., 2017a). Selection rates for these, along with other

combinations achieving accuracies of at least 90% are also included in Table 5.

Table 11 summarizes the rates of selection of 18 predictors in the prognosis model for the

gaming and CI

Table 11: Rates of selection of 18 predictors in the prognosis model for the gaming and CI therapies, and the combined approach

Predictors R 3. #C AC # 1. SA 2. BT 4. EE 5. EW 6. FC 7. FB 8. FT 9. HB 10. HT 11. KY 12. LC 13. PL 14. SC 15. RR 16. TL 17. BK 18. TM CN Gaming 1 75.0 75.0 12.5 0.0 75.0 0.0 25.0 100.0 37.5 75.0 0.0 0.0 0.0 0.0 25.0 25.0 12.5 37.5 8 94.7 2 58.8 71.3 12.5 45.0 46.3 28.8 35.0 78.8 36.3 48.8 0.0 23.8 3.8 0.0 22.5 67.5 46.3 57.5 80 89.5 A 66.9 73.1 12.5 22.5 60.6 14.4 30.0 89.4 36.9 61.9 0.0 11.9 1.9 0.0 23.8 46.3 29.4 47.5 N/A N/A CI therapy 1 15.4 19.2 17.3 15.4 53.8 17.3 73.1 57.7 57.7 51.9 23.1 75.0 78.8 11.5 65.4 94.2 71.2 36.5 52 100 2 29.1 56.6 44.2 30.1 34.9 29.9 56.1 37.5 67.9 52.6 43.3 55.4 67.6 26.6 76.2 82.7 61.2 33.9 1014 97.1 3 27.7 60.0 49.8 32.7 36.0 31.0 53.5 45.7 61.2 50.7 43.2 56.6 62.0 30.7 66.9 75.3 62.0 35.4 4828 94.3 1505 4 30.4 56.8 54.0 39.3 40.7 37.8 50.7 48.0 54.7 49.8 43.7 56.3 55.3 32.8 60.9 69.5 58.9 35.3 91.4 2 A 25.6 48.2 41.3 29.3 41.4 29.0 58.3 47.2 60.4 51.3 38.3 60.8 65.9 25.4 67.3 80.4 63.3 35.3 N/A N/A Combined 1 100 100 0 0.0 100 0.0 100 100 0.0 100 0.0 0.0 0.0 0.0 0.0 0.0 0.0 100 1 94.5 2 100 100 0. 0.0 50.0 50.0 100 100 0.0 50.0 0.0 0.0 0.0 0.0 0.0 50.0 0.0 100 1 93.3 3 100 100 0.0 0.0 66.7 33.3 66.7 100 33.3 66.7 0.0 0.0 0.0 0.0 0.0 33.3 0.0 100 1 93.1 4 66.7 66.7 0.0 0.0 33.3 33.3 50.0 66.7 66.7 66.7 0.0 33.3 16.7 0.0 33.3 66.7 50.0 100 3 92.1 5 60.0 50.0 0.0 10.0 30.0 40.0 50.0 50.0 60.0 40.0 0.0 30.0 20.0 0.0 20.0 80.0 50.0 100 4 91.9 6 66.7 58.3 0.0 8.3 41.7 33.3 41.7 58.3 58.3 50.0 0.0 25.0 16.7 0.0 16.7 66.7 41.7 91.7 2 91.7 7 59.7 29.0 6.5 9.7 37.1 35.5 61.3 51.6 54.8 41.9 0.0 54.8 35.5 0.0 48.4 83.9 62.9 74.2 50 90.7 8 59.2 33.8 5.6 9.9 38.0 33.8 60.6 53.5 54.9 42.3 0.0 53.5 33.8 0.0 43.7 84.5 64.8 73.2 9 90.5 9 58.3 33.3 6.9 9.7 37.5 33.3 59.7 54.2 55.6 41.7 0.0 52.8 33.3 0.0 44.4 84.7 63.9 72.2 1 90.2 A 74.5 63.5 2.1 5.3 48.3 32.5 65.5 70.5 42.6 55.5 0.0 27.7 17.3 0.0 22.9 61.1 37.0 90.1 N/A N/A EPNN: Enhanced probabilistic neural networks; R#: Row number; SA: Stroke affected side; BT: WMFT basket; CN: WMFT can; EE: WFMT extend elbow; EW: WMFT extend elbow weight; FC: WMFT flip cards; FB: WMFT forearm to box; FT: WMFT forearm to table; HB: WMFT hand to box; HT: WMFT hand to table; KY: WMFT key; LC: WMFT lift paper clip; PL: WMFT pencil; SC: WMFT stack checkers; RR: WMFT reach retrieve; TL: WMFT towel; BK: BKT; WMFT: Wolf motor function test; TM: Touch monofilament; OT: Output; TS: Times selected; AC: Accuracy percentage; #C: Number of combinations associated with AC; A: Average. +

Robust predictors across intervention type

To identify the most robust sensorimotor predictors of motor restoration irrespective of training modality, 72 combinations from the combined CI therapy and gaming therapy data sets with average accuracies above 90%, were identified. To determine the direction of effect for each predictor, scatterplots of baseline scores versus change in the WMFT are created and shown in Fig. 16. This figure demonstrates how the selection rates of various predictors vary with overall accuracy of the model. The direction of effect was consistent for all predictors whereby poorer baseline performance yielded greater improvement. To illustrate this point, Fig. 17 shows the change in fine and gross motor performance as a function of the initial motor performance.

Figure 16: a) Change in average of natural logarithm of all WMFT gross motor tests from pre therapy to post therapy (R2 = 0.55) versus average of natural logarithm of all WMFT gross motor tests before therapy; b) Change in average of natural logarithm of

Figure 17: Change in both fine and gross motor performance as a function of initial motor performance for each task.

Parsimonious combinations

In the current study using gaming therapy, among the 8 combinations with the maximum accuracy of 94.7%, three were found to be the most parsimonious, meaning they achieved the highest accuracy with the fewest predictors. They are combination numbers 4, 5, and 7 in Table 9 (bolded and shaded). Combination number 7 includes only four predictors: WMFT forearm to table (gross motor), WMFT reach retrieve (gross motor), WMFT towel (ﬁne motor), and BKT (somatosensory). Combination number 5 includes 5 predictors: WMFT forearm to table (gross motor), WMFT hand to box (gross motor), WMFT reach retrieve (gross motor), WMFT can (ﬁne motor), and WMFT towel

(ﬁne motor). Combination number 4 also includes 5 predictors. Interestingly, this combination is contained in the remaining 5 combinations in Table 9. These predictors are: WMFT extend elbow weight (gross motor), WMFT forearm to table (gross motor),

WMFT hand to table (gross motor), WMFT basket (ﬁne and gross motor), and stroke aﬀected side.

Note that a parsimony analysis is irrelevant for the combined gaming and CI therapy approach because only one combination was found to have the maximum accuracy of

94.5% (ﬁrst row in Table 11 under the Combined section).

Sensitivity analysis on the parsimonious combinations

In the current study, in order to identify the influence of each selected predictor on the prediction accuracy, another sensitivity analysis was performed on the three most parsimonious combinations for gaming therapy. This sensitivity analysis is similar to that reported in the earlier work of the authors (George et al., 2017a). For each combination of interest, each selected predictor is removed, one at a time, and the 87 classification accuracy is computed each time. Those accuracies are then compared with the one combining all 18 predictors. For the first combination (combination number 4 in

Table 9), the model was run 5 times, each time removing one of the 5 included predictors. The accuracy of the prediction drops from 94.7% to 63.2%, 78.9%, 84.2%,

84.2, and 68.4%, by removing each predictor in turn one at a time: WMFT extend elbow weight (gross motor), WMFT forearm to table (gross motor), WMFT hand to table (gross motor), WMFT basket (ﬁne motor), and stroke aﬀected side respectively.

For the second combination (combination number 5 in Table 4), the model was run 5 times, each time removing one of the 5 included predictors. The accuracy of the prediction drops from 94.7% to 89.4%, 89.4%, 63.2%,73.7%, and 84.2%, by removing

WMFT forearm to table (gross motor), WMFT hand to box (gross motor), WMFT reach retrieve (gross motor), WMFT towel (ﬁne motor), and WMFT can (ﬁne motor), respectively.

For the third combination (combination number 7 in Table 4), the model was run 4 times, each time removing one of the 4 included predictors. The accuracy of the prediction drops from 94.7% to 84.2%, 63.2%, 73.7%, and 84.2% by removing WMFT forearm to table (gross motor), WMFT reach retrieve (gross motor), WMFT towel (ﬁne motor), and BKT (somatosensory), respectively.

A similar sensitivity analysis was performed on the single combination with the highest average accuracy of the combined approach. This combination includes 7 predictors. The accuracy of the prediction drops from 94.5% to 78.7%, 91.7%, 85.2%,

85.0%, 85.2%, 88.8%, and 78.5%, by removing WMFT extend elbow weight (gross motor), WMFT forearm to box (gross motor), WMFT forearm to table (gross motor),

WMFT hand to table (gross motor), WMFT basket (ﬁne and gross motor), TM

(somatosensory), and stroke aﬀected side, respectively. Again, among the motor predictors, the accuracy tends to drop equally or more upon removal of a gross motor predictor compared to ﬁne motor predictors.

DISCUSSION

The enhanced probabilistic neural network was used 1) to predict the extent of motor recovery following gaming therapy and 2) to investigate which baseline sensorimotor characteristics were robust predictors of motor restoration, irrespective of therapeutic modality. Concerning the first point, and consistent with our hypothesis, poorer baseline ability on the characteristics most heavily trained during the intervention (i.e., fine motor tasks for in-clinic CI therapy and gross motor tasks for gaming therapy) predicted greater motor restoration. That is, if a participant has a poor baseline score on gross motor tasks and undergoes gaming therapy, our models suggest that person will be a good responder to therapy and have good motor restoration. Accordingly, poorer baseline scores on fine motor tasks predicted better response to CI therapy. This effect does not result from different magnitudes of fine or gross motor improvement between the two interventions.

Concerning the second point, baseline ability on the gross motor tasks of the WMFT, compared to the fine motor tasks, seemed to be the most robust predictors of motor restoration across both datasets. This finding is consistent with Lee et al. (2015), who reported that in 174 chronic stroke patients, proximal joint movement at baseline could significantly predict improvement after both CI therapy (emphasizes fine motor movements) and a specialized robot-assisted therapy (emphasizes gross motor

89 movements). A possible explanation is that the most robust predictors of outcome are those that are most consistently measured over time (least susceptible to variability in performance). However, according to Fritz et al. (Fritz, Blanton, Uswatte, Taub, & Wolf,

2009) the most robust WMFT predictors in the combined approach had reliability

(intraclass correlation) coeﬃcients that were rank-ordered amongst the middle of the 15 items, lending little support for this potential explanation.

One partial explanation appears to be a ceiling effect, whereby poorer performers had a larger range of possible improvement. In support of this explanation, those with less gross motor ability at baseline were able to achieve the largest motor gains, which appear to be accounted for by dramatic gains in gross motor ability (Fig. 5). It appears that many individuals had not yet achieved their maximum potential even several years post-stroke, whereas those with better baseline ability may have already approached their maximal possible recovery. However, although ceiling effects explain why poorer baseline gross motor ability predicts better gross motor gains (i.e. the direction of prediction), they do not account for why baseline gross motor ability was a more robust predictor of motor restoration than baseline fine motor ability. In fact, there was more potential for motor restoration of fine-motor ability than gross motor ability; participant performance on fine motor tasks was worse overall and, moreover, fine-motor tasks on the WMFT are least susceptible to ceiling effects (Woodbury et al., 2010).

