University of Southern California Dissertations and Theses

Development and implementation of a modular muscle-computer interface for personalized motor rehabilitation after stroke

(USC Thesis Other)

Development and implementation of a modular muscle-computer interface for personalized motor rehabilitation after stroke

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Development and Implementation of a Modular Muscle-Computer Interface
for Personalized Motor Rehabilitation After Stroke

by

Octavio Marin-Pardo

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(BIOMEDICAL ENGINEERING)

May 2023

Copyright 2023 Octavio Marin-Pardo

ii
Dedication

To my grandparents, my parents, and my sister.
¡Gracias!

iii
Acknowledgments
Many (if not all) PhDs are forged with the help, guidance, and support of small and big
communities; mine was not the exception. As it would not be possible to name all the people who
have helped me directly and indirectly in this journey, I will try my best to recognize here those
who played the most significant roles in this achievement.
First and foremost, I would like to thank my outstanding advisor, Dr. Sook-Lei Liew. Words
are not enough to express my utmost gratitude, respect, and admiration. Thank you for giving me
the opportunity to work in your laboratory, providing me with vast resources to develop my skills
and scientific career, and showing me the true meaning of “standing on the shoulders of giants”.
Under your kind and patient guidance, I got to understand the myriad of aspects that entail doing
“good science”, including being a critical thinker, communicating clearly and effectively, and
working with the highest standards. I am incredibly fortunate and will be forever grateful for having
a kind and caring mentor who tirelessly went above and beyond to help her students grow. Thank
you for your continuous support and encouragement and for making this experience one of the
greatest of my life.
Next, I thank my dissertation committee: Dr. James Finley, Dr. Kristan Leech, and Dr. Nicholas
Schweighofer. Thank you so much for the enriching discussions inside and outside our Journal
Club. Those conversations helped me deeply to gain a better understanding of the fascinating
field of human motor control and to sharpen my critical thinking as a scientist. I would like to also
extend my sincere appreciation to Dr. Francisco Valero Cuevas for providing the golden ticket to
this incredible adventure. Thank you for all your time, effort, and help in pursuing my graduate
studies at USC. The summer research experience you developed constituted the cornerstone
upon which my doctoral studies rest, and for that, I will remain grateful. My gratitude as well Dr.
Jesús Manuel Dorador González for your aid and support during that summer, and Dr. Krishna
iv
Nayak for giving me a space in your laboratory and introducing me to the fascinating world of
academic research.
I also thank all past and present fellow NPNL members, with special thanks to Lily, Julia,
Artemis, Stephanie, David, Meghan, Anisha, Sharon, Miranda, Coralie, Bethany, Kira, Jessica,
and Erendiz. You truly made this a fantastic and enjoyable experience. Thank you for all the fun,
support, and insight and for sharing delicious meals. I could not have wished for a more caring
and warmer lab, and I am glad and thankful for our friendship. Special thanks to Thanos and
Chris, who not only played a significant role in contributing to the REINVENT project but also to
my development as a researcher and engineer. Thank you for all the lessons, encouragement,
and laughs. Likewise, I thank our friend laboratories in the Biokinesiology Division and all the
great people in them, particularly Pouria, Natalia, Chang, Aram, Victor, Akira, Rini, Yannick, and
Sarah. Thank you for all the discussions and fun times inside and outside HSC. Major thanks to
all the wonderful participants who very generously contributed with their time and patience to this
work. I genuinely enjoyed our time together and will forever carry the lessons I learned from you.
Beyond academic advising and support, this section would not be complete without
acknowledging the work and efforts of the Biomedical Engineering Department, the Viterbi School
of Engineering, and the Chan Division of Occupational Science and Occupational Therapy.
Special thanks to William Yang, Mischal Diasanta, Tracy Charles, Andy Chen, Patty Gutierrez,
and Sonia De Mesa for their invaluable help throughout different stages of this process.
Part of my journey through grad school was also marked by amazing people I met outside of
the lab and who helped keep me sane. Rodrigo, Darío, Arturo, Víctor, José, and Santiago, thank
you for making my time away from Mexico more memorable with great conversations and fun
beyond compare. I will always cherish the great moments we shared and look forward to many
more. Thank you for your friendship and for bringing “a little bit of home” into our many meetings.
I would also like to thank all the funding sources that supported me and the REINVENT project
over the years: the USC Graduate School, the Mexican National Council for Science and
v
Technology (CONACyT), the USC Stevens Center of Innovation, the American Heart Association,
the U.S. Army Research Office, and the National Institutes of Health.
Last but definitely not least, I thank my family, without whom none of this would have been
possible. To my parents, Octavio and Amelia, thank you for your continued support, care, and
love throughout my life, for providing me with the tools to care for myself and others, and for
creating an environment where I could quench my thirst for knowledge and achieve my goals. To
my sister Tania, thank you for being my constant companion and source of inspiration, support,
and encouragement. To my grandparents, Amelia, Armando, and Octavio thank you for your
unconditional love and for providing the strong roots upon which my achievements rest.

vi
Table of Contents
Dedication .................................................................................................................................. ii
Acknowledgments ..................................................................................................................... iii
List of Tables ............................................................................................................................. ix
List of Figures.............................................................................................................................x
Abstract ..................................................................................................................................... xi
Chapter 1: Introduction ............................................................................................................. 1
1.1 Introduction ...................................................................................................................... 1
1.1.1 Overview ................................................................................................................... 1
1.1.2 Organization of the Dissertation ............................................................................... 4
1.2 Aim One Background ...................................................................................................... 5
1.2.1 Using Virtual Reality and Brain-Computer Interfaces to Engage the Action-
Observation Network .................................................................................................5
1.2.2 REINVENT, A Brain-Computer Interface for Virtual Reality-Based
Neurofeedback ..........................................................................................................6
1.2.3 Electromyography Biofeedback ............................................................................... 7
1.4.1 EEG-EMG Coherence in Rehabilitation ................................................................... 8
1.2.4 Aim One Purpose ..................................................................................................... 9
1.3 Aims Two and Three Background .................................................................................. 9
1.3.1 Challenges in Clinic-based and Home-based Rehabilitation .................................. 9
1.3.2 Technology-based Rehabilitation Implementation Challenges ............................. 10
1.3.3 Aims Two and Three Purpose................................................................................ 11
Chapter 2: A Virtual Reality Muscle–Computer Interface for Neurorehabilitation in Chronic
Stroke: A Pilot Study ............................................................................................. 12
2.1 Abstract ......................................................................................................................... 12
2.2. Introduction................................................................................................................... 13
2.3 Materials and Methods .................................................................................................. 16
2.3.1 Participants ............................................................................................................. 16
2.3.2 Study Timeline ........................................................................................................ 17
2.3.3 Clinical Assessments (Sessions 1 and 10) ............................................................ 19
2.3.4 Additional Data Acquired ........................................................................................ 20
2.3.5 Physiological Recordings and Analysis ................................................................. 20
2.3.6 Static Hold Task: Characterization of Muscle Control during EMG Amplitude
Target Tracking (Sessions 1 and 10) ..................................................................... 21
2.3.7 Wrist Extensor Training in Virtual Reality (Sessions 2–9) ..................................... 23
2.3.8 Statistical Analyses ................................................................................................ 25
2.4 Results........................................................................................................................... 27
2.4.1 Feasibility ................................................................................................................ 27
2.4.2 Behavioral Changes Following Training ................................................................ 27
vii
2.4.3 Changes of Muscle Control during EMG Amplitude Target Tracking ................... 29
2.4.4 Neuromuscular Changes Following Training ......................................................... 30
2.4.5 Changes Across Training Sessions ....................................................................... 33
2.5 Discussion ..................................................................................................................... 34
2.5.1 Summary ................................................................................................................ 34
2.5.2 Feasibility and Acceptability ................................................................................... 35
2.5.3 Clinical Assessments ............................................................................................. 36
2.5.4 Neuromuscular Control .......................................................................................... 38
2.5.5 Training Effects versus Task Performance ............................................................ 40
2.5.6 Limitations and Conclusions .................................................................................. 41
Chapter 3: Development of a Low-Cost, Modular Muscle–Computer Interface for At-Home
Telerehabilitation for Chronic Stroke .................................................................... 43
3.1 Abstract ......................................................................................................................... 43
3.2 Introduction .................................................................................................................... 44
3.3 Materials and Methods .................................................................................................. 46
3.3.1 Participants ............................................................................................................. 46
3.3.2 System Architecture ............................................................................................... 46
3.3.3 Acquisition .............................................................................................................. 49
3.3.4 EMG Signal Processing and Biofeedback ............................................................. 49
3.3.5 Example of a Tele-REINVENT Telerehabilitation Session with a Clinician .......... 53
3.3.6 Signal Quality Validation Example ......................................................................... 55
3.3.7 Case Study and Feasibility ..................................................................................... 58
3.4 Results........................................................................................................................... 60
3.4.1 Signal Quality Validation Example ......................................................................... 60
3.4.2 Feasibility, Safety, and Adherence ........................................................................ 61
3.5 Discussion ..................................................................................................................... 62
Chapter 4: Functional and Neuromuscular Changes Induced via a Low-Cost, Muscle-
Computer Interface for Telerehabilitation: A Feasibility Study in Chronic Stroke 67
4.1 Abstract ......................................................................................................................... 67
4.2 Introduction .................................................................................................................... 68
4.3 Materials and Methods .................................................................................................. 71
4.3.1 Participants ............................................................................................................. 71
4.3.2 Study Timeline ........................................................................................................ 72
4.3.3 At-Home Training (Sessions 2–31) ........................................................................ 73
4.3.4 Clinical Assessments (Sessions 1 and 32) ............................................................ 76
4.3.5 Characterization of Muscle Control During EMG Amplitude Tracking (Sessions
1 and 32)................................................................................................................. 77
4.3.6 Statistical Analyses ................................................................................................ 80
4.4 Results........................................................................................................................... 82
4.4.1 Feasibility, Safety, and Acceptability...................................................................... 82
4.4.2 Behavioral Changes in Clinical Assessments ....................................................... 83
4.4.3 Muscle Control Changes During EMG Amplitude Tracking .................................. 84
4.4.4 Changes in Corticomuscular Coherence Following Training ................................ 86
4.4.5 Increased EMG Individuation Across Sessions ..................................................... 88
viii
4.5 Discussion ..................................................................................................................... 89
4.5.1 Clinical Assessments ............................................................................................. 89
4.5.2 Improvements in Muscle Group Individuation ....................................................... 92
4.5.3 Changes in Corticomuscular Coherence ............................................................... 93
4.5.4 Limitations and Future Directions .......................................................................... 95
Chapter 5: Discussion ............................................................................................................. 97
5.1 Summary of Key Findings ............................................................................................. 97
5.1.1 Chapter 2 Summary – Aim One ............................................................................. 97
5.1.2 Chapter 3 Summary – Aim Two ............................................................................. 98
5.1.3 Chapter 4 Summary – Aim Three .......................................................................... 98
5.2 Implications and Significance ....................................................................................... 99
5.2.1 Clinical Implications ................................................................................................ 99
5.2.2 Research and Technological Implications ........................................................... 100
5.3 Limitations and Future Directions ............................................................................... 101
References ............................................................................................................................ 104
Appendix A: List of Relevant Publications ............................................................................ 121

ix
List of Tables
Table 2.1 Participant demographics and baseline evaluations. .................................................. 16
Table 2.2 Statistical comparisons for clinical assessments. ........................................................ 28
Table 2.3 Group level analysis of within-game performance and grip strength. ......................... 34
Table 3.1 Score likelihood for the SkeeBall game as a function of the extensor ratio (ER). ...... 52
Table 3.2 Sensor specifications.................................................................................................... 55
Table 3.3 Muscle activity correlation with grip force while building-up to a sustained grip. ........ 60
Table 4.1 Participant demographics and baseline evaluations. .................................................. 71
Table 4.2 Clinical assessments before and after training. ........................................................... 84
Table 4.3 Muscle individuation during extension and flexion tracking. ........................................ 86
Table 4.4 Changes of muscle activity over time during remote training sessions....................... 89

x
List of Figures
Figure 1.1 REINVENT system. ....................................................................................................... 7
Figure 2.1 Experimental protocol.................................................................................................. 18
Figure 2.2 Clinical assessments. .................................................................................................. 28
Figure 2.3 Performance changes in EMG amplitude control. ...................................................... 30
Figure 2.4 Corticomuscular coherence during static flexion and extension. ............................... 32
Figure 2.5 Muscle activity and performance during training. ....................................................... 33
Figure 3.1 Tele-REINVENT architecture, prototype, and interfaces. .......................................... 48
Figure 3.2 Examples of EMG biofeedback games. ...................................................................... 52
Figure 3.3 Feedback displayed during a session guided by an occupational therapist. ............. 54
Figure 3.4 Test of muscle control. ................................................................................................ 57
Figure 3.5 Muscle activity during REINVENT sessions. .............................................................. 62
Figure 4.1 Experimental protocol.................................................................................................. 73
Figure 4.2 Clinical assessments before and after training. .......................................................... 84
Figure 4.3 Muscle group individuation during EMG amplitude control. ....................................... 85
Figure 4.4 Corticomuscular coherence (CMC) during static extension and flexion. ................... 87
Figure 4.5 Muscle activity during remote training sessions. ........................................................ 88

xi
Abstract
Stroke is a leading cause of long-term adult disability. People with severe motor impairment
(e.g., little to no residual hand or wrist movement) in the chronic phase (more than six months
after stroke onset) have limited rehabilitation options. High doses of repeated movement can
improve upper limb function. However, it remains a challenge to provide such doses in standard
clinical practice. Therefore, there is an urgent need for alternative therapeutic interventions for
chronic stroke survivors with severe motor impairment. In addition, engaging and goal-directed
exercise can be implemented via gamified training programs. Specifically, providing feedback of
muscle activity to avoid unintended coactivation of antagonistic muscles has been shown to
improve motor function. Therefore, in this dissertation I present and discuss the development and
test of REINVENT, a low-cost, portable, modular, mixed-reality, biofeedback system for in-person
and remote neurorehabilitation. The main goal is to evaluate the feasibility of using REINVENT
with different feedback modalities (e.g., immersive and screen-based) in various training
environments (e.g., in-person and at-home). Accordingly, we asked ten severely impaired chronic
stroke survivors to use REINVENT to practice individuated use of wrist muscles during three
separate case studies. Our combined results suggest that this approach is feasible, safe, and
capable of inducing positive outcomes in stroke rehabilitation across varied functional domains.
Furthermore, these improvements were presented along indications of neuroplastic changes and
improved muscle control. Together, this work supports the expansion of using inexpensive and
customizable technology to aid clinicians in improving people’s quality of life after stroke.