The most compelling reason for the selection of gross motor tasks being the most robust predictors of motor restoration, in the authors’ opinion, is that loss of ﬁne-motor control is inherently more diﬃcult to rehabilitate than loss of gross motor function. This idea is supported by the fact that amongst individuals who had considerable potential for

90 improvement at baseline, the extent of improvement on fine motor tasks was smaller than it was for the gross motor tasks. Using the formulation for computing improvement percentage proposed by Lin et al. (2009), amongst all participants, the average improvement on gross motor tasks was 28.91%, whereas average improvement on fine motor tasks was only 9.12%. It was important, however, to consider that some of our population started with baseline values near normal values (Wolf et al., 2006a) for some tasks, particularly for gross motor items. After excluding participants who performed within one standard deviation of normal ability at baseline, mean improvements on individual gross motor tasks were between about 40% and 60%, whereas mean improvements on individual fine motor tasks were typically between about 8% and 20%.

This trend is consistent with other reports in the literature (Lee et al., 2015; Myrhaug

& Ostensjo, 2014).

One physiological explanation for gross motor ability being easier to rehabilitate is that gross motor function can be mitigated by multiple neural pathways originating in diﬀerent regions of the brain (Baker et al., 2015; Lawrence & Kuypers, 1968), whereas ﬁne motor function is thought to be more locally controlled in the motor cortex and descending corticospinal tract (Hoffman & Strick, 1995; Kobayashi, Hutchinson, Schlaug, & Pascual-

Leone, 2003; Lang & Schieber, 2004). The reticulospinal tract is the most widely studied alternative pathway for gross motor control of the upper limbs (Davidson & Buford,

2004; Drew & Rossignol, 1984b; Hirschauer & Buford, 2015; Schepens & Drew, 2004).

In the cases of neurological insult to the corticospinal motor tract, this alternative pathway may facilitate recovery of gross motor function (Herbert, Powell, & Buford, 2015;

Hulbert, S. Adeli, H., Buford, J., 2015; Ortiz-Rosario et al., 2014). However, the

91 reticulospinal tract appears inefficient for producing fine motor movements, at best demonstrating involvement only in whole-hand grasping (Baker et al., 2015). Another alternative pathway for control of gross motor movements is the uncrossed rubrospinal tract, which may serve as a potential mechanism for the less affected hemisphere to contribute to recovery of the more affected upper extremity; however, its projections cannot be traced to spinal segments below C3 in humans, suggesting that the rubrospinal tract is more likely to influence motor neurons involved in proximal movement (Nathan

& Smith, 1982). In sum, there appears to be considerably more neurological substrate

(corticospinal pathways, reticulospinal pathways, and rubrospinal pathways) that can be harnessed for rehabilitation of gross motor function, whereas recovery of ﬁne motor ability is primarily mitigated by corticospinal pathways alone.

Study limitations

The main limitation is the small sample size of this study. Another limitation when comparing these findings with the authors’ earlier research (George et al., 2017a) is that both datasets involved a prospective cohort design and, as such, there is risk of selection bias when directly comparing findings from the two studies. Given these limitations, these findings will need to be replicated on a larger sample, such as that being currently collected in a multisite randomized controlled trial of the gaming therapy system versus CI therapy (L. V. Gauthier et al., 2017b). Despite these limitations, there is some consistency between the two datasets, namely that gross motor function is most strongly influenced by both treatments and that tactile information does not appear to be a robust predictor of neurorestoration following either game-based or CI therapy.

CONCLUSION

Results of this study suggest that those with near-normal gross motor function at baseline are least likely to benefit from motor restorative training. Though small improvements in fine-motor functioning can be realized through the motor restorative interventions studied here, the potential for dramatic improvement of fine motor abilities appears more limited than for gross motor abilities. These findings suggest that for individuals with near-normal proximal movement and residual distal impairment, interventions that focus primarily on overcoming non-use (such as the transfer package of CI therapy (Morris et al., 2006), rather than reducing impairment, may be the most appropriate. For individuals with mild to moderate impairment on gross motor tasks (and at least some distal upper extremity movement), interventions emphasizing intense motor practice in conjunction with overcoming non-use would appear to be highly beneficial.

ACKNOWLEDGEMENTS

Financial support for data analysis was obtained through The Ohio State University

Oﬃce of the Provost Chronic Brain Injury Discovery Theme initiative. Data collection was supported by American Heart Association 12SDG12200013 and Patient Centered

Outcomes Research Institute. Additional support was obtained from Grant

UL1TR001070 from the National Center For Advancing Translational Sciences. The content is solely the responsibility of the authors and does not necessarily represent the oﬃcial views of the funding sources. The computations were performed on the

93 supercomputers at the Ohio Supercomputer Center.

CHAPTER 6: DECODING NEURAL SIGNATURES ASSOCIATED WITH THE QUALITY OF ARM MOVEMENTS USING HETEROGENEOUS KINETIC AND EEG DATA CONCURRENTLY

ABSTRACT

Neural decoding of non-invasive electroencephalogram (EEG) has far reaching implications in the realms of brain computer interface (BCI) and neurofeedback, particularly for motor rehabilitation. In the past, several studies have demonstrated relatively robust neural decoding efforts for motor imagery, physical movements, and even movement prediction. However, more nuanced signal characteristics, such as indicators for the quality of movement, have not been addressed. Therefore, in this study, state-of-the-art combinations of machine learning techniques are employed to extract and classify features in the neural signal that are indicative of the quality of movement a person is about to make. Methods: The data are collected from a combination of position tracking data during self-paced game play of Recovery Rapids ™, a game-based motor restorative therapy platform, and non-invasive EEG recording. Wavelet transform is employed for denoising. A Deep Boltzmann Machines (DBM) is employed for feature extraction and is coupled with a Probabilistic Neural Network (PNN) and an Enhanced

Probabilistic Neural Networks (EPNN) for classification. Combinations of these networks are compared for their utility in accurately extracting and classifying neural signatures associated with movement quality. Results: A combination of DBM for feature extraction and EPNN for classification provide the highest classification accuracies of

70.39% for predicting “good” quality vs. “poor” quality before movement onset. The results presented here are promising for future neurofeedback investigations.

INTRODUCTION

According to the CDC, over 795,000 people suffer from a stroke every year and around 93% of those strokes will affect their ability to move. Specifically, it is reported that this life altering event “reduces mobility in more than half of stroke survivors age 65 and over” (Centers For Disease Control and Prevention, 2015). Moreover, in the United

States alone, billions of dollars are spent to treat and manage these brain injuries over the course of just one year (Center for disease control and prevention.2017). It is clear, then, that more resources and new methods for managing these injuries are desperately needed to help reduce costs and improve quality of life for survivors and care takers.

Thanks to recent advances in computing technology and utilization of these advances by highly interdisciplinary research teams, with backgrounds in engineering, physics, computer science, physical therapy and rehabilitation, new approaches to motor rehabilitation and brain computer interface (BCI) (Liu, Wang, Newman, Thakor, & Ying,

2017; Schudlo & Chau, 2018) are being developed for those with motor deficits (Burns,

Adeli, & Buford, 2014; Ortiz-Rosario & Adeli, 2013). Specifically, with recent advances in machine learning (Guo, Wang, Cabrerizo, & Adjouadi, 2017; Nagaraj, Lamperski, &

Netoff, 2017; Y. Zhang, Wang, Jin, & Wang, 2017), BCI (Fernández-Soto, Martínez-

Rodrigo, Moncho-Bogani, Latorre, & Fernández-Caballero, 2018; Jiao et al., 2018) and

EEG technologies (Chen, Zhao, Wang, Xu, & Gao, 2018; Wostyn et al., 2017), non- invasive, neural decoding (Karimimehr et al., 2017) is being utilized to improve several motor-based treatments, including motor imagery therapies (Kranczioch, Cornelia, Zich,

Catharina, Schierholz, Irina,Sterr, Annette, 2014; Taylor & Schmidt, 2012b) , motor

96 practice/rehabilitation (Bradberry, Gentili, & Contreras-Vidal, 2010; Ofner & Müller-

Putz, 2012), and movement intent (Bhagat et al., 2014; Hulbert, S. Adeli, H., Buford, J.,

2015; Sankari & Adeli, 2011). Specifically, decoding brain signals during motor imagery has seen some success. For example, Taylor et al. (2012b) classified neural signatures of single-trial based mental imagery with 87% accuracy. During EEG recording, these mental images consisted of various imagined motor activities including, push, pull, rotate, and lift.

While the utility of such motor imagery has been proven in BCI and robotic control applications (Chen et al., 2018; Lu, Chen, Zhang, Tong, & Zhou, 2017) to serve those with severe intractable motor deficits (Frolov et al., 2017), decoding neural signatures during actual motor practice may be useful for those with less sever motor deficits, such as in mild to moderate hemiplegia. In 2010, (Bradberry et al.) Sought to decode neural signatures associated with self-initiated reaching movement using EEG.

They reported peak correlations between measured and reconstructed velocity profiles in the x, y, and z directions to be 0.19, 0.38, and 0.32. Two years later, Ofner et al. (2012) fared even better when simply asking participants to move their hand in space, instead of reaching toward a specific target. With position tracking being completed using a consumer-based Microsoft Kinect, they reported correlations between the measured and reconstruction position profiles to be 0.70, 0.77, and 0.62 in the x, y, and z directions, respectively. Taken together, these studies have demonstrated the feasibly of decoding both real and imagined movements from the neural signature.

Taken a step further, the feasibility of decoding movement intent using EEG- based neural decoding prior to initiating physical movements (as opposed to imaginary

97 movements) has also been demonstrated. Ibáñez et al. (2015) demonstrated the feasibility of predicting movement intent (i.e. before movement onset) within 6 healthy individuals, using a combination of EEG and surface electromyographic (EMG) measurements for data collection, a genetic algorithm (Padillo, Luna, Herrera, & Ventura, 2018;

Valenzuela, Jiang, Carrillo, & Rojas, 2018; Wright & Jordanov, 2017) for feature extraction, and a Bayesian classifier, a type of probabilistic classifier (Attema,

Kosgodagan Acharige, Morales-Nápoles, & Maljaars, 2017; S. Y. Zhang, Jiang, & Neild,

2017), for movement classification. They report an average accuracy for decoding 7 different movements within the pre-movement EEG of 62.9%.

Therefore, analysis of non-invasive EEG has been well established for decoding real and imagined movements, as well as predicting intended movements of the upper limbs, not to mention a myriad of other neural decoding applications that range from robot-assisted gait for post-stroke rehabilitation (Contreras-Vidal et al., 2018), to perception of auditory signals (Billig, Davis, & Carlyon, 2018) to prediction of memory performance (Weidemann & Kahana, 2017). A useful next step that may be relevant during motor restorative therapies, is to determine if specific characteristics of the upcoming movements, such as movement quality, can be extracted from the signal.

Specifically, decoding movement quality before movement onset would mean decoding how well someone is about to execute a particular movement compared to normal movement as demonstrated by those with no motor deficit. If these characteristics can be extracted in the pre-movement onset period, they could be a basis for valuable feedback about the quality of motor planning (Li et al., 2017).

It has been shown that even within a motor practice session, a person’s ability to perform a task can fluctuate and the magnitude of trial-to-trial fluctuations typically exceeds the treatment effect for improvement in quality (Gauthier, 2018; Lang et al.,

2016). This means that there are times when an individual performs much better than others, a possible indication that their neural capacity to perform the movement well may also be fluctuating over time. The ability to explain this variation based on the quality of the motor planning that occurs before movement can allow a person to receive feedback on the quality of the motor plan itself, a substantial advance towards personalized rehabilitation.

Therefore, based on the success of aforementioned studies with decoding movements from non-invasive EEG signals, the authors hypothesize that there exist unique features of each person’s neural signal that also characterize the quality of movement a person is about to make. If these features exist and can be extracted and classified accurately, they will have broad implications for the future of personalized, neuro-feedback during motor practice. Specifically, through neurofeedback training based on identification of features associated with “good” movement and features associated with “poorer” movement during the motor planning stages, a person may learn to produce the brain activity patterns for motor planning that are associated with their best movement, thereby addressing fluctuations in neural capacity.

In this study, state-of-the-art combinations of machine learning techniques are employed to extract and classify features in the neural signal during self-paced movements of a person with chronic stroke that are indicative of the quality of movement a person is about to make. The data are collected from a combination of position tracking

99 data during self-paced game play of Recovery Rapids ™, a game-based motor restorative therapy platform, and non-invasive EEG recording. Specifically, a Deep Boltzmann

Machine was coupled with two different classification algorithms, Probabilistic Neural

Network (PNN) and an Enhanced Probabilistic Neural Network (EPNN) to determine the best combination of techniques for classifying movement quality.

METHODS

Participant: Case Study

The participant was a male with chronic (> 6 months) moderate hemiplegia on his left side after an ischemic stroke due to a basilar artery aneurysm.

Game Play

The participant played Recovery Rapids (Games That Move You, PBC), a motor restorative rehabilitation video game that makes use of the consumer-based motion capture technology, the Microsoft Kinect™. From a seated position, the participant controlled an avatar by making a series of pre-defined arm movements (Figure 18). These movements included shoulder flexion, shoulder abduction and adduction, elbow extension and flection, and pronation and supination of the wrist. These movements create the in-game gestures of rowing down the river, steering to avoid obstacles, catching bottles out of the river, and catching descending parachutes (Figure 18). Game play was conducted from a chair with no barriers to arm movement (e.g. arm rests).

Data Collection

A flowchart of the methodologies used in this research is presented in Figure 19.

Kinematic Data:

100

During game play, the Microsoft Kinect captures a time series of Euclidean spatial coordinates for 24 different locations spanning the entire body, at a sampling rate of up to 30Hz (Figure 20). 14 of these locations were used in the analysis that are relevant for upper limb movement, including hand (x2), thumb(x2), wrist(x2), elbow(x2), shoulder(x2), base of spine, torso, chest, and head (Figure 3). Once this raw skeletal data were collected, each gesture was parsed out from the time series using a series of custom

MATLAB scripts. Figure 21 shows a representative example of how the row gesture was parsed from a continuous recording of position data collected by the Kinect. In this figure, the small circle indicates the time that the gesture was recognized in the game

(triggered the game mechanic). The red curve demonstrates the complete movement from start to end. The blue line is position data captured before and after the movement to give context. The gray rectangle indicates the pre-movement time period investigated here.

This analysis is described in detail in Yang et al. (2018).

EEG Data:

Concurrently, EEG data were collected using the Cognionics HD-72 dry, wireless headset at 500 samples/sec. The cap consists of 64 active electrodes (A schematic of the electrode positions is provided in Figure 22. Once applied to the head, the electrodes were adjusted to make better contact with the scalp until the impedances were at or below

5,000 OHMS with most electrodes being below 2,500 OHMS, as indicated by the

Cognionics acquisition software. Because this cap uses dry electrodes, no saline gel or abrading of the scalp was necessary to lower impedance.

Data was transmitted from the cap through wireless high-speed Bluetooth to a nearby computer. A wireless trigger box was used to sync game play and EEG recording

101 such that the timestamps of gestures triggered by the game were marked automatically on the EEG recording.

Data Analysis

Kinect

Once each gesture was parsed from the continuous time series of Kinect skeleton data, a normalcy score was computed to compare the measured movement of the participant to a so-called “ideal” gesture generated previously by healthy controls with no motor deficit (Yang et al., 2018). The kinematic parameters to compute ideal movements included shape, speed, and range of motion. These normalcy indices were computed for every gesture and normalized to a number in the range of 0 to 1, where 1 indicates normal movement. Classes for “good” versus “poor” quality movement were chosen based on the distribution of normalcy residuals from the dose response curves shown for each of four gestures (row, bottle, parachute, and raisearm gestures) in Figure 23. The red curve in each subplot is a logarithm best-fit line for that person’s performance for each movement type. That means that the best-fit curve is indicative of a person’s expected performance over time. Note, however, that this best-fit curve is not indicative of the best quality of movement. Rather, larger values on the y-axis (larger normalcy indices) correspond to better movement. Therefore, residuals above the curve, those closer to normal (“good”) movement, were assigned class 1 and the residuals below the curve were assigned class 2. To create classes that were well-defined, the data points in figure 23 were Gaussian smoothed by convolving a 10 point Gaussian window with the data. The smoothed data are indicated by the green curve in figure 23.

EEG

102

EEG data were analyzed using a combination of custom MATLAB scripts and the open source software, EEGLAB (Delorme A., 2004). First, to ensure the efficacy of the

Fast Fourier Transform used later in the analysis, all data were resampled to 1500 samples/sec. Next, because the EEG electrodes are not adhered to the scalp and the participant performed large, gross motor movements, often with compensatory trunk motion, some electrodes that did not make good contact with the scalp were particularly prone to movement artifacts. Using a combination of visual inspection and the built-in kurtosis method in EEGLAB, the number of channels used in the analysis was reduced from 64 to 21 to reduce both these movement artifacts and blink artifacts. Since kurtosis measures changes in the shape of the probability distribution of the signal, it was useful for removing artifacts that cause abrupt changes in the signal, such as movement artifact

(Dai & Cao, 2017). The remaining 21 channels are highlighted in Figure 5. Then, they were re-referenced to the average of the remaining 21 channels (aka: average reference).

In this data set, it was important to reduce low frequency skin potential artifacts/drift and high frequency EMG or muscle activity. These EMG artifacts were particularly prevalent given high muscle tone in this participant (aka: involuntary muscle contraction). Therefore, the data were filtered using a wavelet transform. The utility of wavelets for filtering a wide variety of signals has recently been established in the literature (Adeli & Jiang, 2006; Dai, Zhang, & Wang, 2015; Li, Park, & Adeli, 2017;

Ortiz-Rosario, Adeli, & Buford, 2015). Briefly, a wavelet transform compares the signal of interest to the so-called mother wavelet. In so doing, it breaks the signal into different levels of detail and approximation. In this way, the signal is filtered of the noise and the resulting approximation signal is used for further analysis. In this study, a Daubechies 4

103 mother wavelet was chosen (Daubechies, 1990) and a wavelet analysis with 10 levels was conducted. This means that 10 levels of detail and one approximation signal were computed. With a 1500Hz sampling frequency, the approximation signal roughly corresponds to the 1.5 Hz component of the original signal and, therefore, largely represents any slow-wave, baseline shifts present in the data. This approximation signal was subtracted from the original signal, effectively removing these baseline shifts. Next, a 4 level wavelet analysis using the same mother wavelet was performed. The fourth level of detail coefficients roughly corresponds to signal components of 94Hz. Therefore, this method can be thought of as a sophisticated way to filter the signal from about 1.5Hz to 94Hz. The remaining approximation signal that was obtained after the second wavelet transform was then utilized for the rest of the analysis. Note that a bandpass filter from

0.5Hz to 50Hz was also tried as a filtering methodology, but was unable to reliably remove movement artifacts and blink artifacts, unlike the wavelet methodology described above. It also resulted in fewer data epochs that were available for classification at the end. Therefore, the results presented here are based on the data processed using the wavelet methodology for filtering.

Half second, pre-movement epochs were extracted from each session of EEG data. The corresponding kinematic data (see Figure 4 for an example) were used to define the EEG epochs as associated with “good” quality movement or “poor” quality movement, as defined by the normality index described above (class 1 or class 2, respectively).

Finally, every epoch of data was transformed from the time domain into the frequency domain and amplitude values were obtained using a fast Fourier transform

104

(FFT) for feature extraction. With a sampling frequency of 1500Hz and an epoch of 0.5 seconds, 750 total amplitudes were obtained from the two-sided FFT with 1Hz frequency resolution. Because each side of a two-sided FFT is a mirrored version of the other side, only values from the one-sided (first half) FFT were used. Specifically, inputs to the

DBM were the amplitude values of the first 250 amplitude values representing frequencies from the delta band well into the upper gamma band. Note that models containing all 375 frequencies from the one sided FFT were also tested, but found to produce the same classification accuracies. Therefore, to reduce the dimensionality of the data set and increase the computational efficiency of the DBM, only the first 250 P1 values from the FFT were used. Epochs of data were compiled into a single sample.

Figure 24 illustrates the entire EEG preprocessing procedure.

Machine Learning

EEG data were compiled into a single data set by combining all epochs associated with

“good” quality (class 1) movement and “poor” quality (class 2) movement. To avoid selection bias, the data were randomly permuted, separated into training and testing sets, and then run through a supervised Deep Boltzmann Machine for feature extraction.

(Hinton & Salakhutdinov, 2006; Rafiei, Khushefati, Demirboga, & Adeli, 2017) (Figure

8). DBM belongs to a class of machine learning algorithms known as deep neural network that has received a lot of attention in recent years for solution of different types of complicated dynamic pattern recognition problems from health monitoring of smart structures (Lin, Nie, & Ma, 2017; Rafiei et al., 2017), crack detection in pavements (A.

Zhang et al., 2017), and analysis of transportation networks (Nabian & Meidani, 2018) to image recognition (Koziarski & Cyganek, 2017; Ortega-Zamorano, Jerez, Juárez, &

105

Franco, 2017), mapping of scalp to intracranial EEG (Antoniades et al., 2018), assessment of deep brain stimulation in epileptic patients (Martín-López et al., 2017), early diagnosis of Alzheimer’s Disease (Ortiz, Munilla, Gorriz, & Ramirez, 2016), and differentiation of early-stage Creutzfeld -Jakob disease and rapidly progressive dementia

(Morabito et al., 2017). Briefly, DBM consists of an encoder and decoder, each with several layers (Figure 25). Each layer consists of a set of neurons that serve to extract features from the previous layer and reduce the dimensionality of the data. Training takes place as the encoder breaks down the input into a reduced dimensionality feature space and the decoder rebuilds the input from those features. Once the network stabilizes

(learns the best features for reconstruction) or the maximum number of learning iterations is reached (set at 400 in this research), the features are extracted and can be used for classification either within a softmax layer of the DBM itself, or a different classification network

In addition to the softmax layer of the DMB, two more classification models were employed in this study: probabilistic neural networks (PNN) (S. Y. Zhang et al., 2017) and Enhanced probabilistic neural networks (EPNN) (Ahmadlou & Adeli, 2010; George et al., 2017; George, Rafiei, Borstad, Adeli, & Gauthier, 2017) to classify the features extracted from DBM in order to determine which classification algorithm produces the highest classification accuracies (Figure 26). It consists of four layers. Layer one is the input layer and consists of one neuron for every input. In this case, that number is equal to the number of neurons in the last layer of the DBM (25 in this research). Layer 2 is called the pattern layer with a number of neurons equal to the amount of training samples.