1
Chapter 1: Introduction
1.1 Introduction
1.1.1 Overview
Almost 800,000 people have a new or recurrent stroke in the United States each year (Tsao
et al., 2022). This disease is defined as a vascular accident where brain tissue is permanently
damaged due to the lack of oxygenation and nutrients caused by the burst of blood vessels
(hemorrhagic stroke) or a clot impeding regular blood flow (ischemic stroke). Thanks to medical
advances, the mortality rate due to stroke has decreased over the last decades (Gittler & Davis,
2018). However, most survivors will deal with impaired brain processes, in which the location and
size of the lesioned brain area are typically correlated with the type and severity of impairment
(Langhorne et al., 2011; Lindenberg et al., 2010). This makes stroke the third leading cause of
global disability and the leading cause of disability in the United States (Feigin et al., 2021; Tsao
et al., 2022). One of the primary disabilities caused by a stroke is motor impairment, especially in
one of the upper extremities (Gadidi et al., 2011; Rathore et al., 2002). Given the burden this
imposes on patients, caregivers, and the healthcare system, it is critical to find therapeutic
solutions that can effectively help improve motor function.
Despite great effort invested in understanding and overcoming motor impairments after
stroke, further research is required to consistently induce recovery. Typically, recovery after stroke
is defined as “the change (mostly improvement) of a given outcome that is achieved by an
individual between two (or more) time points or the mechanism underlying this improvement in
terms of behavioral restitution or compensation strategies” (Bernhardt et al., 2017). These
changes are thought to be induced by a combination of spontaneous and learning-dependent
processes that aim to regain different levels of motor function (Bernhardt et al., 2017; Kwakkel et
al., 2004; Langhorne et al., 2011). Previously, it was assumed that recovery plateaus after
reaching the chronic phase (more than six months after the vascular accident) since most of the
2
neuroplastic processes that allow for spontaneous recovery take place within the first weeks after
onset (Bernhardt et al., 2017; Langhorne et al., 2011). Thus, most rehabilitation services are
focused on that recovery window, limiting the available options for chronic stroke survivors. For
this population, clinical practice usually focuses on compensatory strategies. However, recent
studies have shown that it is possible to promote recovery beyond the 6-month mark (Ballester et
al., 2019). For example, evidence suggests that high doses of repeated task-specific practice
constitute a suitable non-invasive intervention to encourage functional recovery (Lohse et al.,
2014; Stinear, 2017; Ward et al., 2019). However, several treatments that incorporate principles
of repeated practice and have been shown to be effective in restoring motor ability (e.g.,
constrained-induced movement therapy) often restrict their inclusion criteria to participants that
present with a minimum level of function (e.g., more than 10 degrees of active joint motion)
(Winstein et al., 2003; Wolf et al., 2010). Therefore, there is a dire need to provide people with
severe and chronic impairments with more restorative options to improve their movement abilities
and quality of life.
Biofeedback is a non-invasive intervention that allows for task-specific training without
necessitating overt movement. In short, biofeedback consists of making a person aware of a
mechanical or physiological process so they can learn to control it voluntarily (Giggins et al., 2013;
Huang et al., 2006). It requires direct or indirect measurement of said process (e.g., breathing,
muscle, or brain activity) and uses external cues (e.g., visual, auditory, or haptic) to give inform
the person how their intention and behavior alter the measured variable (Wolf, 1983). Then,
practice allows the person to find a strategy that gives them control of the targeted process.
Previous work in stroke rehabilitation has demonstrated that using biofeedback to gain control
over brain activity (i.e., neurofeedback) has positive benefits on motor function (Bai et al., 2020;
Cervera et al., 2018; Ramos-Murguialday et al., 2013; Soekadar et al., 2015). This can be
accomplished through different modalities, for example, using functional magnetic resonance
imaging (fMRI) or electroencephalography (EEG). In fMRI neurofeedback, researchers use
3
powerful electromagnets to measure changes in the brain’s blood flow to quantify the involvement
of cortical and subcortical brain areas in the task of interest (Liew et al., 2016; Weiskopf et al.,
2003). In EEG neurofeedback, researchers use arrays of electrodes to measure the electrical
activity produced by populations of neurons in regions of the cerebral cortex and quantify their
activity during the execution of a task (Pfurtscheller et al., 2003; Vourvopoulos & Bermúdez i
Badia, 2016). Research with both modalities has shown that, when measuring the activity of brain
areas involved in motor control (e.g., the sensorimotor cortex), participants are capable of learning
how to modulate this activity and that this training induces positive outcomes (Liew et al., 2016;
Matarasso et al., 2021; Ramos-Murguialday et al., 2013). It has been proposed that these forms
of biofeedback may allow closing the feedback loop between a neural instruction that the brain
generates and the subsequent outcome that it is expecting to happen. This is based on the
knowledge that some motor control mechanisms rely on internal (e.g., proprioception) and
external feedback (e.g., visual stimuli) to produce accurate movements (Krakauer et al., 2019).
However, these types of neurofeedback introduce limitations in terms of accessibility, cost, and
personalization since most of their training paradigms require using expensive equipment in
research facilities and with limited personalization parameters. Measuring the electrical activity of
the muscles via electromyography (EMG) has the potential to overcome these barriers while
providing feedback of a signal that also originates in the brain.
Extensive research has focused on developing and validating behavioral assessments to
quantify motor impairment objectively. Examples frequently used in upper limb evaluations
include the Fugl-Meyer assessment, the Action Research Arm Test, the Box and Blocks test, and
quantification of the range of motion of different joints (e.g., the wrist) (Fugl-Meyer et al., 1975;
Lyle, 1981; Mathiowetz et al., 1985; Ryu et al., 1991). Together, these clinical assessments
provide reliable, stable, and accurate measurements that allow us to quantify and track
impairment over time in terms of overt quality of the movements, coordination, reflexes, and
functional performance (Gladstone et al., 2002; Platz et al., 2005). However, these metrics may
4
fail to capture more subtle and covert changes induced by training (e.g., variations of brain or
muscle activity) that could precede observable functional changes. To address this, previous
research has explored the relationship between changes in brain and muscle activity and changes
in behavioral measurements of impairment. Examples include evaluating changes in resting-state
brain activity, event-related desynchronization during movement imagery and execution, brain-
muscle connectivity, and muscle-muscle connectivity (Fong et al., 2021; Krauth et al., 2019; H.
Liu et al., 2022; Saes et al., 2019; Sebastián-Romagosa et al., 2020). Overall, getting a better
understanding of the neuroplastic processes that allow for the restoration of lost neural activity
could allow us to get further insights into the recovery process and enhance current rehabilitation
efforts.

1.1.2 Organization of the Dissertation
In this dissertation, I describe the development of an EMG-based biofeedback system that
aims to promote and quantify stroke recovery in severely impaired populations. The proposed
design builds on literature that shows the potential of using EMG signals to encourage motor
recovery after stroke and further considers the current needs of stroke survivors, clinicians, and
researchers. We combined this previous knowledge with our findings during pilot tests to create
a modular system that targets the specific needs of its users and allows for continued
development and upgrades. Specifically, in this thesis I aim to:
1. Determine whether it is feasible to use a virtual reality-based EMG biofeedback
system to improve upper limb control in people after stroke. We found that it was
feasible and safe to use the system during seven training sessions. We evaluated
motor control improvement with clinical and behavioral assessments and showed
moderate improvements across participants.
5
2. Develop a system for at-home EMG biofeedback post-stroke that incorporates
telerehabilitation services and test its feasibility. We found that it was feasible and
safe to use the system during 40 training sessions.
3. Evaluate the feasibility and efficacy of at-home EMG biofeedback training to
improve upper limb control in people after chronic severe stroke. We found that
higher doses of training in a home environment were tolerable and capable of induce
moderate improvements in our target population.
Overall, this dissertation presents seminal work that shows alternatives to better take
advantage of using technology to provide engaging training experiences and give clinicians more
tools to improve people’s quality of life after stroke. Furthermore, I emphasize the importance of
proactively incorporating users’ needs and input in the design of rehabilitation devices to provide
more suitable solutions to current healthcare challenges. In the remainder of this chapter, I will
present literature reviews that support our design choices and the rationale of each aim.

1.2 Aim One Background
1.2.1 Using Virtual Reality and Brain-Computer Interfaces to Engage the Action-Observation
Network
Human movement is complicated. Multiple systems in the brain are in charge of different
motor stages, from planning, selection, and prediction, to execution, control, and supervision, to
name a few (Doya, 2000; Krakauer et al., 2019; Spampinato & Celnik, 2020). Extensive literature
has studied how these and other systems interact to produce effective movements in neurotypical
and impaired populations. However, there is much to learn about motor control and how to use
this knowledge to produce more effective therapeutic interventions. Of particular interest is the
Action Observation Network (AON), which corresponds to the brain regions that are active when
an individual executes, observes, or imagines an action (Celnik et al., 2008; Garrison et al., 2013).
Previous research has shown the feasibility of improving motor performance via activation of the
6
AON in people who do not have accurate motor control after stroke (Franceschini et al., 2012;
Sugg et al., 2015; Zhang et al., 2018). Additionally, recent advances in computer interaction have
allowed the development of VR therapeutic experiences with high levels of embodiment (i.e., the
sense that visual perception of limb movements corresponds to the subject’s own limbs) and
presence (i.e., the sense that the subject is inside the virtual environment). Research suggests
that engaging embodiment and presence could be beneficial for stroke survivors and that the
AON may be involved in processing such feedback (Borrego et al., 2019; Holper et al., 2010;
Perez-Marcos, 2018). Furthermore, research has shown that VR can induce neuroplastic
changes after stroke and that using VR with other therapeutic approaches might be more
beneficial than using VR alone (Hao et al., 2022; Laver et al., 2017). Therefore, incorporating
measurement and feedback of brain activity via a brain-computer interface (BCI) with VR
technology may also add benefit to neurorehabilitation efforts. As mentioned in Section 1.1,
people can learn to modulate their brain activity and this training can induce positive outcomes
after stroke. Similarly, BCIs have also been suggested to improve motor function (Bai et al., 2020;
Cervera et al., 2018; Matarasso et al., 2021). Therefore, by integrating current evidence that
supports the activation of the AON via BCIs and VR, our research group developed REINVENT
(Rehabilitation Environment using the Integration of Neuromuscular-based Virtual Enhancements
for Neural Training), a first-of-its-kind BCI for VR-based stroke rehabilitation neurofeedback
(Spicer et al., 2017).

1.2.2 REINVENT, A Brain-Computer Interface for Virtual Reality-Based Neurofeedback
Preliminary work from our group showed that REINVENT is capable of inducing embodiment
during immersive neurofeedback training and that this correlated with improved performance in
neurotypical participants (Juliano et al., 2020). After upgrading the system (shown in Figure 1.1),
we investigated the feasibility of using it in a small sample (n=4) of stroke survivors with varying
levels of impairment (Vourvopoulos, Marin-Pardo, et al., 2019; Vourvopoulos, Pardo, et al., 2019).
7
Our results showed that the participant with greater impairment had good performance in the
training task (between 65% and 95% of successful trials) and showed modest changes in cortical
activity. However, participants with mild impairment showed greater variability in task control (20%
to 90% of success) and neural and behavioral assessments. We concurrently recorded EMG
signals from their wrist muscles while the participants were training with EEG biofeedback. A post
hoc analysis showed that the mildly-impaired participants might have had better task performance
had they used EMG to control the virtual environment (Marin-Pardo et al., 2019). Overall, our
preliminary analyses suggested that biofeedback after stroke might be more successful if it is
tailored for the participants’ level of impairment and that, for those with residual muscle activity,
EMG might be a more suitable control signal.