The third layer is the summation layer and has as many neurons as there are classes, in

106 this research 2. The last layer is the output layer and consists of a single decision neuron that assigns the classification to the inputs. The advantage of EPNN is that it takes into account the local heterogeneity of the training data through the introduction of the local circles and can use that information to improve the classification accuracy (Hirschauer,

Adeli, & Buford, 2015).

Several parameters of the DBM were tested in order to optimize the results, including the number of features extracted, learning rate, pre-training iterations, fine- tuning iterations, and the size of batches. These parameters are highly problem dependent. In this study, the parameters resulting in the highest classification accuracies were a 10% testing-to-training ratio and a hidden layer with 25 neurons (Figure 25).

Other DBM architectures with a larger number of layers were also tested, but they yielded the same classification accuracy at a higher computational cost.

RESULTS

In this case study, the participant attended 10 sessions of simultaneous gameplay and EEG recording. A total of 7,208 gestures were parsed from the kinematic skeleton data. They included 724 bottle gestures, 546 parachute gestures, 1,329 navigate left gestures, and 1,952 rowing gestures.

After pre-processing, a total of 1481 epochs (each associated with a single gesture) were extracted for analysis. Using the normality index described above, these gestures were comprised of 907 “good” quality gestures and 574 “poor” quality pre- movement epochs.

107

After transformation to the frequency domain using FFT and initial dimensionality reduction of the frequency amplitudes, the results of testing accuracies are based on 1481 samples each containing 250*21 = 5,250 data points for feature extraction.

After training and testing on 50 random permutations of the data set, the highest average testing classification accuracy achieved was 70.39% with a combination of

DBM for feature extraction and EPNN for classification of the features. The next highest average classification accuracy was 70.00% obtained from PNN on features extracted by the DBM. A combination of DBM with the softmax layer achieved the lowest average accuracy of 67.95%. A validation test was performed to determine if the assigned classes were really capturing meaningful patterns within the data. To perform this test, the class labels for each sample were randomly assigned so that the original classes corresponding to “good” vs. “poor’ epochs were no longer matched to the data. These data were used as input into the DBM and the classification accuracies were calculated on

50 random permutation of this newly labeled data. As expected, the classification accuracies were much lower, performing right around chance level. With these randomly assigned classes, average accuracies were 53.32% with EPNN, 51.03% with PNN, and

49.35% with DBM.

DISCUSSION

These results demonstrate the feasibility of extracting and classifying features of a neural signal associated, specifically, with the quality of movement encoded in the pre- movement signal. Classifications accuracies shown here were produced from a combination of a Deep Boltzmann Machine for feature extraction and an Enhanced

108

Probabilistic Neural Network for classification are on par with previous neural decoding studies related to movement (Bradberry et al., 2010; Chen et al., 2018; Ofner & Müller-

Putz, 2012; Ortiz-Rosario & Adeli, 2013; Rafiei & Adeli, 2018; Taylor & Schmidt,

2012b). While some motor imagery studies have seen higher classification accuracies for neural decoding (Taylor & Schmidt, 2012a) , the most obvious difference between those studies and the one performed here is that they, by definition, do not involve any physical movement. Therefore, they are notoriously easier to capture with EEG as there are no associated movement artifacts. However, the wavelet filtering utilized in this study was able to successfully remove a vast majority of the movement artifacts that appeared as large baseline shift. What was less easy to cleanly remove with either the wavelet filtering or the bandpass filtering were the high frequency EMG artifacts. The participant in this study presented with high amounts of tone, or involuntary muscle contractions, during movements. This EMG activity was also captured by the EEG system thereby producing a lot of artifact in the data. This can be seen as high frequency activity just before each movement (red line) in figure 24. While wavelet filtering was able to remove much of this high frequency artifact from the data, a balance had to be made between removing the artifact but also preserving any real high frequency components of the EEG signal. For this reason, it could be that not all EMG artifacts were clearly removed from the data which may have contributed to lower classification accuracies. The lack of EMG artifact in healthy participants from previous studies may have been a contributor to their classification success.

Another potential contributor to the success of previous studies, are the easily defined classes associated with movement decoding. When trying to classify neural

109 signals associated with certain movements, each movement type represents a single, discrete class. On the other hand, in this study, we investigated the quality of a person’s movement. Quality of movement inherently exists on a continuous scale. Therefore, defining an optimal division between class 1, “good”, and class 2, “poor”, is a challenge.

Future studies of this kind may choose to define the discrete classes by a different set of criteria or utilize a different classification method that allows for continuous classification, such as convolutional neural networks (Acharya, Oh, Hagiwara, Tan, &

Adeli, 2017; Acharya et al., 2018; Krizhevsky, Sutskever, & Hinton, 2012). This consideration of reclassification may contribute to higher classification accuracies in future studies.

However, unlike previous studies, the results presented here do reflect the feasibility of decoding this specific characteristic related to movement quality instead of seeking to decode movements themselves. The utility of this type of decoding is revealed in applications of real-time feedback to enhance or improve the quality of a person’s movement. Ultimately, these results may have resounding implications for real-time neurorehabilitation and self-regulated neuromodulation during motor restorative training after stroke, and indeed, any sort of movement quality training (e.g..: sports, military).

These fields of neuromodulation and neurofeedback are rapidly growing and are piquing the interest of the scientific community. A simple Google Scholar search of

“neurofeedback” and “neuromodulation” will return over 9,000 results from 2018 alone.

It should come as no surprise, then, that the success of neurofeedback and neuromodulation in both healthy persons (Gruzelier, 2014; Jurewicz et al., 2018) as well as those with a myriad of neurological disorders has been demonstrated (Thibault, Robert

110

T., Lifshitz, Michael,Raz, Amir, 2016). For example, Jurewicz et al. (2018) demonstrated that healthy persons, utilizing EEG neurofeedback, could train their brain to up-regulate both alpha (8-12 Hz) and partial beta (12-22 Hz) activities. Similarly, neurofeedback training has been shown to improve emotional regulation, (Zotev, Phillips, Yuan, Misaki,

& Bodurka, 2014), attention deficit hyperactivity disorder (ADHD) symptoms

(Micoulaud-Franchi et al., 2014; Van Doren et al., 2018), post-traumatic stress disorder

(PTSD) (Zotev et al., 2018), and headache (Moshkani Farahani, Tavallaie, Ahmadi, &

Fathi Ashtiani, 2014), thereby demonstrating the feasibility and utility of neurofeedback.

The results presented in this study show that unique features of a person’s brain signal that characterize their upcoming movement performance can be extracted and classified. In line with the fast growing fields of neuromodulation and neurofeedback, future studies should consider utilizing these features for feedback to improve outcomes of motor restorative rehabilitation. For example, during concurrent collection of continuous EEG data with Recovery Rapids game play, the brain signal can be monitored to detect these characterizing features. Once features that are indicative of an upcoming

“good” quality or “poor quality” movement are identified within the brain signal, feedback can be provided to alert the participant of the particular brain state they are in.

This feedback can take many forms. One possibility is utilizing a simple visual display with a stoplight red-to-green color bar indicating neural activity associated with “poor” quality versus “good” quality movements. The goal would be to produce or maintain the

“green” color bar associated with better motor planning during motor training.

111

STUDY LIMITATIONS

Although this study represents a first step toward this type of real-time neurofeedback, there are several design parameters that may be further optimized. Most obviously, this was a case study. Given that neurofeedback would always need to be individualized, a case by case analysis is relevant. Nonetheless, future work should incorporate a larger number of individuals with a variety of stroke presentations to demonstrate feasibility in a larger population. Because the participant in this study had a basilar artery infarction, his overall cortical function was largely intact, making this a best-base scenario. It is expected that the neural features for each participant may be different, and extensive cortical damage could provide different conditions for this approach. Therefore, generalizability of the computational model needs to be tested with a variety of cases to ensure high accuracy for several participants.

Another area of optimization should be consideration of the feature space.

Because there are nearly an infinite number of features that can be extracted from the

EEG signal, the choice of feature space is an important one for uncovering the most uniquely characterizing features. In this study, the frequency spectrum spanning the traditional 5 EEG bands (Delta (0-4 Hz), Theta (4-7 Hz), Alpha (8-12Hz), Beta (12-30

Hz), and Gamma (30-100+Hz) was used for feature extraction. Similar approaches have been used in other deep learning algorithms for successful feature extraction from EEG during motor imagery (Chen et al., 2018). However, other feature spaces could be investigated that range from those associated with specific electrodes or EEG bands to features that are extracted from purely mathematical constructs.

112

For example, features extracted by individual EEG electrodes have the potential to reveal specific channels that contain the most differential information about movement quality. This approach may then lend itself to subsequent source localization studies using other techniques such as low resolution brain electromagnetic tomography

(LORETA) (Bradberry et al., 2010). Source localization techniques as well as coherence methods (Sankari & Adeli, 2011) would not only provide a new feature space in which to explore, but could also lead to deeper neurophysiological understanding of the underlying neural processes as has been demonstrated in other applications. (Lin et al., 2017; Sotero

& Trujillo-Barreto, 2008).

On the other hand, features extracted from pure mathematical transforms of the

EEG data might also provide utility in increasing classification accuracy. For example, creating a feature space consisting of principal components extracted from principal component analysis (PCA) is one avenue for exploration (Ortiz-Rosario et al., 2015).

Another may be utilizing wavelet analysis not only for filtering, but feature extraction as well. For example, wavelet coefficients associated with specific levels of detail (i.e.: associated with certain frequency bands) could be utilized as a feature space in their own right (Li et al., 2017; Perez-Ramirez et al., 2016; Qarib & Adeli, 2015). If a high number of coefficients are computed, dimensionality reduction can be performed using DBM.

Otherwise, these coefficients may be fed directly into EPNN for classification. In this regard, using Wavelet Packet Transform (WPT) would be superior to the standard discrete wavelet transform (Jiang & Adeli, 2004). In WPT, both approximations and details are decomposed at each level which provides a more powerful time-frequency analysis capability (Jiang, Mahadevan, & Adeli, 2007).

113

With a larger population size and exploration into other feature spaces of the signal, higher classification accuracies may be reached. With these high accuracies, features that are present within the pre-movement neural signal encoding an upcoming

“good” quality movement, or “poor” quality movement, can be used within a clinical neurofeedback setting. Specifically, by looking for these features within the neural signal during game-play or other motor restorative treatments, real-time neurofeedback can be disseminated based on the appearance of a “good” signal pattern or a “poor” signal pattern with the aim that the individual undergoing treatment may learn to produce those brain signal patterns associated with best possible movement.

CONCLUSION

This study demonstrated the feasibility of extracting features of a neural signal associated with predictive movement quality. These results are achieved using a combination of advanced machine learning techniques, including Deep Boltzmann

Machine for feature extraction and Enhanced Probabilistic Neural Networks for classification. While the classification accuracies achieved in this study were on par with other neural decoding studies, there is room for improvement. One of the major sources of improvement is likely to come from exploring different feature spaces. Ultimately, these techniques should be used to implement real-time neurofeedback for motor performance.

114

Row or Parachutes Navigate Bottles Figure 18: Screen shot of Recovery Rapids

The avatar the participant controls is shown on the big middle screen. The motion captured skeleton that captures the participant’s movement, is shown on the smaller screen in the upper right. The accompanying115 , pre-defined movements for navigating down the river are shown in the picture below.

Recovery

Rapids EEG Signal

Kinematic Data Capture

Velocity/ Joint Range of Angular Angles Motion Velocity

Normalcy Preprocessing Score

Machine Learning: Feature Extraction

Classification of movement quality Figure 19: Flowchart of methodologies (Class 1 or 2)

116

Figure 20: Skeleton coordinates captured by Kinect

117

Figure 21: Gesture parsing for a rowing gesture

The red curve indicates the gesture itself. The blue line is skeleton data captured before and after the movement to give context. The small circle indicates the time the gesture was trigged in the game. The gray rectangle indicates the 0.5 second epoch before movement onset.

118

AFp3h AFpz AFp4h AFF5h AF6h

AFF5 AFF6

AFF5h AFF3 AFF1 AFFz AFF2 AFF4 AFF6h

FFC5h FFC3 FFC3h FFC1h FFCz FFC2h FFC4h FFC4 FFC6h

FCC5h FCC3 FCC1 FCC1h FCCz FCC2h FCC2 FCC4 FCC6h

CCP5h CCP3 CCP1 CCP1h CCPz CCP2h CCP2 CCP4 CCP6h

CCP5h CPP3 CPP3h CPP1h CPPz CPP2h CPP4h CPP4 CPP6h

PO5 PO3 PO1 POz PO2 PO4 PO6 PO7 PO8 POO7 POO8

O1h Oz O2h

Figure 22: Map of 64 EEG electrodes. Highlighted electrodes were used for analysis in this research

119

Figure 23: Normalcy Residuals for the row, bottle, parachute, and raisearm gestures The red curve indicates the expected performance for the individual. The residuals indicated the actual performance, as compared to "normal”.

120

A B C

D E F

Figure 24: Illustration of EEG processing. In panels A-E, the x axis represents time (in seconds) and the Y-axis represents the voltage fluctuations in each channel. Red lines labeled 256 correspond to the beginning of a movement recognized by the Recovery Rapids game. All EEG panels are shown on a 60µV scale. A: 64 channels of raw EEG data. B: 21 Remaining channels after channel removal by kurtosis and visual inspection. C: All channels re-referenced to the average reference. D: Result of wavelet filtering. E: The data are split into 0.5 second epochs. F: Each channel in each epoch are transformed into the frequency domain using the Fast Fourier Transform. The amplitude values (P1) on the y-axis for frequencies up to 250Hz (x-axis), are used as input into the DBM. For each epoch, there are 250(Hz) * 21(Channels) = 5,250 inputs.

121

Soft Max Layer for Features Classification

Inputs OR (P1 Values) 5,250 25 25 5,250

EPNN or PNN for Classification Features

Encoder Decoder Figure 25: Two-layer deep Boltzmann machine for feature extraction consisting of an encoder and decoder The numbers in the hidden layer (dashed lines) correspond to the number of neurons present in that layer. Training takes place as the encoder breaks down the input into a reduced dimensionality feature space and the decoder rebuilds the input from those features. Once the network stabilizes (learns the best features for reconstruction) or the maximum number of learning iterations is reach (400 was used in this research), the features are extracted and used for classification either within a SoftMax layer of the DBM itself, or a different classification network, in this research EPNN or PNN.

122

Figure 26: Architecture of PNN/EPNN. It consists of four layers. Layer one is the input layer and consists of one neuron for every input. In this case, that number is equal to the number of neurons in the last layer of the DBM (25 in this research). Layer 2 is called the pattern layer with a number of neurons equal to the number of training samples. The third layer is the summation layer and has as many neurons as there are classes, in this research 2. The last layer is the output layer and consists of a single decision neuron that assigns the classification to the inputs.

123

CHAPTER 7: SYNTHESIS AND RECOMMENDATIONS

Taken together, the results of this dissertation represent steps toward more robust multi- system analysis of neuro motor control with implications for advances in motor restorative therapies. In Chapter 3, stimulus triggered averaging of EMG signals was used to demonstrate interactions between the motor cortex (CX) and the pontomedullary reticular formation (PMRF) during a reaching task performed by non-human primates. By revealing specific types of interactions between brain regions, including summation, gating, and other complex interactions, this multi-system analysis contributes to closing the gap of knowledge that exists between our understanding of motor control and neurophysiology. Future neurophysiological studies of this kind can build on this work by developing a more robust classification of interaction types as represented in the EMG signal patterns. For example, apart from simple summation and gating, the other major interactions type identified in Chapter 2 was the “complex” interaction pattern.

With the current surge in successful machine learning methodologies for a plethora of applications (Hirschauer, Adeli, & Buford, 2015b; Qarib & Adeli, 2015; Rafiei & Adeli, 2015;

Rafiei & Adeli, 2017; Rafiei, Khushefati, Demirboga, & Adeli, 2017b), future works could consider utilizing machine learning techniques such as those used in the later chapter of the dissertation, including feature extraction and classification, to better quantify this type of signal pattern and potentially elucidate the underlying neurophysiology.

Given these potential directions to investigate in this line of research, the steps toward more complete understanding of multi-system interactions presented here also have the potential to better inform motor rehabilitation efforts in the future. For instance, with the knowledge that there are multiple regions of the brain that participate in muscle recruitment in the upper limbs, 124 more options for motor-restorative therapy might be explored after someone suffers damage to their motor cortex, as is often seen in stroke.

In Chapters 4, machine learning was used to determine if both pre-therapy somatosensory and motor ability were key factors in predicting response to constraint-induced movement therapy. The results presented here indicate that baseline somatosenation, as measured using the

Brief Kinesthesia Test and Touch Test Monofilaments, does not contribute to the overall predication accuracy of CI therapy response. With baseline motor ability considered as a predictor however, accuracies for predicting whether someone would be a “Best”, “Moderate”, or “Non” responder to the treatment reached 100%.

In Chapter 5 motor ability was broken down by specific task types, as defined by the

Wolf Motor Function Test. The results show that combinations of tasks performed using gross movements, as opposed to fine movements, give rise to the highest prediction accuracies of treatment response (up to 100%), regardless of therapy modality (gaming motor restorative therapy or traditional CI therapy). It was also shown that those with poorer gross motor ability at baseline tend to see the greatest response to treatment.

These results have implications for prognostic decision making. For example, if a clinician is trying to assess whether a person is a good candidate for motor restorative CI therapy versus, say, compensatory therapy, they could a) consider the person’s baseline gross motor ability as a factor in their decision making and b) assess that ability using only a small subset of the traditional tasks performed on the WMFT.

Because the results presented in Chapters 4 and 5 were based on a relatively small sample size, an obvious next step for this work will be to test it on more participants. One challenge for

125 machine learning is that, with small datasets, it may perform well on that dataset, but not generalize well to new data. Therefore, to achieve the most robust (most generalizable) models, a larger sample size of participants should be added.

In Chapter 6, the presented case study successfully demonstrated the feasibility of extracting features that are indicative of the quality of movement a person is about to perform from neural signals. Once again, machine learning was employed to extract (Deep Boltzmann

Machine (DBM)) and classify (Enhanced Probabilistic Neural Network (EPNN)) features with an accuracy of 70.39%. For future studies of this kind, that seek to elucidate characterizing features of the neural signal, an area where there is plenty of room for exploration is within the selected feature space for each study. In the presented case study, features were extracted from the amplitude values of the frequency spectrum, which is consistent with other studies (Tabar &

Halici, 2016b). However, features can be extracted from any number of domains associated with the signal, from the time domain to specific frequency bands within the frequency spectrum, to individual or subsets of channels that may show high coherence or source localization values.

Solving the so called “EEG Inverse Problem”, or determining source localizations associated with the features that are indicative of movement quality is another promising area for investigation. Through utilization of standardize low resolution brain electromagnetic tomography (sLORETA) or other source localization methodologies, it might be possible to uncover the underlying neurophysiology and specific neural activity associated with these features.

Ultimately, by achieving high accuracies of feature classification found in neural signatures associated with movement quality, another future pathway is into the clinic. Features

126 the can be recognized and classified as being associated with a certain type of brain pattern that is indicative of an upcoming “good” quality movement or “poor” quality movement can be exploited for use in neurorehabilitation and neurofeedback. For example, imagine a person is undergoing the game-based motor restorative therapy discussed in chapters 4-6, during simultaneous EEG monitoring. It is feasible that with the appropriate automated set-up for comparing the current features present in pre-movement neural signals to the pre-determined

“good” features and “poor” features for that individual that matches to either type of feature class could be used to give neurofeedback. For example, if, just before a person is about to make a rowing gesture, “good” features show up in the EEG signal, that information can be relayed to the person to provide encouragement and feedback that they are in the optimal brain state.

Different feedback would be given in the presence of “poor” features. This feedback could take many forms, but even something as simple as a color bar display with a red bar indicating the presence of “poor” features and a green bar indicating the presence of “good” features could be a useful form of neurofeedback. In this way, a person may learn to alter their brain state to reflect patterns associated with their best movement.

Each of these studies and the results produced therein represent the utility and importance of multi-system analysis. Moreover, from the study of electrical signals produced by the muscles and the brain, to the movement kinematics of real-time motor-restorative game play, the natural lens through which to conduct these experiments was biophysics. With these multi-system, biophysical approaches, gaps in our understanding between how our arms move and how our brain produces those movements has been lessened; prediction of motor-restorative treatment response has expanded the pool of supplemental information that is available in order to

127 determine the most appropriate type of therapy for a person with upper-limb motor deficits; and another step toward real-time, personalized neurofeedback for motor restoration has been taken with the successful identification and classification of uniquely characterizing neural features associated with movement quality.

128

References

Adeli, Hojjat,,Hung, Shih-Lin,,. (1995). Machine learning : Neural networks, genetic algorithms,

and fuzzy systems. New York: Wiley.

Ahmadlou, M., & Adeli, H. (2010a). Enhanced probabilistic neural network with local decision

circles: A robust classifier. Integrated Computer-Aided Engineering, 17(3), 197-210.

Ahmadlou, M., & Adeli, H. (2010b). Enhanced probabilistic neural network with local decision

circles: A robust classifier. Integrated Computer-Aided Engineering, 17(3), 197-210.

Anderson, K. R., Woodbury, M. L., Phillips, K., & Gauthier, L. V. (2015). Virtual reality video

games to promote movement recovery in stroke rehabilitation: A guide for clinicians.

Archives of Physical Medicine and Rehabilitation, 96(5), 973-976.

doi:10.1016/j.apmr.2014.09.008 [doi]

Antoniades, A., Spyrou, L., Martin-Lopez, D., Valentin, A., Alarcon, G., Sanei, S., & Took, C.