Figure 1.1 REINVENT system.
(Left) Hardware consisting of A) electroencephalography (EEG) electrodes and amplifier, B) immersive
head-mounted display for virtual reality (VR) with hand-held controllers, C) electromyography (EMG)
sensors, D) virtual environment. (Right) Software architecture consisting of interfaces for signal acquisition,
processing, and storage (connected via the Labstreaming Layer network protocol (LSL)) and feedback
visualization in VR. Figure adapted from (Vourvopoulos, Marin-Pardo, et al., 2019).

1.2.3 Electromyography Biofeedback
EMG, which represents the electrical activity generated by contracting muscles, has been
utilized to strengthen the affected limb of patients with limited voluntary movement. Specifically,
EMG biofeedback has been used since the late 1960s and was shown to improve joint range
8
motion and muscle strength, function, and relaxation after stroke (Huang et al., 2006; Wolf, 1983;
Woodford & Price, 2007). There is increasing evidence that EMG biofeedback might also be
beneficial to improve motor function in patients with residual but limited movement (Armagan et
al., 2003; Doğan-Aslan et al., 2012; Kim, 2017; Schleenbaker & Mainous, 1993; Woodford &
Price, 2007; Wright et al., 2014). In neurotypical populations, co-contraction of antagonist muscles
can assist in finer motor control, especially when learning a novel task (Latash, 2018). However,
research suggests that antagonistic coactivation after stroke may be detrimental to motor function
(Beer et al., 2000; Dewald et al., 1995; Zackowski, 2004). Therefore, specific training to avoid
unwanted coactivation of antagonist muscles might induce positive motor outcomes, as shown
by the works of others (Donoso Brown et al., 2014; Mugler et al., 2019; Wright et al., 2014).
Furthermore, using EMG to provide biofeedback training and quantify recovery could provide
additional information regarding brain processes that control the signal.

1.4.1 EEG-EMG Coherence in Rehabilitation
Much research has aimed to develop and validate behavioral assessments of impairment
(e.g., the Fugl-Meyer assessment and the action research arm test) (Fugl-Meyer et al., 1975;
Lyle, 1981; Page et al., 2012; van der Lee, Beckerman, et al., 2001). However, although these
assessments have demonstrated desirable properties (e.g., low intra- and inter-rater variability,
good quantification of functional impairment (Gladstone et al., 2002; Platz et al., 2005)), there is
also a need to capture more subtle and covert changes in motor recovery that could precede
observable functional changes (e.g., neurological changes due to functional reorganization).
Previous research has explored correlations between motor impairment after stroke and direct or
indirect measurements of brain activity (Fong et al., 2021; Krauth et al., 2019; Saes et al., 2019).
Of particular interest is the use of EEG-EMG coherence (i.e., corticomuscular coherence) and
EMG-EMG coherence (i.e., intermuscular coherence) due to their relatively low cost and high
portability. In short, corticomuscular coherence assesses the synchrony of muscles and the neural
9
signals that originate their activity (i.e., the corticospinal pathway from the primary motor cortex)
(J. Liu et al., 2019). Similarly, intermuscular coherence quantifies the extent of the shared neural
drive (i.e., control rhythms of cortical and spinal origin) common to a pair of muscles (Farina et
al., 2014). Previous research has attempted to characterize how stroke affects these
measurements and to use them for therapeutic neurofeedback approaches (T. Boonstra, 2013;
Guo et al., 2020; Krauth et al., 2019; Larsen et al., 2017). However, it is not clear how this proxy
of neuroplastic reorganization changes due to EMG biofeedback in chronic stroke populations.

1.2.4 Aim One Purpose
The objective of this aim is to develop and evaluate the feasibility of an EMG-based variation
of our REINVENT VR biofeedback rehabilitation system to increase volitional muscle activity while
reducing unintended co-contractions (Marin-Pardo et al., 2020). The main hypotheses are: (1)
reinforcement of EMG activity in participants with severe movement deficits would be feasible,
safe, and provide a positive experience for participants; (2) training would produce modest
improvements in clinical assessments comparable to what we have previously observed with
EEG-based neurofeedback.

1.3 Aims Two and Three Background
1.3.1 Challenges in Clinic-based and Home-based Rehabilitation
As mentioned in Section 1.1, research suggests that high doses of intensive repeated task-
specific practice can promote functional recovery in severely impaired populations (Lohse et al.,
2014; Stinear, 2017; Ward et al., 2019). However, adopting such approaches in standard clinical
practice comes with challenges. Several constraints have been identified, primarily in terms of
space (e.g., in-clinic and at-home physical space is often restricted), mobility (e.g., not all patients
can travel on a regular basis to attend therapy sessions), time (e.g., sessions provide an average
of 32 repetitions of upper extremity functional movements), and cost (e.g., insurance companies
10
often limit the procedures or technologies a therapist can incorporate into the session) (Health
Resources and Services Administration, 2022; Lang et al., 2009; Marzolini et al., 2016).
Telerehabilitation (i.e., rehabilitation services provided by a clinician via information and
communication technologies) has been identified as a potential solution to overcome many of
these limitations. Specifically, recent studies suggest that telerehabilitation incorporated into
stroke rehabilitation is feasible and can be as effective as in-person therapy (Cramer et al., 2019;
Dodakian et al., 2017). However, technology-based rehabilitation has introduced new challenges
that prevent the adoption of such solutions.

1.3.2 Technology-based Rehabilitation Implementation Challenges
Multiple portable systems for at-home rehabilitation after stroke have been developed over
the last few years to facilitate telerehabilitation services, improve quality of life, provide stroke
education, and improve movement function (Cramer et al., 2019; Dodakian et al., 2017).
Specifically, EMG biofeedback devices have been developed to improve accessibility and training
time (Donoso Brown et al., 2014; Jian et al., 2021; Mugler et al., 2019). However, previous studies
have identified barriers to incorporating EMG biofeedback, and other technology-based solutions,
in home environments (McFarland & Vaughan, 2016; Peters et al., 2015). Examples include:
participant’s low adherence, high economic costs (for stroke survivors and clinicians to acquire
the technology and to train clinicians on how to use it), and limitations in terms of required space
(e.g., the equipment can be cumbersome or requires specific environments to use it), setup time
(e.g., it can take as much time to set up and configure the system as the effective training time),
and technical literacy (e.g., physical and graphical interfaces are complicated and without “quick
set up” settings) (Chen et al., 2019; Donoso Brown et al., 2014; Feldner et al., 2019, 2020;
Hochstenbach-Waelen & Seelen, 2012). Furthermore, it has been proposed that having a clinician
continuously involved and allowing them to track their patient’s progress are essential in home-
based rehabilitation programs (Cramer et al., 2019; Feldner et al., 2019). Therefore, it is critical
11
to consider all these factors to develop not only scientific-based rehabilitation solutions (i.e., based
on proven principles that involve the appropriate mechanisms to induce recovery) but also to
proactively design the devices targeting the specific needs of their users. Thus, we accounted for
these elements during the development of Tele-REINVENT, a portable iteration of our REINVENT
system that incorporates telerehabilitation capabilities to provide EMG feedback for individuated
muscle training in home environments (Marin-Pardo et al., 2021).

1.3.3 Aims Two and Three Purpose
Overall, the objective of these aims is to determine the efficacy of EMG biofeedback
incorporated with home-based telerehabilitation to improve upper limb control in people with
severe chronic stroke. The main hypotheses are: (1) this system will be feasible, safe, and provide
a positive user experience for participants over extensive use (e.g., 30-40 days of training); (2)
training will induce modest improvements in clinical and physiological assessments, similar or
greater to those seen in Aim 1; and (3) participants will show a high adherence to the training
protocol and will have an overall positive experience.

12
Chapter 2: A Virtual Reality Muscle –Computer Interface for
Neurorehabilitation in Chronic Stroke: A Pilot Study
This section is adapted from:
Marin-Pardo, O., Laine, C. M., Rennie, M., Ito, K. L., Finley, J., & Liew, S.-L. (2020). A Virtual
Reality Muscle–Computer Interface for Neurorehabilitation in Chronic Stroke: A Pilot Study.
Sensors.

2.1 Abstract
Severe impairment of limb movement after stroke can be challenging to address in the chronic
stage of stroke (e.g., greater than 6 months post stroke). Recent evidence suggests that physical
therapy can still promote meaningful recovery after this stage, but the required high amount of
therapy is difficult to deliver within the scope of standard clinical practice. Digital gaming
technologies are now being combined with brain–computer interfaces to motivate engaging and
frequent exercise and promote neural recovery. However, the complexity and expense of
acquiring brain signals has held back widespread utilization of these rehabilitation systems.
Furthermore, for people that have residual muscle activity, electromyography (EMG) might be a
simpler and equally effective alternative. In this pilot study, we evaluate the feasibility and efficacy
of an EMG-based variant of our REINVENT virtual reality (VR) neurofeedback rehabilitation
system to increase volitional muscle activity while reducing unintended co-contractions. We
recruited four participants in the chronic stage of stroke recovery, all with severely restricted active
wrist movement. They completed seven 1-hour training sessions during which our head-mounted
VR system reinforced activation of the wrist extensor muscles without flexor activation. Before
and after training, participants underwent a battery of clinical and neuromuscular assessments.
We found that training improved scores on standardized clinical assessments, equivalent to those
previously reported for brain–computer interfaces. Additionally, training may have induced
13
changes in corticospinal communication, as indexed by an increase in 12–30 Hz corticomuscular
coherence and by an improved ability to maintain a constant level of wrist muscle activity. Our
data support the feasibility of using muscle–computer interfaces in severe chronic stroke, as well
as their potential to promote functional recovery and trigger neural plasticity.

2.2. Introduction
A growing body of evidence suggests that movement rehabilitation in the chronic phase of
stroke (greater than 6 months post stroke) can be effective even for those with severe impairment,
provided that the intensity and duration of therapy is much higher than what is commonly
assessed in studies of rehabilitation (Ballester et al., 2019; Lohse et al., 2014; McCabe et al.,
2015; Ward et al., 2019). Accordingly, there is a pressing need for automated or at-home training
tools that can guide therapeutic practice while motivating the required high frequency of training
(Winstein et al., 2016). However, for these systems to gain widespread application, it is critical to
understand which movement-related signals are the most practical to monitor and effective to
reinforce.
Brain–computer interfaces (BCIs) that trigger game activity upon detection of movement-
related brain signals, measured via electroencephalography (EEG), have received much attention
over the last two decades (Bai et al., 2020; Cervera et al., 2018). These systems do not require
active, volitional movement and can often just be driven by the imagination of movement, allowing
them to be of benefit to those whose severe deficits preclude direct reinforcement of overt
movement by closing the loop between the brain and the environment (Remsik et al., 2016;
Soekadar et al., 2015). However, it is known that not all individuals are able to learn how to
modulate their brain activity, and much research is needed to predict who can control a BCI, even
in the absence of a clinical condition (Alkoby et al., 2018). Further, the practicalities of at-home,
self-administered EEG remain a significant challenge (McFarland & Vaughan, 2016; Peters et al.,
2015).
14
An alternative bio-signal to reinforce is muscle activity, recorded through electromyography
(EMG). Much effort has been dedicated to developing systems and protocols that use EMG to
assist in stroke rehabilitation, for example, controlling exoskeletons (Mulas et al., 2005; Ngeo et
al., 2013) and providing feedback as a complement to traditional interventions (Armagan et al.,
2003). Research has previously shown that even those with little or no active range of motion can
often activate muscles, albeit weakly and with involuntary co-activation of antagonist muscles,
which prevents the intended movement (Beer et al., 2000; Dewald et al., 1995; Zackowski, 2004).
While EMG-based biofeedback training has often been reported to have positive effects (Arpa &
Ozcakir, 2019; Glanz et al., 1995; Kim, 2017; Schleenbaker & Mainous, 1993; Woodford & Price,
2007), relatively few studies have attempted to use EMG feedback as a way to monitor and
suppress unintended co-contractions (Mugler et al., 2019; Wright et al., 2014), which could
ultimately prevent gains in motor recovery even if muscle strength is increased. Thus, the goal of
the current system is to examine the use of EMG feedback to specifically train individuals to
reduce unintended co-contractions.
Recently, we tested the feasibility of a BCI rehabilitation system (REINVENT), which used
EEG signals to trigger the movement of a realistic virtual arm within an immersive virtual reality
(VR) environment (Vourvopoulos, Pardo, et al., 2019). The use of a realistic arm matches intent
with outcome and may engage the purported action observation network, similar to mirror therapy
(Hsieh et al., 2020; Sugg et al., 2015; Thieme et al., 2018). The EEG version of the system
(Vourvopoulos, Pardo, et al., 2019) showed promising results in terms of user satisfaction and
produced modest improvements in clinical assessments of deficits. However, for individuals who
could activate their muscles, we noted the possibility that reinforcement of even trace muscle
activity may have been as effective, if not more, compared with EEG.
Therefore, in the current pilot study, we tested the feasibility of an EMG-based variation of the
REINVENT training system and explored training-related changes in clinical presentation and
neuromuscular control. Specifically, we recruited four individuals in the chronic stage of recovery,
15
who had less than 15 degrees of voluntary wrist extension and unintended flexor-extensor co-
coactivation during attempted movement. They completed a series of seven 1-hour training
sessions during which our system reinforced extensor activation without concurrent activation of
flexors.
A battery of standardized clinical assessments was administered before and after training to
monitor generalized improvement in motor function. In addition, before and after training, we
probed changes in neural control of flexor and extensor muscles (separately) as participants
attempted to hold a steady level of muscle activation using EMG feedback and a visual target.
We quantified training related changes in task performance as well as in corticomuscular
coherence, which measures synchronization between EEG and EMG oscillations.
Corticomuscular coherence in the 12–30 Hz frequency range has been used to probe
corticospinal tract integrity and neural recovery after stroke (Krauth et al., 2019; J. Liu et al., 2019;
Rossiter et al., 2013; von Carlowitz-Ghori et al., 2014; Zheng et al., 2018).
We hypothesized that that (1) reinforcement of EMG activity in participants with severe
movement deficits would be feasible, safe, and provide a positive user experience for participants;
(2) training would produce modest improvements in clinical assessments comparable to what we
have previously observed for EEG-based neurofeedback (i.e., variable improvement across
individuals, but with some participants showing clinically meaningful effects); and (3) that we
would observe evidence of improved neuromuscular control as indexed by task performance and
by enhanced beta band (12–30 Hz) corticomuscular coherence. Finally, we expected that
participants who showed strong post-training changes in clinical assessments and neuromuscular
control would also show large improvements in task performance during training, assuming that
within-task performance is sensitive to training-induced neural plasticity.