C. (2018). Deep neural architectures for mapping scalp to intracranial EEG. International

Journal of Neural Systems, , 1850009.

Asanuma, H. (1973). Cerebral cortical control of movement. The Physiologist, 16(2), 143-166.

Attema, T., Kosgodagan Acharige, A., Morales-Nápoles, O., & Maljaars, J. (2017). Maintenance

decision model for steel bridges: A case in the netherlands. Structure and Infrastructure

Engineering, 13(2), 242-253.

129

Baker, S. N., Zaaimi, B., Fisher, K. M., Edgley, S. A., & Soteropoulos, D. S. (2015). Pathways

mediating functional recovery. Progress in Brain Research, 218, 389-412.

Bhagat, N. A., French, J., Venkatakrishnan, A., Yozbatiran, N., Francisco, G. E., O'Malley, M.

K., & Contreras-Vidal, J. L. (2014). (2014). Detecting movement intent from scalp EEG in a

novel upper limb robotic rehabilitation system for stroke. Paper presented at the

Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual International

Conference of the IEEE, 4127-4130.

Bird, M., Cannell, J., Jovic, E., Rathjen, A., Lane, K., Tyson, A., . . . Smith, S. (2017). A

randomized controlled trial investigating the efficacy of virtual reality in inpatient stroke

rehabilitation. Archives of Physical Medicine and Rehabilitation, 98(10), e27.

Borstad, A. L., Crawfis, R., Phillips, K., Pax Lowes, L., Maung, D., McPherson, R., . . .

Gauthier, L. V. (2018). In-home delivery of constraint-induced movement therapy via

virtual reality gaming. Journal of Patient-Centered Research and Reviews, 5(1), 6-17.

Borstad, A., & Nichols-Larsen, D. S. (2016a). The brief kinesthesia test is feasible and sensitive:

A study in stroke. Brazilian Journal of Physical Therapy, 20(1), 81-86.

Borstad, A., & Nichols-Larsen, D. S. (2016b). The brief kinesthesia test is feasible and sensitive:

A study in stroke. Brazilian Journal of Physical Therapy, 20(1), 81-86.

Botella, C., Serrano, B., Banos, R. M., & Garcia-Palacios, A. (2015). Virtual reality exposure-

based therapy for the treatment of post-traumatic stress disorder: A review of its efficacy,

130

the adequacy of the treatment protocol, and its acceptability. Neuropsychiatric Disease and

Treatment, 11, 2533-2545. doi:10.2147/NDT.S89542 [doi]

Bradberry, T. J., Gentili, R. J., & Contreras-Vidal, J. L. (2010). Reconstructing three-

dimensional hand movements from noninvasive electroencephalographic signals. The

Journal of Neuroscience : The Official Journal of the Society for Neuroscience, 30(9), 3432-

3437. doi:10.1523/JNEUROSCI.6107-09.2010 [doi]

Brinkman C, P. R. (1979). Supplementary motor area in the monkey: Activity of neurons during

performance of a learned motor task. Journal of Neurophysiology, 42(3), 681-709.

Burns, A., Adeli, H., & Buford, J. A. (2014). Brain–computer interface after nervous system

injury. The Neuroscientist, 20(6), 639-651.

Callahan, A. D., Hunter, J. M., & Mackin, E. (1995). Rehabilitation of the hand: Surgery and

therapy Mosby.

Canedo, A., & Lamas, J. A. (1993). Pyramidal and corticospinal synaptic effects over

reticulospinal neurones in the cat. The Journal of Physiology, 463, 475-489.

Center for disease control and prevention. (2017).

Centers For Disease Control and Prevention. (2015). CDC injury center research priorities.

Retrieved from https://www.cdc.gov/injury/pdfs/researchpriorities/cdc-injury-research-

priorities.

131

Daubechies, I. (1990). The wavelet transform, time-frequency localization and signal analysis.

IEEE Transactions on Information Theory, 36(5), 961-1005.

Davidson, A. G., & Buford, J. A. (2006). Bilateral actions of the reticulospinal tract on arm and

shoulder muscles in the monkey: Stimulus triggered averaging. Experimental Brain

Research, 173(1), 25-39.

Davidson, A. G., & Buford, J. A. (2004). Motor outputs from the primate reticular formation to

shoulder muscles as revealed by stimulus-triggered averaging. Journal of Neurophysiology,

92(1), 83-95. doi:10.1152/jn.00083.2003 [doi]

Debener, S., Minow, F., Emkes, R., Gandras, K., & Vos, M. (2012). How about taking a low‐

cost, small, and wireless EEG for a walk? Psychophysiology, 49(11), 1617-1621.

Delorme A., M. S. (2004). EEGLAB: An open source toolbox for analysis of single-trial EEG

dynamics

. Journal of Neuroscience Methods, (134), 9.

Drew, T. (1991). Functional organization within the medullary reticular formation of the intact

unanesthetized cat. III. microstimulation during locomotion. Journal of Neurophysiology,

66(3), 919-938.

132

Drew, T., & Rossignol, S. (1984a). Phase-dependent responses evoked in limb muscles by

stimulation of medullary reticular formation during locomotion in thalamic cats. Journal of

Neurophysiology, 52(4), 653-675.

Drew, T., & Rossignol, S. (1984b). Phase-dependent responses evoked in limb muscles by

stimulation of medullary reticular formation during locomotion in thalamic cats. Journal of

Neurophysiology, 52(4), 653-675.

Drew, T., & Rossignol, S. (1990). Functional organization within the medullary reticular

formation of intact unanesthetized cat. II. electromyographic activity evoked by

microstimulation. Journal of Neurophysiology, 64(3), 782-795.

EEGLAB. (2011). Chapter 01: Rejecting artifacts. Retrieved from

https://sccn.ucsd.edu/wiki/Chapter_01:_Rejecting_Artifacts

Fisher, K. M., Zaaimi, B., & Baker, S. N. (2012). Reticular formation responses to magnetic

brain stimulation of primary motor cortex. The Journal of Physiology, 590(16), 4045-4060.

Fregosi, M., & Rouiller, E. M. (2017). Ipsilateral corticotectal projections from the primary,

premotor and supplementary motor cortical areas in adult macaque monkeys: A quantitative

anterograde tracing study. European Journal of Neuroscience, 46(8), 2406-2415.

Fritz, S. L., Blanton, S., Uswatte, G., Taub, E., & Wolf, S. L. (2009). Minimal detectable change

scores for the wolf motor function test. Neurorehabilitation and Neural Repair, 23(7), 662-

667.

133

Frolov, A. A., Mokienko, O., Lyukmanov, R., Biryukova, E., Kotov, S., Turbina, L., . . .

Bushkova, Y. (2017). Post-stroke rehabilitation training with a motor-imagery-based brain-

computer interface (BCI)-controlled hand exoskeleton: A randomized controlled multicenter

trial. Frontiers in Neuroscience, 11, 400.

Gabran, S., Salam, M. T., Dian, J., El-Hayek, Y., Velazquez, J. P., Genov, R., . . . Mansour, R.

R. (2013). High-density intracortical microelectrode arrays with multiple metallization

layers for fine-resolution neuromonitoring and neurostimulation. IEEE Transactions on

Neural Systems and Rehabilitation Engineering, 21(6), 869-879.

Gauthier, L. (2018). (2018). Remote monitoring of motor recovery using game-based motion

capture. Paper presented at the Brain Health and Performance Summit, Columbus, OH.

Gauthier, L. V., Kane, C., Borstad, A., Strahl, N., Uswatte, G., Taub, E., . . . Mark, V. (2017a).

Video game rehabilitation for outpatient stroke (VIGoROUS): Protocol for a multi-center

comparative effectiveness trial of in-home gamified constraint-induced movement therapy

for rehabilitation of chronic upper extremity hemiparesis. BMC Neurology, 17(1), 109.

Gauthier, L. V., Kane, C., Borstad, A., Strahl, N., Uswatte, G., Taub, E., . . . Mark, V. (2017b).

Video game rehabilitation for outpatient stroke (VIGoROUS): Protocol for a multi-center

comparative effectiveness trial of in-home gamified constraint-induced movement therapy

for rehabilitation of chronic upper extremity hemiparesis. BMC Neurology, 17(1), 109.

134

George, S. H., Rafiei, M. H., Borstad, A., Adeli, H., & Gauthier, L. V. (2017). Gross motor

ability predicts response to upper extremity rehabilitation in chronic stroke. Behavioural

Brain Research, 333, 314-322.

George, S. H., Rafiei, M. H., Gauthier, L., Borstad, A., Buford, J. A., & Adeli, H. (2017a).

Computer-aided prediction of extent of motor recovery following constraint-induced

movement therapy in chronic stroke. Behavioural Brain Research, 329, 191-199.

George, S. H., Rafiei, M. H., Gauthier, L., Borstad, A., Buford, J. A., & Adeli, H. (2017b).

Computer-aided prediction of extent of motor recovery following constraint-induced

movement therapy in chronic stroke. Behavioural Brain Research, 329, 191-199.

Georgopoulos, A. P., Kalaska, J. F., Caminiti, R., & Massey, J. T. (1982). On the relations

between the direction of two-dimensional arm movements and cell discharge in primate

motor cortex. The Journal of Neuroscience : The Official Journal of the Society for

Neuroscience, 2(11), 1527-1537.

Govender, M., Bowen, R. C., German, M. L., Bulaj, G., & Bruggers, C. S. (2015). Clinical and

neurobiological perspectives of empowering pediatric cancer patients using videogames.

Games for Health Journal, 4(5), 362-374.

Gromala, D., Tong, X., Choo, A., Karamnejad, M., & Shaw, C. D. (2015). (2015). The virtual

meditative walk: Virtual reality therapy for chronic pain management. Paper presented at the

Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing

Systems, 521-524. 135

He, X., & Wu, C. (1985). Connections between pericruciate cortex and the medullary

reticulospinal neurons in cat: An electrophysiological study. Experimental Brain Research,

61(1), 109-116.

Herbert, W. J., Powell, K., & Buford, J. A. (2015). Evidence for a role of the reticulospinal

system in recovery of skilled reaching after cortical stroke: Initial results from a model of

ischemic cortical injury. Experimental Brain Research, 233(11), 3231-3251.

Hinton, G. E., & Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural

networks. Science (New York, N.Y.), 313(5786), 504-507. doi:313/5786/504 [pii]

Hirschauer, T. J., Adeli, H., & Buford, J. A. (2015a). Computer-aided diagnosis of parkinson’s

disease using enhanced probabilistic neural network. Journal of Medical Systems, 39(11),

179.

Hirschauer, T. J., Adeli, H., & Buford, J. A. (2015b). Computer-aided diagnosis of parkinson’s

disease using enhanced probabilistic neural network. Journal of Medical Systems, 39(11),

179.

Hirschauer, T. J., & Buford, J. A. (2015). Bilateral force transients in the upper limbs evoked by

single-pulse microstimulation in the pontomedullary reticular formation. Journal of

Neurophysiology, 113(7), 2592-2604. doi:10.1152/jn.00852.2014 [doi]

Hoffman, D. S., & Strick, P. L. (1995). Effects of a primary motor cortex lesion on step-tracking

movements of the wrist. Journal of Neurophysiology, 73(2), 891-895.

136

Hulbert, S. Adeli, H., Buford, J. (2015). (2015). Interactions between corticospinal and

reticulospinal outputs determine muscle response in the upper limbs and trunk as revealed

with stimulus triggered averaging. Paper presented at the Neuroscience Meeting Planner,

Society for Neuroscience, Chicago, Illinois.

Hulbert, S., & Adeli, H. (2015). Spotting psychopaths using technology. Reviews in the

Neurosciences, 26(6), 721-732.

Ibáñez, Jaime, Serrano, J. I., del Castillo, M. D., Minguez, J., Pons,J.L.,. (2015). Predictive

classification of self-paced upper-limb analytical movements with EEG. Med Biol Eng

Comput Medical & Biological Engineering & Computing, 53(11), 1201-1210.

Ibáñez, J., Monge-Pereira, E., Molina-Rueda, F., Serrano, J., del Castillo, M. D., Cuesta-Gómez,

A., . . . Miangolarra-Page, J. C. (2017). Low latency estimation of motor intentions to assist

reaching movements along multiple sessions in chronic stroke patients: A feasibility study.

Frontiers in Neuroscience, 11, 126.

Jankowska, E., & Edgley, S. A. (2006). How can corticospinal tract neurons contribute to

ipsilateral movements? A question with implications for recovery of motor functions. The

Neuroscientist : A Review Journal Bringing Neurobiology, Neurology and Psychiatry,

12(1), 67-79. doi:12/1/67 [pii]

Jurewicz, K., Paluch, K., Kublik, E., Rogala, J., Mikicin, M., & Wróbel, A. (2018). EEG-

neurofeedback training of beta band (12–22Hz) affects alpha and beta frequencies – A

137

controlled study of a healthy population

doi:https://doi.org/10.1016/j.neuropsychologia.2017.11.021

Kasser, R. J., & Cheney, P. D. (1985). Characteristics of corticomotoneuronal postspike

facilitation and reciprocal suppression of EMG activity in the monkey. Journal of

Neurophysiology, 53(4), 959-978.

Keizer, K., Kuypers,H.G.J.M.,. (1989). Distribution of corticospinal neurons with collaterals to

the lower brain stem reticular formation in monkey (macaca fascicularis). Exp Brain Res

Experimental Brain Research, 74(2), 311-318.

Kermadi, I., CALCIATI, T., & ROUILLER, E. (1998). Neuronal activity in the primate

supplementary motor area and the primary motor cortex in relation to spatio-temporal

bimanual coordination. Somatosensory & Motor Research, 15(4), 287-308.

Kobayashi, M., Hutchinson, S., Schlaug, G., & Pascual-Leone, A. (2003). Ipsilateral motor

cortex activation on functional magnetic resonance imaging during unilateral hand

movements is related to interhemispheric interactions. NeuroImage, 20(4), 2259-2270.

Koziarski, M., & Cyganek, B. (2017). Image recognition with deep neural networks in presence

of noise–dealing with and taking advantage of distortions. Integrated Computer-Aided

Engineering, 24(4), 337-349.

138

Kranczioch, Cornelia, Zich, Catharina, Schierholz, Irina,Sterr, Annette,. (2014). Mobile EEG and

its potential to promote the theory and application of imagery-based motor rehabilitation.

INTPSY International Journal of Psychophysiology, 91(1), 10-15.

Kranczioch, C., Zich, C., Schierholz, I., & Sterr, A. (2014). Mobile EEG and its potential to

promote the theory and application of imagery-based motor rehabilitation. International

Journal of Psychophysiology, 91(1), 10-15.

Lacroix, S., Havton, L. A., McKay, H., Yang, H., Brant, A., Roberts, J., & Tuszynski, M. H.

(2004). Bilateral corticospinal projections arise from each motor cortex in the macaque

monkey: A quantitative study. Journal of Comparative Neurology, 473(2), 147-161.

Lang, C. E., Strube, M. J., Bland, M. D., Waddell, K. J., Cherry‐Allen, K. M., Nudo, R. J., . . .

Birkenmeier, R. L. (2016). Dose response of task‐specific upper limb training in people at

least 6 months poststroke: A phase II, single‐blind, randomized, controlled trial. Annals of

Neurology, 80(3), 342-354.

Lang, C. E., & Schieber, M. H. (2004). Reduced muscle selectivity during individuated finger

movements in humans after damage to the motor cortex or corticospinal tract. Journal of

Neurophysiology, 91(4), 1722-1733. doi:10.1152/jn.00805.2003 [doi]

Laver, K. E., Lange, B., George, S., Deutsch, J. E., Saposnik, G., & Crotty, M. (2018). Virtual

reality for stroke rehabilitation. Stroke, 49(4), e160-e161.

139

Lawrence, D. G., & Kuypers, H. G. (1968). The functional organization of the motor system in

the monkey 1: I. the effects of bilateral pyramidal lesions. Brain, 91(1), 1-14.

Lee, Y., Hsieh, Y., Wu, C., Lin, K., & Chen, C. (2015). Proximal fugl-meyer assessment scores

predict clinically important upper limb improvement after 3 stroke rehabilitative

interventions. Archives of Physical Medicine and Rehabilitation, 96(12), 2137-2144.

Levin, M. F., Weiss, P. L., & Keshner, E. A. (2015). Emergence of virtual reality as a tool for

upper limb rehabilitation: Incorporation of motor control and motor learning principles.

Physical Therapy, 95(3), 415-425.

Lin, K., Huang, Y., Hsieh, Y., & Wu, C. (2009). Potential predictors of motor and functional

outcomes after distributed constraint-induced therapy for patients with stroke.

Neurorehabilitation and Neural Repair, 23(4), 336-342.

Lin, Y., Nie, Z., & Ma, H. (2017). Structural damage detection with automatic feature‐extraction

through deep learning. Computer‐Aided Civil and Infrastructure Engineering, 32(12), 1025-

1046.

Liu, D., Chen, W., Lee, K., Chavarriaga, R., Iwane, F., Bouri, M., . . . Millán, J. d. R. (2018).

EEG-based lower-limb movement onset decoding: Continuous classification and

asynchronous detection. IEEE Transactions on Neural Systems and Rehabilitation

Engineering,

140

Llorens, R., Noe, E., Colomer, C., & Alcaniz, M. (2015). Effectiveness, usability, and cost-

benefit of a virtual reality-based telerehabilitation program for balance recovery after stroke:

A randomized controlled trial. Archives of Physical Medicine and Rehabilitation, 96(3),

418-425.e2. doi:10.1016/j.apmr.2014.10.019 [doi]

Luccarini, P., Gahery, Y., & Pompeiano, O. (1990a). Cholinoceptive pontine reticular structures

modify the postural adjustments during the limb movements induced by cortical stimulation.

Archives Italiennes De Biologie, 128(1), 19-45.

Luccarini, P., Gahery, Y., & Pompeiano, O. (1990b). Cholinoceptive pontine reticular structures

modify the postural adjustments during the limb movements induced by cortical stimulation.

Archives Italiennes De Biologie, 128(1), 19-45.

Lv, J., Li, Y., & Gu, Z. (2010). Decoding hand movement velocity from electroencephalogram

signals during a drawing task. Biomedical Engineering Online, 9(1), 64.

Martín-López, D., Jiménez-Jiménez, D., Cabañés-Martínez, L., Selway, R. P., Valentín, A., &

Alarcón, G. (2017a). The role of thalamus versus cortex in epilepsy: Evidence from human

ictal centromedian recordings in patients assessed for deep brain stimulation. International

Journal of Neural Systems, 27(07), 1750010.

Martín-López, D., Jiménez-Jiménez, D., Cabañés-Martínez, L., Selway, R. P., Valentín, A., &

Alarcón, G. (2017b). The role of thalamus versus cortex in epilepsy: Evidence from human

ictal centromedian recordings in patients assessed for deep brain stimulation. International

Journal of Neural Systems, 27(07), 1750010. 141

Massion, J. (1992). Movement, posture and equilibrium: Interaction and coordination. Progress

in Neurobiology, 38(1), 35-56.

Matsuyama, K., & Drew, T. (1997). Organization of the projections from the pericruciate cortex

to the pontomedullary brainstem of the cat: A study using the anterograde tracer phaseolus

vulgaris‐leucoagglutinin. Journal of Comparative Neurology, 389(4), 617-641.

Matsuyama, K., Mori, F., Nakajima, K., Drew, T., Aoki, M., & Mori, S. (2004). Locomotor role

of the corticoreticular–reticulospinal–spinal interneuronal system. Progress in Brain

Research, 143, 239-249.

Maung, D., Crawfis, R., Gauthier, L. V., Worthen-Chaudhari, L., Lowes, L. P., Borstad, A., &

McPherson, R. J. (2013). (2013). Games for therapy: Defining a grammar and

implementation for the recognition of therapeutic gestures. Paper presented at the Fdg, 314-

321.

Maung, D., Crawfis, R., Gauthier, L. V., Worthen-Chaudhari, L., Lowes, L. P., Borstad, A., . . .

Adams, J. (2014). (2014). Development of recovery rapids-A game for cost effective stroke

therapy. Paper presented at the Fdg,

McKiernan, B. J., Marcario, J. K., Karrer, J. H., & Cheney, P. D. (1998). Corticomotoneuronal

postspike effects in shoulder, elbow, wrist, digit, and intrinsic hand muscles during a reach

and prehension task. Journal of Neurophysiology, 80(4), 1961-1980.

142

Montgomery, L. R., Herbert, W. J., & Buford, J. A. (2013a). Recruitment of ipsilateral and

contralateral upper limb muscles following stimulation of the cortical motor areas in the

monkey. Experimental Brain Research, 230(2), 153-164.

Montgomery, L. R., Herbert, W. J., & Buford, J. A. (2013b). Recruitment of ipsilateral and

contralateral upper limb muscles following stimulation of the cortical motor areas in the

monkey. Experimental Brain Research, 230(2), 153-164.

Montgomery, L. R. (2013). Role of the reticulospinal and corticoreticular systems for the control

of reaching in non human primates.

Morabito, F. C., Campolo, M., Mammone, N., Versaci, M., Franceschetti, S., Tagliavini, F., . . .

Labate, A. (2017). Deep learning representation from electroencephalography of early-stage

creutzfeldt-jakob disease and features for differentiation from rapidly progressive dementia.

International Journal of Neural Systems, 27(02), 1650039.

Morris, D., Taub, E., & Mark, V. (2006). Constraint-induced movement therapy: Characterizing

the intervention protocol. Europa Medicophysica, 42(3), 257.

Myrhaug, H. T., & Ostensjo, S. (2014). Motor training and physical activity among preschoolers

with cerebral palsy: A survey of parents' experiences. Physical & Occupational Therapy in

Pediatrics, 34(2), 153-167. doi:10.3109/01942638.2013.810185 [doi]

Nathan, P. W., & Smith, M. C. (1982). The rubrospinal and central tegmental tracts in man.

Brain : A Journal of Neurology, 105(Pt 2), 223-269.

143

O’Sullivan, J., Chen, Z., Herrero, J., McKhann, G. M., Sheth, S. A., Mehta, A. D., & Mesgarani,

N. (2017). Neural decoding of attentional selection in multi-speaker environments without

access to clean sources. Journal of Neural Engineering, 14(5), 056001.