16
2.3 Materials and Methods
2.3.1 Participants
We recruited four stroke survivors for this pilot study. Inclusion criteria required that each
participant was in the chronic phase of recovery (>6 months since stroke onset); presented with
upper extremity hemiparesis; was not taking anti-spasticity medication; and had no receptive
aphasia, significant vision loss (corrected vision was acceptable), secondary neurological
disease, or hand contractures. We specifically sought individuals with limited active wrist
extension, as control over wrist extensor muscle activity was to be trained within our intervention.
All participants gave written informed consent, and the protocol was approved by the Institutional
Review Board of the University of Southern California (reference number: HS-17-00916,
approved on 9/24/2019). Furthermore, none of the participants were receiving standard physical
therapy and all participants that were part of other exercise programs agreed to pause such
exercises for the duration of the study. We also assume that most spontaneous biological
recovery had plateaued for all of our participants because they had their stroke at least 2 years
prior to our intervention (Bernhardt et al., 2017; Cortes et al., 2017). Participant characteristics
are listed in Table 2.1.

Table 2.1 Participant demographics and baseline evaluations.
Participant Sex Age Onset (months) Paresis FMA-UE MOCA
1 Male 66 34 Left 19 23
2 Male 42 34 Right 22 17
3 Male 64 56 Left 14 22
4 Female 53 28 Left 20 22
FMA-UE, Fugl–Meyer assessment for the upper extremity; MOCA, Montreal cognitive assessment.

17
2.3.2 Study Timeline
Each participant visited the lab for ten sessions (1–2 h each) over the course of two weeks.
An outline of the study elements is shown in Figure 2.1, with each element detailed further below.
Briefly, in sessions 1 and 10, we performed clinical assessments of upper limb function, grip
strength, and wrist mobility (Figure 2.1A). We also performed an assessment of muscle control
(static hold), in which participants used feedback of their wrist EMG amplitude from flexors and
extensors (separately) to match a target level of activation for 16 trials of 4 s each (Figure 2.1B).
During this test, we also recorded EEG over the ipsilesional and contralesional motor cortices to
evaluate corticomuscular coherence. During session 2, participants were familiarized with our VR
wrist-extensor training system. During sessions 3–9, the training intervention was provided for 1
h each.

18

Figure 2.1 Experimental protocol.
(a) Timeline of the 10-session program. Sessions 1 and 10 comprised a battery of standard clinical
assessments and a test of neuromuscular control (static hold with electromyography (EMG) and
electroencephalography (EEG) recording). Session 2 allowed for familiarization of the participants with the
equipment and training paradigm. Our training program spanned seven sessions (session 3 through 9),
focused on targeted wrist extensor activation of the more affected limb. (b) Muscle control test (static hold).
In sessions 1 and 10, we tested control of wrist flexor and extensor muscles using a task in which
participants were to maintain a constant level (15% maximal voluntary contraction (MVC)) of muscle activity
for sixteen 4-second epochs using feedback of wrist muscle EMG and a target displayed on a computer
screen. During this test, task performance and coherence between EEG and EMG signals were measured.
(c) REINVENT training task. A single trial consisted of 7 s of rest and 5 s of attempted movement where,
upon success, a virtual arm pushed a beach ball off a table. A successful trial required 2 s of extensor EMG
activity that (1) exceeded 30% of the maximal activity as recorded during a power grip, and (2) was
proportionally larger than unintended flexor activity, as determined by an extension ratio that adaptively
increased or decreased depending on subject performance. A total of 120 trials were performed per
session. (d) REINVENT system. Consisting of (1) acquisition and processing computer, (2) VR headset,
and (3) EEG and (4) EMG sensors placed over the flexor carpi radialis (FCR), flexor carpi ulnaris (FCU),
extensor carpi radialis (ECR), and extensor carpi ulnaris (ECU).

19
2.3.3 Clinical Assessments (Sessions 1 and 10)
An occupational therapist performed clinical cognitive and motor assessments as part of the
pre- and post-training evaluations. These assessments included the following:
• Fugl–Meyer assessment for the upper extremity (FMA-UE). This scale measures
sensorimotor impairment of the upper limb following a hemiplegic stroke, including movement,
coordination, and reflexes, and provides a score that ranges from 0 (greatest impairment) to
66 (least impairment) (Fugl-Meyer et al., 1975).
• Action research arm test (ARAT). This scale measures functional performance of the upper
limb in terms of the ability to functionally manipulate objects with different sizes, weights, and
shapes, and provides a score that ranges from 0 (greatest impairment) to 57 (least
impairment) (Lyle, 1981).
• Montreal cognitive assessment (MOCA). This is an assessment of cognitive impairments
evaluating visuospatial abilities, memory, attention, concentration, language, and orientation,
and provides a score that ranges from 0 (greatest impairment) to 30 (least impairment)
(Nasreddine, Phillips, Bédirian, et al., 2005).
• Sixteen-question stroke impact scale (SIS-16). This assessment consists of a series of self-
reported questions evaluating quality of life as related to strength, hand function, mobility, and
activities of daily living, and provides a total score that ranges from 16 (greatest impairment)
to 80 (least impairment) (Duncan et al., 2003).
• Wrist range of motion (ROM). Using a goniometer, we recorded the maximum degrees of
passive and active wrist extension, wrist flexion, ulnar deviation, and radial deviation. Activities
of daily life usually require 40 degrees of wrist extension, 40 degrees of wrist flexion, and 40
degrees of combined ulnar and radial deviation (Ryu et al., 1991).

20
2.3.4 Additional Data Acquired
• Grip strength (GS). In each session, we recorded maximal grip force from the more affected
hand using an analog dynamometer, while recording the associated EMG.
• Simulator sickness questionnaire (SSQ). In sessions 2 and 9, we evaluated each participant’s
comfort with the VR environment using this 16-question survey covering oculomotor
discomfort, disorientation, and nausea. The total score ranges from 0 (no sickness induced)
to 63 (highest values of sickness) (Kennedy et al., 1993).
• Finally, we qualitatively evaluated the participants' overall experience and feedback in terms
of enjoyment and ease of use with a free-form questionnaire at the end of the experiment.

2.3.5 Physiological Recordings and Analysis
For all sessions, we measured surface EMG signals from four muscles at 2000 Hz using a
Delsys Trigno Wireless System (Delsys Incorporated, Natick, USA). The Delsys Trigno EMG
sensors were taped to the skin above the flexor carpi radialis (FCR), flexor carpi ulnaris (FCU),
extensor carpi radialis longus (ECR), and extensor carpi ulnaris (ECU) of the more affected limb.
The skin was cleaned with isopropyl alcohol and electrodes were positioned using double-sided
tape and wrapped with a bandage. Proper positioning was confirmed via palpation and
observation of EMG during attempted wrist extension, flexion, radial, and ulnar deviation, and
light grip. These signals were down-sampled to 1000 Hz for offline storage and analysis.
Additionally, in the first and last sessions (1 and 10), we also recorded EEG at 500 Hz over
the right and left motor cortices using a Starstim 8 System (Neuroelectrics, Barcelona, Spain).
Electrodes were positioned at frontal-central (FC3, FC4), central (C3, C4, C5, C6), and central-
parietal (CP3, CP4) scalp locations. This is the same system and electrode montage we used to
provide neurofeedback in our previous study (Vourvopoulos, Pardo, et al., 2019). Here, we use
EEG only to assess changes in corticospinal connectivity after EMG-based training. These signals
were interpolated to 1000 Hz for offline storage and analysis.
21
2.3.6 Static Hold Task: Characterization of Muscle Control during EMG Amplitude Target Tracking
(Sessions 1 and 10)
We sought to determine whether training influenced the degree to which participants could
control their wrist muscle activity, as distinct from performance during the training task. Therefore,
participants were asked to maintain a constant level of extensor EMG during attempted wrist
extension and flexor EMG during wrist flexion. For each task, the flexor or extensor muscle with
the largest signal to noise ratio during voluntary activation was chosen to provide EMG feedback.
The chosen EMG signal was smoothed and rectified in a 1-second moving window to control the
height of a feedback cursor that moved left to right across the computer screen for 10 s before
looping back. The target participants were required to reach was the 4-second hold phase
(plateau) of a trapezoid spanning 6 s (Figure 2.1B). The target’s hold phase was set to 15% of
the tracked muscle’s maximal EMG, as established during the power grip. Two 90-second trials
were completed for wrist extension and again for wrist flexion. At least 1 min of practice was
provided for each task prior to recordings, and the first trial was removed so that all analyzed
static holds were preceded by 4 seconds of rest. This resulted in 16 total holds per direction
(flexion or extension), per participant. The order of flexion versus extension for each session was
randomized. The flexion trials were executed with the hand in a supine position and the extension
trials were executed with the hand in the prone position. The task is quasi-isometric for these
participants as the required level of muscle contraction resulted in little if any overt movement of
the wrist. Participants were asked to rest their more affected arm on a pillow, and they were
provided with a low-stiffness stress-relief ball to keep their fingers from involuntarily curling
uncomfortably as they attempted the task. We discouraged participants from actively attempting
to grip the ball.
Error quantification: We quantified the accuracy with which participants could maintain a
stable level of muscle activation by calculating the median absolute deviation of the feedback
cursor from the target during the last 3 s of each hold phase.
22
Corticomuscular coherence: Along with calculation of tracking error, we measured
synchronization between EEG and EMG signals during the same time epochs. This measure,
called corticomuscular coherence, is a frequency-domain correlation where a value of 0 marks no
correlation between signals at a given frequency, and 1 indicates perfect correlation. We were
particularly interested in evaluating any changes in 12–30 Hz range (e.g., beta band) neural drive
to muscles, as this is commonly detected during static muscle contractions and used to probe the
integrity of corticospinal communication (Belardinelli et al., 2017; Fang et al., 2009; Fisher et al.,
2012; Mima et al., 2001; Norton & Gorassini, 2006; Pan et al., 2018; Rossiter et al., 2013; von
Carlowitz-Ghori et al., 2014; Zheng et al., 2018).
EEG signals were first bandpass filtered between 5 and 100 Hz using a sixth order, zero-
phase Butterworth filter, and re-referenced to the common average after removing any noisy or
bad channels, which were identified via manual inspection and using an artifact reconstruction
method within Matlab’s EEGLAB toolbox (Delorme & Makeig, 2004; Kothe & Jung, 2016).
Channels C3 and C4 were evaluated for coherence with wrist EMG signals. If one of these
channels had to be excluded, FC3 and FC4 were used instead. The choice is unlikely to have
influenced our results, as these electrodes are close to each other, and previous studies have
shown that corticomuscular coherence is not precisely localized (even to the contralateral
hemisphere) during unimanual actions in this study population (Krauth et al., 2019; Rossiter et
al., 2013). EMG signals for each epoch were bandpass filtered between 15 and 450 Hz, the
amplitude envelope extracted using the Hilbert transform (T. W. Boonstra & Breakspear, 2012;
Farina et al., 2013; Mehrkanoon et al., 2014), and the resulting signals were normalized to have
zero-mean and unit variance (S. N. Baker, 2000; Halliday & Rosenberg, 2000). For each task,
pooled coherence (Amjad et al., 1997) was calculated between the EEG (both ipsilesional and
contralesional) and EMG signals from the two active extensors or flexors. Trial epochs were first
concatenated, and then coherence was calculated using the mscohere function in Matlab,
23
specifying 512 ms Hann-windowed segments with 75% overlap. A 95% confidence level for each
coherence profile was calculated using Equation 2.1 (Carter, 1987; Rosenberg et al., 1989):
CL = 1 − 0 .05
1
L − 1

(2.1)
where L is the number of segments used to calculate coherence, adjusted for tapering, and
overlap as in (Kattla & Lowery, 2010). For a group-level analysis, we repeated the above
procedure after concatenating data from all four participants. Using this technique, the same
statistical methods can be used to evaluate group-level data as used for individual coherence
profiles (Amjad et al., 1997).

2.3.7 Wrist Extensor Training in Virtual Reality (Sessions 2 –9)
We utilized the same task in sessions 2 through 9; however, we excluded session 2 from our
analyses as it was a familiarization session during which participants were allowed to ask
questions during the task. Our training paradigm (sessions 3–9) utilized an Oculus Rift CV1
(Facebook Technologies, Menlo Park, USA) head-mounted display system with 3D-audio
headphones. We used the lab streaming layer (LSL) protocol (Swartz Center for Computational
Neuroscience, 2022) for data synchronization between acquisition and feedback systems. The
VR task was programmed in the Unity game engine (v2017.4.1, Unity Technologies, San
Francisco, USA) and rendered with the Oculus SDK, as per our previous work (Vourvopoulos,
Pardo, et al., 2019). EMG signals were processed and analyzed online with custom scripts in
Matlab (R2014a, The Mathworks, Natick, USA).
Participants were shown a visualization of their two arms resting on a table, as shown in Figure
2.1C. The virtual arms were chosen from a set of models to best match each individual’s physical
characteristics. Each trial consisted of 7 s of rest and a 5-second movement attempt window
during which participants were required to maintain a threshold of EMG activation for 2 s for the
virtual hand to push the ball off the table. EMG feedback was given in real-time during the
24
movement attempt window, such that a decrease in the EMG signal led to the hand moving
towards the start position and an increase led to the hand moving towards the end goal. This was
repeated for six blocks of 20 trials for each session, lasting approximately 1 hour. In this training
paradigm, we sought to reinforce wrist extensor activation without simultaneous flexor activation.
As mentioned previously, this is advantageous over simple EMG activation as a feedback signal
because unintended co-contraction of antagonists may be just as, if not more, detrimental to
voluntary movement than total paralysis (Ellis et al., 2017; Zackowski, 2004), and we found this
to be the case with our study participants as well. We thus calculated an extensor ratio (ER),
shown in Equation 2.2:
ER =
E M G
e x t e n s o r s
E M G
e x t e n s o r s
+ E M G
f l e x o r s
(2.2)
as the sum of the extensor activity divided by the sum of the extensors and flexors. This
calculation was made at 250 Hz, using a 480 ms moving window of EMG. Each signal was
rectified, averaged, and normalized to the maximal activity recorded during a maximal power grip.
For a successful trial, two thresholds had to be exceeded. First, the summed extensor activity
was required to exceed 30% of its maximal level during the power grip, and second, an adaptive
ER threshold had to be exceeded. The ER threshold was initially set to 0.5 (equal flexor and
extensor activity) at the beginning of each session and would increase or decrease in increments
of 0.3 (within the range of 0.3 to 0.97) if the previous three trials were all successes or failures,
respectively. This was implemented to comply with current recommendations that training tasks
should be challenging and progress in difficulty to encourage continuous and adaptive learning
(Winstein et al., 2008). Participants were to rest their more affected arm on a pillow, and they
were provided with a low-stiffness stress-relief ball to keep their fingers from involuntarily curling
uncomfortably as they attempted the task. We discouraged participants from actively attempting
to grip the ball. Importantly, such a strategy would lead to task failure because this grip would
result in co-contraction of flexors and extensors, leading to a low ER.
25
2.3.8 Statistical Analyses
We utilized custom scripts in Matlab (R2019a, The Mathworks, Natick, USA) and R (R
Foundation for Statistical Computing, Vienna, Austria) for offline signal processing and statistical
analyses, respectively.
2.3.8.1 Behavioral and Neuromuscular Changes Following Training
Clinical assessments: We performed paired t-tests (N = 4) to identify the strongest and most
consistent changes in FMA-UE, ARAT, SIS-16, SSQ, ROM, and grip strength. These group-level
tests were considered separate, and thus significant at the p < 0.05 level without correction. In
addition, owing to the small number of subjects, we also report non-significant trends (p < 0.1) as
an exploratory analysis.
Static hold (EMG tracking performance): The mean and standard deviation of tracking error
were calculated across the 16 attempts for each participant before and after the training
intervention. Group-level effects were identified using a paired t-test to compare pre- versus post-
training error across individuals. Additionally, a paired t-test was used to evaluate the significance
of any training-induced changes in performance at the individual level. The significance level for
each individual was set to p < 0.05.
Corticomuscular coherence during tracking: We evaluated changes in corticomuscular
coherence at the group level, with a Z-score difference of coherence (Rosenberg et al., 1989)
(pre versus post) at each frequency of the group-level coherence profiles constructed for each
task (flexion or extension) and hemisphere (ipsilesional and contralesional) using the formula
given in Equation 2.3:
Z
di f f
=
FZ
po s t
− FZ
pr e
√ 1 / L

(2.3)
where FZ is the Fisher-transformed coherence value (i.e., atanh (sqrt (Coh)) and L represents the
degrees of freedom, calculated as described for Equation 1. This provides a standard Z score for
the difference in coherence between sessions 1 and 10, for every frequency. Then, we created a
26
composite Z-score for the 12–30 Hz beta band using Stouffer’s Z-score method (Kilner et al.,
1999). We chose the lower bound (12 Hz) for better comparability with a previous study (Krauth
et al., 2019), in which beta band corticomuscular coherence was found to differ according to
recovery status. Composite Z-scores with an absolute value above 1.96 are considered significant
at the 5% confidence level. For completeness, two other bands, alpha (8–12) and gamma (30–
50), were tested as well.
2.3.8.2 Changes Across Training Sessions
Performance: We calculated the average within-game ER-threshold and % successful trials
to track changes in task performance over training sessions for each participant. A Spearman’s
rank correlation was used to determine if these measures increased or decreased consistently
across days, and paired t-tests of the first versus last session were used to test for any consistent
group-level effects.
EMG during training: The same procedure was utilized for three metrics of EMG activity (flexor
amplitude, extensor amplitude, and ER). For the analysis of flexor and extensor amplitude
changes over sessions, we normalized the mean rectified voltage for each EMG signal by the
mean and standard deviation of those recorded in session 3 (Z-score). This normalization allows
for better comparability between participants. Although the method is susceptible to small
variations in electrode placement and signal quality across sessions, it avoids dependence on
EMG activity during maximal voluntary effort. Furthermore, for our study population, this maximal
voluntary effort does not represent the full capacity of the muscles, and thus cannot be assumed
to be stable across sessions. We assessed changes over time as before, using Spearman’s rank
correlation for individuals, and a paired t-test for group-level change.
EMG during maximal power grip: Finally, we assessed grip strength and EMG activity
recorded during the daily maximal power grips. We calculated the amplitude of each EMG signal
as well as the ER at maximal grip force. Similarly, we assessed changes over time using
Spearman’s rank correlation for individuals and a paired t-test for group-level change.
27
2.4 Results
2.4.1 Feasibility
Participants reported minor levels of discomfort after training with the head-mounted display,
assessed with the simulator sickness questionnaire, during both the first and last training sessions
(first: mean = 5.56, SD = 5.27; last: mean = 6.35, SD = 2.59). Changes in such discomfort were
not significant when comparing both sessions (t = 0.32, p = 0.76). Qualitatively, all participants
liked the virtual environment and commented on enjoying the experience. They also reported
using different strategies to improve their control on the task such as concentrating on different
parts of their hands, imagining the movement, or focusing on muscle sensations. All reported that
they would be enthusiastic to use a portable home system if one were available. Although wrist
mobility was severely limited in all participants, all participants were able to activate their muscles
enough to generate detectable activity during EMG-based training and assessments. This is
particularly relevant because participants with this level of severity would often be considered
primarily suitable for EEG-based neurofeedback under the assumption that no direct reinforceable
motor commands exist.

2.4.2 Behavioral Changes Following Training
At the group level, only the SIS-16 showed significant improvements (t = 5.67, p = 0.011;
Figure 2.2). We found non-significant trends in that ARAT (t = 2.61, p = 0.079) and FMA-UE (t =
2.43, p = 0.093). The range of active wrist extension improved for three participants, but the effect
was quite variable across individuals (t = 2.27, p = 0.108). In addition, three participants that
improved their FMA score, with two (participants 2 and 3) showing an improvement meeting the
minimal clinically important difference (MCID) criteria of 4.25–7.25 points (Page et al., 2012).
None of the other measures showed significant group level changes, as shown in Table 2.2.
28

Figure 2.2 Clinical assessments.
Assessments for each participant, as evaluated in sessions 1 and 10. * indicates group-level significance
of p < 0.05. † indicates a group-level nonsignificant trend at p < 0.1. FMA-UE, Fugl–Meyer assessment for
the upper extremity; ARAT, action research arm test; SIS, stroke impact scale.

Table 2.2 Statistical comparisons for clinical assessments.
Assessment t p Pre Post
ARAT 2.61 0.079 5.75 (8.85) 7 (9.38)
Extension 2.27 0.108 6.75 (7.81) 10.5 (8.19)
FMA-UE 2.43 0.093 18.75 (3.40) 23.25 (4.19)
Grip More Imp. 1.25 0.299 8.67 (6.99) 9.44 (5.78)
SIS-16 5.67 0.011* 58.5 (10.08) 62.75 (9.29)
Group level paired t-test results for each assessment, as well as the mean (SD) values for each in sessions
1 and 10 (pre and post, respectively). Bold font represents nonsignificant trends (p < 0.1) and * represents
significant changes (p < 0.05). FMA-UE, Fugl–Meyer assessment for the upper extremity; ARAT, action
research arm test; SIS, stroke impact scale.
ARAT
(0 –57)
Extension
(Deg.)
FMA-UE
(0 –66)
Grip More Imp.
(Kg)
SIS-16
(1 6–80 )
Subject
Evaluation
*
† †
0
5
10
15
0
5
10
15
20
15
20
25
50
60
70
Pre Post
5
10
15
1
2
3
4
29
2.4.3 Changes of Muscle Control during EMG Amplitude Target Tracking
Trends of improved motor control, as measured by a reduction in the median absolute
deviation from the target, were seen in three of four participants after training, for both flexion and
extension tasks (extension: t = -5.69, p < 0.001; t = -2.04, p = 0.06; and t = -3.60, p = 0.003 for
participants 1, 2, and 4, respectively; flexion: t = -5.09, p < 0.001; t = -2.67, p = 0.018; and t = -
2.02, p = 0.061 for participants 1, 3, and 4, respectively). All three statistically significant individual
tests (p < 0.05) were also significant at the Bonferroni-corrected 95% confidence level for eight
tests (p < 0.0063), which argues against the possibility that these results are the incidental
outcome of making eight statistical tests. However, a paired t-test at the group level showed
nonsignificant p-values of 0.21 and 0.22 for extension and flexion tracking error, respectively.
Figure 2.3 shows the tracking error before and after training for each participant.

30

Figure 2.3 Performance changes in EMG amplitude control.
Individual changes (1–4, left to right) in tracking performance (median absolute deviation (MAD) error)
during maintenance of a constant level of extensor (top) or flexor (bottom) activity before and after seven
sessions of wrist extensor training. Note that the training itself did not require maintenance of a constant
level of EMG, nor was direct EMG feedback provided. Improvements were seen in both tasks after training.
* indicates a significant change at the 95% confidence level. † indicates a nonsignificant trend at p < 0.1.
Bar heights display the mean error and error bars display +/-1 standard deviation.

2.4.4 Neuromuscular Changes Following Training
Consistent, significant corticomuscular coherence was observed only during static holding of
wrist extension and not during flexion. Pooled coherence across participants shows that, during
maintained wrist extension, the only significant coherence occurred within the beta band (12–30
Hz) and only after training (Figure 2.4A, top, ipsilesional: Z-score pre–post difference = 2.57, p =
0.010; contralesional: Z-score pre–post difference = 3.29, p = 0.001). After training, the composite
difference of coherence, that is, the averaged coherence within each frequency band, was
Evaluation
Flexion
Error MAD (% MVC)
Extension
Error MAD (% MVC)
* *
S1 S2
* *
S3 S4
†
1
†
0.0
2.5
5.0
7.5
0.0
2.5
5.0
7.5
Pre Post
31
evaluated (Figure 2.4A bottom, horizontal lines at Z-score = 1.96 indicate the threshold for
significance at p < 0.05 level), and showed a significant effect of training only for the beta band.
This change in beta-band corticomuscular coherence was significant at the group level for both
ipsilesional and contralesional EEG. It was also specific to the wrist extension task; there was
little or no coherence between EEG measured from either hemisphere and flexor EMG during the
flexion task (Figure 2.4B). Individual coherence profiles showed similar effects, with significant
peaks in the beta band in bilateral hemispheres post training for wrist extension, and no consistent
effects for other conditions (Figure 2.4C and 2.4D). The magnitude and precise frequencies were
variable, but it is clear that the group-level effects could not have occurred without consistency
across subjects. For example, subject 4 showed a large peak in the gamma band (around 40 Hz)
in the post-training contralesional recordings (Figure 2.4C), and yet this is not reflected in the
pooled coherence (Figure 2.4A).

32

Figure 2.4 Corticomuscular coherence during static flexion and extension.
(a) Group-level pooled coherence of ipsilesional and contralesional EEG to wrist extensor muscles during
extension task. Horizontal lines indicate the 95% confidence level. Bar plots below the spectra represent
the composite difference of coherence before vs. after training within three frequency bands. Beta-band
coherence was increased significantly and bilaterally after training. (b) Coherence during the flexion task
was generally not present either before or after training. Panels (c) and (d) show the individual coherence
spectra for each participant (1–4, left to right). Coherence within 0 and 60 Hz is shown in all plots, including
gray vertical dashed lines indicating the beta band (12–30 Hz).