Ofner, P., & Müller-Putz, G. R. (2012). (2012). Decoding of velocities and positions of 3D arm

movement from EEG. Paper presented at the Engineering in Medicine and Biology Society

(EMBC), 2012 Annual International Conference of the IEEE, 6406-6409.

Ortega-Zamorano, F., Jerez, J. M., Juárez, G. E., & Franco, L. (2017). FPGA implementation of

neurocomputational models: Comparison between standard back-propagation and C-mantec

constructive algorithm. Neural Processing Letters, 46(3), 899-914.

Ortiz-Rosario, A., & Adeli, H. (2013). Brain-computer interface technologies: From signal to

action. Reviews in the Neurosciences, 24(5), 537-552.

Ortiz-Rosario, A., Adeli, H., & Buford, J. A. (2015). Wavelet methodology to improve single

unit isolation in primary motor cortex cells. Journal of Neuroscience Methods, 246, 106-

118.

Ortiz-Rosario, A., Berrios-Torres, I., Adeli, H., & Buford, J. A. (2014). Combined corticospinal

and reticulospinal effects on upper limb muscles. Neuroscience Letters, 561, 30-34.

OSC. (1987). Ohio supercomputer center .

144

Padillo, F., Luna, J. M., Herrera, F., & Ventura, S. (2018). Mining association rules on big data

through MapReduce genetic programming. Integrated Computer-Aided Engineering,

(Preprint), 1-19.

Peterson, B. W., Anderson, M. E., & Filion, M. (1974). Responses of ponto-medullary reticular

neurons to cortical, tectal and cutaneous stimuli. Experimental Brain Research, 21(1), 19-

44.

Polli, A., Moseley, G. L., Gioia, E., Beames, T., Baba, A., Agostini, M., . . . Turolla, A. (2017).

Graded motor imagery for patients with stroke: A non-randomized controlled trial of a new

approach. European Journal of Physical and Rehabilitation Medicine, 53(1), 14-23.

doi:10.23736/S1973-9087.16.04215-5 [doi]

Qarib, H., & Adeli, H. (2015). A new adaptive algorithm for automated feature extraction in

exponentially damped signals for health monitoring of smart structures. Smart Materials

and Structures, 24(12), 125040.

Rafiei, M. H., & Adeli, H. (2015). A novel machine learning model for estimation of sale prices

of real estate units. Journal of Construction Engineering and Management, 142(2),

04015066.

Rafiei, M. H., & Adeli, H. (2017). NEEWS: A novel earthquake early warning model using

neural dynamic classification and neural dynamic optimization. Soil Dynamics and

Earthquake Engineering, 100, 417-427.

145

Rafiei, M. H., Khushefati, W. H., Demirboga, R., & Adeli, H. (2017a). Supervised deep

restricted boltzmann machine for estimation of concrete. ACI Materials Journal, 114(2)

Rafiei, M. H., Khushefati, W. H., Demirboga, R., & Adeli, H. (2017b). Supervised deep

restricted boltzmann machine for estimation of concrete. ACI Materials Journal, 114(2)

Riddle, C. N., & Baker, S. N. (2010). Convergence of pyramidal and medial brain stem

descending pathways onto macaque cervical spinal interneurons. Journal of

Neurophysiology, 103(5), 2821-2832.

Rowe, D. L., Robinson, P. A., & Rennie, C. J. (2004). Estimation of neurophysiological

parameters from the waking EEG using a biophysical model of brain dynamics. Journal of

Theoretical Biology, 231(3), 413-433.

Sankari, Z., & Adeli, H. (2011). Probabilistic neural networks for diagnosis of alzheimer's

disease using conventional and wavelet coherence. Journal of Neuroscience Methods,

197(1), 165-170.

Schepens, B., & Drew, T. (2003). Strategies for the integration of posture and movement during

reaching in the cat. Journal of Neurophysiology, 90(5), 3066-3086.

doi:10.1152/jn.00339.2003 [doi]

Schepens, B., & Drew, T. (2004). Independent and convergent signals from the pontomedullary

reticular formation contribute to the control of posture and movement during reaching in the

cat. Journal of Neurophysiology, 92(4), 2217-2238. doi:10.1152/jn.01189.2003 [doi]

146

Schepens, B., & Drew, T. (2006). Descending signals from the pontomedullary reticular

formation are bilateral, asymmetric, and gated during reaching movements in the cat.

Journal of Neurophysiology, 96(5), 2229-2252. doi:00342.2006 [pii]

Siddique, N., & Adeli, H. (2013). Computational intelligence: Synergies of fuzzy logic, neural

networks and evolutionary computing John Wiley & Sons.

Sotero, R. C., & Trujillo-Barreto, N. J. (2008). Biophysical model for integrating neuronal

activity, EEG, fMRI and metabolism. NeuroImage, 39(1), 290-309.

Soteropoulos, D. S., Williams, E. R., & Baker, S. N. (2012). Cells in the monkey ponto‐

medullary reticular formation modulate their activity with slow finger movements. The

Journal of Physiology, 590(16), 4011-4027.

Standen, P., Threapleton, K., Richardson, A., Connell, L., Brown, D., Battersby, S., . . . Burton,

A. (2017). A low cost virtual reality system for home based rehabilitation of the arm

following stroke: A randomised controlled feasibility trial. Clinical Rehabilitation, 31(3),

340-350.

Szabo, J., & Cowan, W. (1984). A stereotaxic atlas of the brain of the cynomolgus monkey

(macaca fascicularis). Journal of Comparative Neurology, 222(2), 265-300.

Tabar, Y. R., & Halici, U. (2016a). A novel deep learning approach for classification of EEG

motor imagery signals. Journal of Neural Engineering, 14(1), 016003.

147

Tabar, Y. R., & Halici, U. (2016b). A novel deep learning approach for classification of EEG

motor imagery signals. Journal of Neural Engineering, 14(1), 016003.

Tanji, J., Okano, K., & Sato, K. C. (1988). Neuronal activity in cortical motor areas related to

ipsilateral, contralateral, and bilateral digit movements of the monkey. Journal of

Neurophysiology, 60(1), 325-343.

Taub, E., Uswatte, G., Mark, V., & Morris, D. (2006). The learl1ed nonuse phenomenon:

Implications for rehabilitation. Eura Medicophys, 42, 241-255.

Taub, E. (1976). Movement in nonhuman primates deprived of omatosensory feedback. Exercise

and Sport Sciences Reviews, 4(1), 335-374.

Taub, E., Miller, N. E., Novack, T. A., Cook, E. W.,3rd, Fleming, W. C., Nepomuceno, C. S., . . .

Crago, J. E. (1993). Technique to improve chronic motor deficit after stroke. Archives of

Physical Medicine and Rehabilitation, 74(4), 347-354.

Taub, E., Uswatte, G., King, D. K., Morris, D., Crago, J. E., & Chatterjee, A. (2006a). A

placebo-controlled trial of constraint-induced movement therapy for upper extremity after

stroke. Stroke, 37(4), 1045-1049. doi:01.STR.0000206463.66461.97 [pii]

Taub, E., Uswatte, G., King, D. K., Morris, D., Crago, J. E., & Chatterjee, A. (2006b). A

placebo-controlled trial of constraint-induced movement therapy for upper extremity after

stroke. Stroke, 37(4), 1045-1049. doi:01.STR.0000206463.66461.97 [pii]

148

Taub, E., Uswatte, G., Mark, V. W., Morris, D. M., Barman, J., Bowman, M. H., . . . Bishop-

McKay, S. (2013). Method for enhancing real-world use of a more affected arm in chronic

stroke: Transfer package of constraint-induced movement therapy. Stroke, 44(5), 1383-

1388. doi:10.1161/STROKEAHA.111.000559 [doi]

Taylor, G. S., & Schmidt, C. (2012). Empirical evaluation of the emotiv EPOC BCI headset for

the detection of mental actions. Proceedings of the Human Factors and Ergonomics Society

Annual Meeting, 56(1), 193-197. doi:10.1177/1071181312561017

Thibault, Robert T., Lifshitz, Michael,Raz, Amir,. (2016). The self-regulating brain and

neurofeedback: Experimental science and clinical promise. Cortex, 74

Uswatte, G., & Taub, E. (2013). In Merzenich M. M., Nahum M. and Van Vleet T. M.(Eds.),

Chapter 15 - constraint-induced movement therapy: A method for harnessing

neuroplasticity to treat motor disorders Elsevier. doi:https://doi.org/10.1016/B978-0-444-

63327-9.00015-1

Valenzuela, O., Jiang, X., Carrillo, A., & Rojas, I. (2018). Multi-objective genetic algorithms to

find most relevant volumes of the brain related to alzheimer's disease and mild cognitive

impairment. International Journal of Neural Systems,

Van Doren, J., Arns, M., Heinrich, H., Vollebregt, M. A., Strehl, U., & Loo, S. K. (2018).

Sustained effects of neurofeedback in ADHD: A systematic review and meta-analysis.

European Child & Adolescent Psychiatry, , 1-13.

149

Velliste, M., Perel, S., Spalding, M. C., Whitford, A. S., & Schwartz, A. B. (2008). Cortical

control of a prosthetic arm for self-feeding. Nature, 453(7198), 1098.

Wolf, S. L., Winstein, C. J., Miller, J. P., Taub, E., Uswatte, G., Morris, D., . . . Excite

Investigators. (2006a). Effect of constraint-induced movement therapy on upper extremity

function 3 to 9 months after stroke: The EXCITE randomized clinical trial. Jama, 296(17),

2095-2104.

Wolf, S. L., Winstein, C. J., Miller, J. P., Taub, E., Uswatte, G., Morris, D., . . . Excite

Investigators. (2006b). Effect of constraint-induced movement therapy on upper extremity

function 3 to 9 months after stroke: The EXCITE randomized clinical trial. Jama, 296(17),

2095-2104.

Woodbury, M., Velozo, C. A., Thompson, P. A., Light, K., Uswatte, G., Taub, E., . . . Nichols-

Larsen, D. S. (2010). Measurement structure of the wolf motor function test: Implications

for motor control theory. Neurorehabilitation and Neural Repair, 24(9), 791-801.

Wright, J., & Jordanov, I. (2017). Quantum inspired evolutionary algorithms with improved

rotation gates for real-coded synthetic and real world optimization problems. Integrated

Computer-Aided Engineering, 24(3), 203-223.

Zaaimi, B., Dean, L. R., & Baker, S. N. (2017). Different contributions of primary motor cortex,

reticular formation, and spinal cord to fractionated muscle activation. Journal of

Neurophysiology, 119(1), 235-250.

150

Zhang, S. Y., Jiang, J. Z., & Neild, S. (2017). Optimal configurations for a linear vibration

suppression device in a multi‐storey building. Structural Control and Health Monitoring,

24(3), e1887.

Zhang, Y., Wang, Y., Jin, J., & Wang, X. (2017). Sparse bayesian learning for obtaining sparsity

of EEG frequency bands based feature vectors in motor imagery classification. International

Journal of Neural Systems, 27(02), 1650032.

Zotev, V., Phillips, R., Misaki, M., Wong, C. K., Wurfel, B. E., Krueger, F., . . . Bodurka, J.

(2018). Real-time fMRI neurofeedback training of the amygdala activity with simultaneous

EEG in veterans with combat-related PTSD doi:https://doi.org/10.1016/j.nicl.2018.04.010

151