33
2.4.5 Changes Across Training Sessions
At a group level, we found no significant changes in game performance or muscle activity
during training, although the ER did show a non-significant trend (t = 2.58, p = 0.08). Table 2.3
contains group-level data for training-related muscle activity, game performance, and grip
strength. At the individual level (Figure 2.5), measures of muscle activity were variable, but three
of four participants showed increasing extensor activity over time, significant for participant 4
across sessions (Spearman’s rho = 0.86, p = 0.024). Similarly, participant 3 showed a significant
increase in ER (rho = 0.86, p = 0.024). However, no consistent or significant improvements were
found in the proportion of successful trials or increases in the ER threshold.

Figure 2.5 Muscle activity and performance during training.
Each of the four columns represents the per-session average of one measure (rows) recorded during
training. Top row: extensor EMG activity, normalized to the first session (Z-score). Middle row: extensor
ratio (ER), or ratio of extensor activation relative to total muscle activity, also normalized to the first session.
Bottom row: game performance as the percentage of trials in which participants exceeded the ER threshold,
as well as a minimal activation threshold set to 30% of maximal wrist extension. Best fit lines are included
to visualize trends across sessions. Red lines indicate that a Spearman rank correlation calculated between
session number and the measure of interest, for that participant, exceeded the 95% confidence threshold.
-2
0
2
4
6
-2
0
2
4
6
0
25
50
75
100
4 6 8 3 5 7 9 4 6 8 3 5 7 9 4 6 8 3 5 7 9 4 6 8 3 5 7 9
S1 S2 S3 S4
Extensors (Z) ER (Z) % Success
Session
Muscle Activity and Performance During Training
34
Table 2.3 Group level analysis of within-game performance and grip strength.
Activity t p Pre Post
ER 2.58 0.082 8.94x10
-16
(1.65x10-15) 0.80 (0.62)
Extensors 1.81 0.168 -1.99x10
-16
(1.32x10
-15
) 1.41 (1.56)
Grip 1.53 0.224 7.67 (6.39) 10.13 (4)
Flexors 0.91 0.431 -4.71x10
-16
(7.37x10
-16
) 0.51 (1.11)
Success -0.40 0.719 57.29 (7.65) 55.83 (5.57)
Threshold -0.03 0.981 36.82 (19.35) 36.44 (21.44)
Bold font represents nonsignificant trends (p < 0.1). ER, extensor ratio.

2.5 Discussion
2.5.1 Summary
In this pilot study, we explored the use of EMG feedback within a VR-based rehabilitation
program targeting individuals with chronic, severe movement deficits from stroke. We found that
seven 1-hour training sessions in which participants attempted to activate wrist extensor muscles
without coactivation of flexors was both feasible in this study population and provided an
acceptable overall user experience, with minimal discomfort stemming from the VR environment
or the required task. Despite the relative brevity of the training program, we observed notable
improvements in standard clinical assessments, the strongest being an improvement in the SIS-
16 quality of life measure. We also saw non-significant trends towards improvements in the FMA-
UE and the ARAT. Notably, as observed in previous studies, these improvements were highly
variable across individuals, but they generally matched or exceeded what has been found using
EEG-based neurofeedback. In addition, after training, participants significantly improved their
ability to maintain a constant level of wrist flexor and extensor activation and, importantly, showed
enhanced 12–30 Hz corticomuscular coherence, which we interpret as evidence of neural
reorganization associated with functional recovery. Participants did not improve their performance
during the training task itself, suggesting that post-training improvements in our various
35
assessments were not owing to learned, task-specific behaviors, but instead, a more general
influence of movement training on neural recovery.

2.5.2 Feasibility and Acceptability
Our first hypothesis was that an EMG-based version of the REINVENT system would be
feasible and acceptable to users. Although we assumed that EMG would be detectable in anyone
without severe flaccid paralysis, the utilization of very weak EMG within an entirely EMG-based
training program has not been commonly reported for our targeted study population, leaving the
feasibility of this approach unclear. One pilot study (Wright et al., 2014) and a later follow-up
(Mugler et al., 2019) demonstrated that a muscle–computer interface could be used to uncouple
pairs of shoulder/elbow muscles. However, our task focused specifically on promoting
individuated wrist extensor activity in participants with little to no functional wrist movement.
Accordingly, our task may have been more difficult and would have produced weaker EMG
signals.
We were able to detect useable EMG signals in all participants and confirmed that it was
feasible to use these signals to drive our training paradigm. However, the use of EMG biofeedback
in participants with such severe motor deficits has several important caveats that merit attention.
First, raw EMG voltage varies across individuals and recording methods, and thus EMG
amplitudes are typically normalized to a maximal level of activation determined for each individual.
Our target population could not maximally activate their muscles to provide an unambiguous
normalization. Furthermore, none of our participants could produce a muscle-specific ‘maximal
effort’; as such, efforts always produced some degree of co-activation across all forearm muscles.
We thus normalized our EMG using a power grip, which produces a consistent co-activation of all
muscles, and thus reduced the dependence of our training system on the use of precise units of
EMG amplitudes. This method also mitigates potential problems stemming from electrical cross-
talk between nearby active muscles, as well as the complex association between EMG amplitude
36
and voluntary effort, that is, when the raw EMG signal is dominated by the recruitment or de-
recruitment of a small number of motor unit action potentials, as can be the case at very low levels
of activation.
Moreover, we implemented an adaptive threshold to define success within our training
paradigm (the extensor ratio, ER). This method provides an adaptive level of challenge, which is
desirable for any personalizable rehabilitation program. However, as with any adaptive threshold,
there is a need to balance delivery of progressive challenge with the practical requirement of
maintaining participant engagement and motivation. This can be difficult to predict a priori. We
found that all of our participants remained engaged with the task, and despite average success
rates of just over 50% per session, none reported that the task was too difficult or frustrating.
Anecdotally, most considered each session akin to a workout and felt it natural that the system
would alter difficulty contingent on performance. Simulator sickness and general discomfort with
using the VR headset were minimal, and most suggestions for improvement were directed at the
monotony of the task, or specific preferences regarding the visual environment.
Overall, participants rated the training experience as positive, and despite several potential
difficulties related to the use of EMG to control the virtual arm, our EMG-based variant of
REINVENT was found to be feasible in all participants.

2.5.3 Clinical Assessments
We had initially hypothesized that, after training, we would see improvement in clinical metrics
at least on par with what we had found in our EEG-based REINVENT training program
(Vourvopoulos, Pardo, et al., 2019). This was partly based on our previous finding that three of
four participants might have been more successful had EMG been used to trigger movement of
the virtual arm rather than EEG. In that study, group-level effects were not statistically significant,
but three of four participants showed improved FMA-UE scores, and one improved by 6 points,
which meets the criteria for a clinically meaningful change (4.25–7.25 points (Page et al., 2012)).
37
In the present study, overall improvements may have been even stronger, with a significant group-
level effect for the SIS-16 and positive, although nonsignificant, trends (p < 0.1) exhibited for
ARAT and FMA-UE. For the FMA-UE, we again found improvement in three of four participants,
but now two showed clinically meaningful differences (7 and 8 points). In fact, on average, there
was a 4.5 point improvement in FMA-UE, which even compares well to longer training
interventions (Bai et al., 2020).
Mugler et. al. found that six sessions of EMG-based training to reduce co-contraction was
able to improve FMA-UE scores of 32 moderate-to-severe stroke survivors by an average of about
3 points (Mugler et al., 2019). This is the closest in design to the current study and confirms that
this type of training can be of benefit in a larger population. While there is no direct EEG-only
paradigm for training co-contraction, it is worth noting that many of these studies involve much
higher doses. For example, the study by Bundy et. al. used 37 to 72 sessions to improve ARAT
scores by 10 points on average (Bundy et al., 2017). Likewise, our training paradigm may benefit
from continued use beyond seven sessions.
Because our training task was specific to wrist extension, improvements in ARAT, FMA-UE,
or SIS-16 may indicate a broad, generalized effect of training. That is not to say there were no
specific effects on the wrist muscles, as there was improvement in the range of active wrist
extension for several individuals, but it is certain that these changes in wrist function do not explain
changes in clinical metrics. Concerning range of motion, it should be noted that the standard test
of active extension begins from a neutral wrist angle, and thus small gains in muscle strength or
individuation may have been more apparent had we tested the ability to lift the wrist from a
relaxed/hanging position, where the general tone of antagonist flexor muscles may have been
lower at the initiation of extensor activation.

38
2.5.4 Neuromuscular Control
In addition to measuring training-induced changes in standard clinical assessments, we
included a test intended to determine whether wrist extensor training within the REINVENT
system produced generalizable changes in the neural control of either wrist extensor or flexor
muscles. To disambiguate general effects from task-specific ones, we performed our analysis of
neuromuscular control using a task that required fundamentally different behavior compared with
our training paradigm. Specifically, we asked participants to maintain steady activation of a wrist
extensor or flexor muscle given feedback of their EMG amplitude and a target. This allowed two
separate (but related (Kristeva et al., 2007)) evaluations of neural control, one focused on motor
function and the other on neural plasticity.
Our functional assessment was simply task performance. We had initially expected
improvement after training, and indeed, three of four participants became more capable of
maintaining a steady level of either flexor or extensor activity, as indexed by the median absolute
deviation of their feedback cursor from the target. Because we found improved performance for
both flexors and extensors, we interpret this finding to imply that training had a generalizable
influence on voluntary control of muscles.
Given that neurorehabilitation is intended to produce useful changes in neural function, we
also assessed corticomuscular coherence during the same static holding task. Corticomuscular
coherence quantifies synchrony between EEG and EMG activity in the frequency domain, and
high coherence is interpreted as a clear indication of functional communication and connectivity
between the cortex and the motor neurons that innervate a given muscle (T. Boonstra, 2013;
Farmer, 1998; J. Liu et al., 2019; Mima & Hallett, 1999). It is worth noting that previous literature
has shown inter-subject variability when measuring corticomuscular coherence, even in the
absence of a clinical condition (Chwodhury et al., 2015). Nevertheless, corticomuscular
coherence in the beta band (12–30 Hz) has often been used to probe the functional integrity of
corticospinal communication following stroke (Fang et al., 2009; Guo et al., 2020; Krauth et al.,
39
2019; Larsen et al., 2017; Mima et al., 2001; Pan et al., 2018; Rossiter et al., 2013; von Carlowitz-
Ghori et al., 2014; Zheng et al., 2018). Initially after stroke, corticomuscular coherence is reduced
(Fang et al., 2009; Larsen et al., 2017; Mima et al., 2001), but increases either with natural
recovery (Krauth et al., 2019; Rossiter et al., 2013; von Carlowitz-Ghori et al., 2014) or recovery
due to specific rehabilitation efforts (Pan et al., 2018; Zheng et al., 2018). Furthermore, changes
in beta-band corticomuscular coherence correlate positively with corticospinal excitability and
inversely with gamma-aminobutyric acid (GABA) mediated cortical inhibition (M. R. Baker &
Baker, 2003; J. Liu et al., 2019; Matsuya et al., 2017; Power et al., 2006). For stroke survivors in
the chronic stage of recovery, the location of maximal EEG–EMG coherence on the scalp is not
necessarily over the primary motor cortex contralateral to the active limb, but instead, can be
located over a broad area including bilateral supplementary and premotor areas, and may even
be strongest on the contralesional hemisphere during activation of muscles on the more affected
limb (Braun et al., 2007; Krauth et al., 2019; Rossiter et al., 2013).
Our findings strongly suggest that training did impact corticospinal communication in our
participants. First, we found that training produced frequency-specific effects on corticomuscular
coherence. Only the beta band showed consistent, significant coherence across participants, and
only after training. The change in coherence between session 1 and 10 was even statistically
significant at the group level, suggesting that the post-stroke recovery of corticomuscular
coherence (Krauth et al., 2019; Rossiter et al., 2013; Zheng et al., 2018) can be induced, even at
the chronic phase, by a short-term behavioral intervention. Further, our data emphasize that
rehabilitation paradigms targeting neural plasticity do not necessarily require detection or
reinforcement of movement-related EEG oscillations. EMG may provide a sufficient (and natural)
index of motor circuit operation when it can be detected.
Interestingly, training-induced enhancement of corticomuscular coherence was only observed
when participants were asked to maintain wrist extensor activation. There were no changes in
coherence (and more generally, no coherence) during the flexion task. This may relate to the fact
40
that the trained task was designed to produce voluntary control over the extensors (not the
flexors), and/or because wrist extensors may receive a greater proportion of monosynaptic
corticospinal projections compared with the wrist flexor muscles (de Noordhout et al., 1999;
Palmer & Ashby, 1992). Because changes in reticulospinal pathways may ultimately cause
pathological synergistic muscle activation after stroke (Li et al., 2019; McPherson et al., 2018;
Owen et al., 2017), it may be that there is a greater contribution of reticulospinal drive to wrist
flexion compared with extension movements (Andrade & Andrade, 2012). Accordingly, if one aim
of our training task was to reduce unintended flexor activation, then the reinforced neural activity
might have been a reduction of brainstem or reticulospinal output to the flexors. At the same time,
our task may have positively reinforced direct corticospinal activation of the extensors, which,
unlike an alteration in brainstem output, would have been reflected as a change in corticomuscular
coherence.

2.5.5 Training Effects versus Task Performance
We had initially expected to find some association between task performance and the extent
of improvement measured by our post-training assessments. This was premised on the
assumption that task success might itself promote recovery, and further, that neural recovery
induced by a particular task should improve performance of that specific task. Neither assumption
is supported by our results. We found clear evidence of training-induced changes in clinical
metrics and indices of neural recovery that would not be expected to have occurred
spontaneously in this population, or by virtue of a given assessment having been previously tested
two weeks earlier. At the same time, we found no consistent improvement in the performance of
the training task across seven sessions. The training did, however, apparently promote some
form of generalized recovery. Potentially, the training-induced changes that we observed reflect
the early stages of a recovery process that would only later gain the strength to generate
meaningful improvements in task performance. For example, improved utilization of the
41
corticospinal tract may have occurred thanks to the constant movement attempts, and led to
modest improvements in certain clinical assessments. However, given that our task training also
demanded inhibition of involuntary flexor activity, this may have required a different (e.g.,
reticulospinal) circuit that did not, or had not yet, responded to training. Overall, our results
emphasize that performance improvement within a rehabilitation paradigm is not necessarily a
prerequisite of training-induced neural plasticity, nor potential improvements in clinical function.

2.5.6 Limitations and Conclusions
As with any pilot study, it is important to emphasize that our sample size does not allow us to
generalize our findings to the larger population of individuals suffering from movement deficits
due to stroke. Our results, while promising, must be interpreted with the understanding that both
symptom expression and the progression of recovery can be highly variable across individuals.
Even so, our results do suggest that EMG-based rehabilitation using the REINVENT system is
feasible, and at least in some cases, can promote measurable improvements in clinical metrics
and measures of neuromuscular control. Further, our findings emphasize that changes in neural
activity, clinical outcomes, and performance of a training task are likely to occur on different time
scales and to different degrees for different individuals. Understanding the relationships between
these metrics will be critical for uncovering the mechanisms of neural recovery and optimizing
rehabilitation protocols; however, studies with a larger sample size are required. Furthermore,
EMG has potential for many extensions and variations that could provide more realistic
interactions, like predicting body movements and their related forces and torques (Andrade &
Andrade, 2012; Ngeo et al., 2013). However, more research is required before introducing these
mechanisms because, for example, the assumed linear relationships between EMG and force are
lost after a stroke (Suresh et al., 2015). Importantly, another area worth exploring would be
merging different feedback modalities, similar to the works of (Kawase et al., 2017) and (Leeb et
al., 2011). Although our present study focuses on exclusive utilization of EMG for reasons of
42
practicality, it would be of utmost interest to investigate whether, when EEG is feasible, hybrid
systems could further improve behavioral outcomes beyond exclusive EEG- or EMG-based
feedback. Ultimately, personalized training may require systematic characterization of many
interacting factors, including measures of spasticity that could help to disambiguate muscle co-
contraction and synergistic neural drive, for example, modified Ashworth scale, pathological
synergies, muscle strength, attention, motivation, and magnetic resonance imaging scans to
investigate the structural and functional integrity of distinct neural circuits.
While such a high degree of personalization is not yet practical, adaptive rehabilitation
systems that can be self-administered and taken home are within reach. Our study demonstrates
that relatively simple measures of muscle activation, in combination with commercial virtual reality
equipment, may be sufficient to promote recovery even in very severely impaired individuals.
Although utilization of muscle–computer interfaces for stroke rehabilitation is rare, it seems from
our study that this method is not only feasible, but capable of producing neural changes and
improvement on clinical evaluations. However, achieving significant outcomes may require
interventions of a higher dose and with weekly behavioral measurements to capture the evolution
of recovery. Our study, along with other recent efforts (Mugler et al., 2019), supports the idea that,
given the existence of myographic activity, muscle–computer interfaces are feasible and may be
especially practical in terms of cost, simplicity, and eventual application outside of a clinical or
laboratory setting.

43
Chapter 3: Development of a Low-Cost, Modular Muscle –Computer
Interface for At-Home Telerehabilitation for Chronic Stroke
This section is adapted from:
Marin-Pardo, O., Phanord, C., Donnelly, M. R., Laine, C. M., & Liew, S.-L. (2021).
Development of a Low-Cost, Modular Muscle–Computer Interface for At-Home Telerehabilitation
for Chronic Stroke. Sensors.

3.1 Abstract
Stroke is a leading cause of long-term disability in the United States. Recent studies have
shown that high doses of repeated task-specific practice can be effective at improving upper-limb
function at the chronic stage. Providing at-home telerehabilitation services with therapist
supervision may allow higher dose interventions targeted to this population. Additionally, muscle
biofeedback to train patients to avoid unwanted simultaneous activation of antagonist muscles
(co-contractions) may be incorporated into telerehabilitation technologies to improve motor
control. Here, we pre-sent the development and feasibility of a low-cost, portable,
telerehabilitation biofeedback system called Tele-REINVENT. We describe our modular
electromyography acquisition, processing, and feedback algorithms to train differentiated muscle
control during at-home therapist-guided sessions. Additionally, we evaluated the performance of
low-cost sensors for our training task with two healthy individuals. Finally, we present the results
of a case study with a stroke survivor who used the system for 40 sessions over 10 weeks of
training. In line with our previous research, our results suggest that using low-cost sensors
provides similar results to those using research-grade sensors for low forces during an isometric
task. Our preliminary case study data with one patient with stroke also suggest that our system is
feasible, safe, and enjoyable to use during 10 weeks of biofeedback training, and that
improvements in differentiated muscle activity during volitional movement attempt may be induced
44
during a 10-week period. Our data provide support for using low-cost technology for individuated
muscle training to reduce unintended coactivation during supervised and unsupervised home-
based telerehabilitation for clinical populations, and suggest this approach is safe and feasible.
Future work with larger study populations may expand on the development of meaningful and
personalized chronic stroke re-habilitation.

3.2 Introduction
Stroke is a leading cause of long-term disability in the United States with almost 800,000
people experiencing a new or recurrent stroke each year (Virani et al., 2020). While motor
recovery was thought to plateau by the chronic stage after stroke (more than 6 months after the
vascular incident), more recent studies have shown that improvement of upper limb function is
possible at the chronic stage (Ballester et al., 2019; Ward et al., 2019). Recent research suggests
that high dose interventions of repeated task-specific practice are effective at inducing significant
positive outcomes in this population (Lohse et al., 2014; Stinear et al., 2020; Veerbeek et al.,
2014; Ward et al., 2019). However, due to the time and physical constraints of many therapy
sessions, common in-clinic interventions only provide on average 32 repetitions of functional
upper extremity movements per session (Lang et al., 2009). Providing at-home telerehabilitation
services with therapist supervision and input is a potential solution to allow clinicians to deliver
quality, higher dose interventions. Recent studies suggest that telerehabilitation for stroke
rehabilitation is feasible and as effective as in-person therapy (Cramer et al., 2019; Dodakian et
al., 2017).
One effective method for improving upper limb function, which could be combined with
telerehabilitation, is the reinforcement of muscle activity using electromyography (EMG)
biofeedback. Muscle biofeedback has been shown to reduce spasticity and improve post-stroke
arm function, motor control, muscle activity, and strength (Armagan et al., 2003; Doğan-Aslan et
al., 2012). Specifically, previous research has shown that biofeedback training to avoid
45
unintended simultaneous activation of antagonist muscle groups may be particularly beneficial for
reducing un-necessary co-contractions that impede functional motor control (Mugler et al., 2019;
Wright et al., 2014). However, further research is required to address remaining fundamental
questions in biofeedback investigations—for example, what is the required intensity and dosage
to significantly improve long-term outcomes?
Recently, portable systems have been developed for at-home use to improve accessibility
and training time with EMG biofeedback (Donoso Brown et al., 2014; Mugler et al., 2019).
However, proper implementation of home-based EMG biofeedback is critical to prevent low
participant adherence, avoid high costs, and account for limitations in terms of required physical
space, time, and technical literacy (Chen et al., 2019; Donoso Brown et al., 2014; Feldner et al.,
2019, 2020; Hochstenbach-Waelen & Seelen, 2012). Specifically, it has been suggested that the
ability to track patient progress in real-time and the continued involvement of a clinician in the
intervention are key factors that could improve patient motivation and adherence to at-home
rehabilitation programs (Cramer et al., 2019; Feldner et al., 2019).
To address these needs, we developed and tested a low-cost, portable telerehabilitation
biofeedback system called Tele-REINVENT. Tele-REINVENT builds upon our previous work in
which we developed and tested a system (REINVENT) that could provide biofeedback of brain or
muscle activity on a computer screen or in immersive virtual reality (VR) with a head-mounted
display (HMD) in a laboratory or clinic setting (Marin-Pardo et al., 2020; Vourvopoulos, Pardo, et
al., 2019). Tele-REINVENT uses the same modular platform and incorporates a telerehabilitation
component for live video and audio conferencing with a clinician who meets regularly with the
participant to monitor progress. The clinician can also provide technical support, ensure the
electrodes are placed correctly, and monitor EMG signals in real-time to ensure adequate signal
quality. In addition, Tele-REINVENT uses a portable laptop with commercial low-cost EMG
sensors for greater affordability and accessibility in the home environment. Lastly, Tele-
REINVENT has new gamified elements to encourage greater engagement, motivation, and
46
adherence to a home-based program. Overall, we aimed to incorporate benefits from both
literature on telerehabilitation and EMG biofeedback into our current system.
In this paper, we present a detailed description of the Tele-REINVENT system. Additionally,
we provide a validation example with two healthy individuals showing that, for our purposes, the
performance of the low-cost sensors can be considered com-parable to that of research-grade
sensors. Finally, we present the feasibility and results of a case study with a chronic stroke
survivor who tested the system for 40 sessions over 10 weeks.

3.3 Materials and Methods
3.3.1 Participants
To confirm that low-cost EMG sensors produced valid and appropriate measurements, for our
purposes, as compared with research-grade equipment, we compared measurements collected
from 2 healthy male right-handed participants (ages 28 and 37 years old) using both systems.
For the case study, we recruited a 67-year-old male stroke survivor, 11 years after stroke
onset, to test our developed system for 10 weeks. The participant had upper extremity
hemiparesis, was not taking anti-spasticity medication, had no receptive aphasia, had corrected
vision, and did not have a secondary neurological disease. The participant had less than 15
degrees of active wrist or finger extension in the more impaired hand and was unable to grasp
and release a ball unassisted.
Protocols were approved by the University of Southern California Health Sciences Campus
Institutional Review Board (IRB) and written informed consent was obtained from all participants
in accordance with the Declaration of Helsinki.

3.3.2 System Architecture
Tele-REINVENT is a portable stroke telerehabilitation system for at-home, therapist-driven,
personalized, and gamified training. We designed the system to acquire reliable EMG signals and
47
provide realistic feedback as game control. We designed and developed hardware and software
that allowed us to acquire, process, and store the participant’s EMG signals on a laptop computer
(Razer Blade RZ09-01953; Operating System: Windows, Processor: Intel Core i7 7700, RAM: 16
GB; Razer Inc., Irvine, CA, USA). Furthermore, the architecture of the system allows for remote
update, control, and data retrieval. Figure 3.1a shows the overall architecture of the system.
Briefly, a C# ap-plication developed in Visual Studio (Community 2019, Microsoft, Redmond, WA,
USA) controlled the information flow between the system modules and graphical user interfaces
(GUI), while the Labstreaming Layer protocol (LSL) (Swartz Center for Computational
Neuroscience, 2022) transmitted data between modules for processing, game interaction, and
storage. Importantly, connecting modules through the LSL network provides an architecture that
allows for the functional independence of each component. This is advantageous for continued
development since different configurations of sensors, processing pipelines, and environments
can be implemented without necessitating updates to other modules. Specifically, while we only
tested EMG biofeedback on a laptop screen in the current study, the system builds on the modular
architecture we developed in previous work, allowing for the input of electroencephalography or
movement data, and visual output to either a laptop screen or HMD-VR system as well as
proprioceptive feedback via handheld controllers (Marin-Pardo et al., 2020; Vourvopoulos, Pardo,
et al., 2019). Additionally, we incorporated a NeuroPype script (Intheon, San Diego, CA, USA) for
real-time visualization of the digitized EMG signals.
We designed Tele-REINVENT with several GUI configuration screens to accommodate
different user needs and allow for user-friendly set-up depending on the user’s role as a
participant, clinician, or researcher. In this way, Tele-REINVENT can be used as a system for
stroke survivors to use on their own at home, with minimal interaction and configuration required
for set up, as shown in Figure 3.1b. Alternatively, as tested here, we also believe that effective
training requires a trained clinician to specify parameters based on the participant’s profile and,
thus, we developed a second GUI specifically for clinicians to use and customize, shown in Figure
48
3.1e. This interface also helps to quantify clinical data for the clinician to monitor progress. Finally,
researchers may also desire to use this system and tune more specific parameters to test different
hypotheses; there is a third GUI with more detailed parameters to configure.

Figure 3.1 Tele-REINVENT architecture, prototype, and interfaces.
(a) Software architecture consisting of an acquisition client for electromyography (EMG), signal processing
scripts, and a game engine for feedback visualization. Each module is managed by the main application
and all data is internally streamed and stored using the Labstreaming Layer (LSL) protocol. This
architecture supports both remote data retrieval and updates. Green arrows represent data and blue arrows
configurations and commands. (b) Participant graphical interface. Here, the participant can select the game
to play and set user configurations accordingly. (c) Case and Myoware sensor with connectors for
disposable EMG electrodes. (d) Tele-REINVENT prototype consisting of 2 Myoware EMG sensors, a
Teensy 4.0 development board, and a USB isolator enclosed in 3D-printed cases. (e) Clinician graphical
interface. Here, the clinician can set up configurations for the biofeedback game, such as number of trials,
trial duration, and difficulty. Additionally, the interface can show data from previously recorded sessions.

49
3.3.3 Acquisition
We used a Teensy 4.0 development board (PJRC.COM LLC, Sherwood, OR, USA) and a pair
of Myoware muscle sensors (Advancer Technologies LLC, Raleigh, NC, USA) to acquire signals
from the wrist extensor and flexor muscle groups. We designed and manufactured 3D-printed
cases to enclose all hardware components to provide electrical insulation and improve durability.
First, each Myoware sensor amplified the signal registered between a pair of differential
electrodes. Then, the Teensy board digitized the signals and sent them to the computer through
a USB cable via serial communication. Finally, a C# application connected the board with the
computer and streamed the signals to a local network using the LSL protocol. All scripts were
custom made using Visual Studio and Arduino Software (Arduino AG). This configuration allowed
us to acquire up to 4 channels of EMG signals at 2000 Hz with 12-bit ADC resolution. Importantly,
the Myoware sensors are not capable of in-board filtering; raw EMG signals are only differentially
amplified. The Myoware board was designed to be used with microcontrollers that require positive
voltages for analogic acquisition, e.g., an Arduino or a Teensy board. Thus, Myoware centers the
amplified signal around half the voltage used to power it, that is, about 1.65 V when powering the
Myoware with 3.3 V. This is done by the electronics of the board as part of the signal conditioning.

3.3.4 EMG Signal Processing and Biofeedback
We developed custom scripts in Matlab (R2020a, The Mathworks, Natick, MA, USA) to
process EMG signals in real-time and use them for game control. Digitized signals were filtered,
rectified, and normalized to a prerecorded maximum grip. The filter we implemented also removes
the DC offset that may be present in the digitized signal. Thus, a value of normalized activity close
to 0 corresponds to no volitional activity registered by the sensor and a value close to 1 represents
an amplitude similar to that seen during the attempted grip. The clinician or researcher can modify
the specifics of the processing algorithms based on the study needs. Each game was developed
in the Unity game engine (v2019.3.12f1, Unity Technologies, San Francisco, CA, USA) to provide
50
feedback in the form of different games and stream game interactions (e.g., current score and
trial number) through the LSL protocol. Games currently in the Tele-REINVENT suite are
described below:
• SeeEMG — The purpose of this game is to provide quick visual feedback of real-time extensor
and flexor muscle activity as independent continuous streams. After ap-plying the signal
processing steps described above, EMG signals are plotted in each frame as a continuous
line for each of the muscles. The participant is allowed to select one of two configurations.
Option 1 displays EMG smoothed with a 250-ms Hann window, down-sampled to the refresh-
rate of the screen. This option uses a heatmap color scale to map high amplitudes in red and
low amplitudes in blue (Figure 3.2c). The default range sets “high” as the same amplitude as
the maximum grip used to calibrate the system, and “low” as the lowest amount of activity
(typically, none) recorded during calibration. This range can be dynamically adjusted by the
clinician based on the average amplitude of the signals to ensure the whole range of colors is
rendered. Option 2 displays rectified EMG, down-sampled to the refresh rate of the screen,
as white lines. A dynamic green horizontal line is also plotted to represent the maximum
amplitude reached during that session (Figure 3.2d).
• SkeeBall — The purpose of this game is to improve wrist extensor activity and control while
reducing abnormal coactivation of wrist flexors during active extension at-tempts. In this game,
we use the ratio of activity from the wrist extensors and flexors to move a ball to different
targets. First, we apply the signal processing steps de-scribed above. Then, we calculate an
extensor ratio (ER) value (shown in Equation 3.1), defined as the sum of the mean extensor
activity divided by the sum of the mean activity from extensor and flexor muscles.
ER =
e x te n so r
e x te n so r + f le x o r
(3.1)
ER values closer to 1 indicate more individuated extension, closer to 0.5 more
coactivation, and closer to 0 more individuated flexion. Previous research has suggested
51
different effective methods to calculate the contribution of antagonist activity in different
movements and muscle groups (Kellis et al., 2003; Lobov et al., 2018; Luo et al., 2018; Wright
et al., 2014). We selected the current approach for its simplicity, low computational
requirements, and its ability to capture the relationship of two antagonistic muscles. As noted
above and mentioned in our previous work (Marin-Pardo et al., 2020), we expect that using
this ratio is advantageous to encourage wrist extension without simultaneous unintended wrist
flexion. For each trial, a score is assigned as a function of the ER value and a probability
likelihood as shown in Table 3.1. That is, an ER value of 0.7 would result in 20 points in 60%
of trials and 30 points in 40% of trials. Calculated point values trigger the corresponding
animation of the hand hitting the ball and the ball moving to the appropriate ring (Figure 3.2a).
At the end of the game, the final score is calculated as the cumulative points for each trial.
Importantly, the probability values shown in Table 3.1 are only used to determine the score of
the trial. That is, a calculated value of individuated extension (ER > 0.5) will always be
positively reinforced, providing higher scores to higher individuation values.
• Blinko — The purpose of this game is to improve both wrist extensor and wrist flexor activity
while reducing abnormal coactivation. Similar to SkeeBall, we use the ratio of activity from the
wrist extensors and flexors, the extensor ratio (ER), to move a round flat disc, or chip, across
the top of a board. The board was modeled after the popular gameshow chance game Plinko,
as shown in Figure 3.2b. The board consists of rows of pegs, with each row offset from the
one above it. At the bottom of the board, there are nine slots that represent points. These are
labeled as follows: $100, $500, $1000, $0, $10,000, $0, $1000, $500 and $100. When a chip
is dropped at a slot between two pegs at the top of the board, the chip is diverted by the slots
below it. Thus, the chip may take any number of paths to the bottom and land in any of the
nine slots. For each trial, the player attempts to move the chip right with wrist extension or left
with wrist flexion. Then, the player drops the chip and the points assigned correspond with the
52
bottom slot where the chip landed. At the end of the game, the final score is calculated as the
cumulative points for each trial.
Table 3.1 Score likelihood for the SkeeBall game as a function of the extensor ratio (ER).
ER value Points Likelihood
0.9 < ER  0.99 30 100%
0.6 < ER  0.9 30 40%
20 60%
0.4 < ER  0.6 20 40%
10 60%
0.2 < ER  0.4 10 20%
0 80%
ER  0.2 0 100%

Figure 3.2 Examples of EMG biofeedback games.
(a) SkeeBall: adequate extensor ratio muscle activity translates as movement of the arm shooting the ball
to different targets. (b) Blinko: adequate extensor ratio muscle activity translates as movement of the
character left and right, who then drops a chip to win points. (c) SeeEMG: smooth EMG is mapped to show
high amplitudes in red and low amplitudes in blue. A dynamic range to scale amplitudes can be modified
by the clinician. (d) SeeEMG: down-sampled rectified EMG. The green horizontal line represents the
maximum value reached in that session.
53
3.3.5 Example of a Tele-REINVENT Telerehabilitation Session with a Clinician
Before starting the telerehabilitation sessions, a trained clinician administers a series of
behavioral assessments to quantify the participant’s baseline level of impairment to determine
training parameters, e.g., training duration, rest intervals, and the number of repetitions per game.
These parameters are set by the clinician or researcher and used for remote training sessions
but can be modified on a regular basis as sessions advance. Then, the participant undergoes a
detailed orientation showing them step-by-step how to use the system and how to properly place
the sensors. Alternatively, if it is unfeasible to do all procedures in person (e.g., due to logistics or
social distancing protocols), the orientation can also be performed using the incorporated video
conferencing application (Zoom Video Communications, Inc., San Jose, CA, USA). Training
sessions can be done by the participant alone or remotely monitored by the clinician or
researcher. Figure 3.3 shows an example of a guided session in which the screen can be
configured to view real-time performance of SkeeBall game, the participant’s electrode placement
and body movements, the clinician, and real-time EMG signals viewed all at once. Video meetings
with the occupational therapist can take place as often as needed (e.g., in our case study de-
scribed below, they were performed daily for the first week, and afterward on a bi-weekly basis or
if there were technical problems with the system). The participant and clinician also
communicated every other day via email. As shown in previous studies, regular contact with the
clinician is critical for maintaining participant engagement and adherence (Cramer et al., 2019).
54

Figure 3.3 Feedback displayed during a session guided by an occupational therapist.
(a) SkeeBall game rendered on the computer screen. (b) View from the laptop camera of the participant’s
arm with electrode placements. (c) Occupational therapist guiding the session. (d) Real-time visualization
of the EMG signals.

Our Tele-REINVENT system consists of a laptop computer with all necessary pro-grams
preloaded, configured, and displayed in an easy-to-use manner, a pair of EMG sensors with the
enclosed acquisition board, and a package of disposable electrodes. When participants first
receive the equipment, they are also given a printed manual of how to use the system. For
enhanced simplicity and consistency, each session starts with instructional videos that remind the
participant how to start a telerehabilitation session, plug in the acquisition device, and position the
sensors over the targeted muscles. Then, a calibration video guides the participant through wrist
and hand movements, including gross grasp, wrist extension, and wrist flexion. Upon completion
of the recording, a Matlab script filters and rectifies the signals, as described above, and calculates
the mean amplitudes during the grip. These values are later used as a calibration to normalize
real-time signals. Then, the participant selects which game they will play and sets visual
configurations for the game, e.g., color-coded signals or arm model (Figure 3.1b). Once the
55
participant is ready, the game starts. The number of repetitions per game, rest periods between
trials, and the duration of gameplay is determined by the researcher or clinician in collaboration
with the participant. These parameters are influenced by fatigue, spasticity, endurance, and other
personal factors. Finally, after completion of each training session, deidentified recorded data is
securely synchronized to the clinician’s or re-searcher’s computer via Google Drive (Google,
Mountain View, CA, USA).

3.3.6 Signal Quality Validation Example
We present an example in two healthy individuals to validate the use of low-cost Myoware
sensors by comparing calculated values of muscle coactivation with those obtained using a
research-grade Delsys Trigno Wireless System (Delsys Inc., Natick, MA, USA). Table 3.2
presents a summary of the specifications for both sensors.
Table 3.2 Sensor specifications.
Specification Delsys Trigno Avanti Tele-REINVENT
EMG sensors 4 2
Inter-electrode spacing 10 mm 20 mm
EMG sampling rate (max) 4370 sa/sec 2000 sa/sec
Resolution 16 bits 12 bits
In-board amplification 1000 V/V 500 V/V
In-board filter
(max range)
Butterworth bandpass
(10–850 Hz)
NA
1

Estimated price USD 8400
2
USD 150
3

1
NA = not available. This sensor does not incorporate in-board filtering.
2
Price may vary since a quote
from the vendor is required. This price included the proprietary acquisition system and software (Trigno
Avanti Platform) and 4 Trigno Mini Sensors.
3
Estimated cost of the required components to build a Tele-
REINVENT system, including a Teensy 4.0 board and 2 Myoware sensors. Both Delsys and Tele-
REINVENT require a laptop computer (not included in these estimates).

56
As noted previously, for this analysis we recruited 2 male participants and recorded a series
of repeated time-constrained ramp-up and hold isometric grips, similar to what was performed
during the Tele-REINVENT tasks. First, we cleaned the skin with iso-propyl alcohol and
electrodes were placed over the extensor carpi radialis and flexor carpi ulnaris of the dominant
hand. Proper positioning was confirmed via palpation and observation of the EMG signals during
wrist flexion, extension, ulnar deviation, radial deviation, and gross grip. Then, we simultaneously
recorded EMG and exerted force during a 3-second power grip using 2 Trigno Mini sensors and
a digital grip dynamometer (Biometrics Ltd., Ladysmith, VA, USA). All signals were digitized with
a data acquisition system (USB-6210, National Instruments, Austin, TX, USA) at 1000 Hz and
acquired with a custom interface in Matlab.
To quantify the performance of each sensor, each participant sat in front of a computer screen,
their arms rested comfortably on a table and we provided feedback of exerted force during
isometric grip holds. We asked the participant to follow the target shown in Figure 3.4a consisting
of 5 s to reach 25% of the previously recorded maximum grip force, 5 s holding the force level,
and 7 s of rest. We provided visual feedback with a cursor swiping through the screen from left to
right crossing the screen in 17 s and moving up and down according to the current force amplitude
until the completion of 12 trials. Finally, we removed the Trigno Mini sensors, placed the Myoware
sensors over the same location, and repeated the feedback task. We used the same calibrated
grip force for recordings with both sensors.
57

Figure 3.4 Test of muscle control.
Two healthy control participants received feedback of their exerted grip force and were asked to follow an
amplitude-guided 5-second build-up to a hold level (25% of their maximum voluntary contraction (MVC))
during 5-second epochs, followed by 7 s of rest. The activity of both muscles was analyzed during the ramp-
up and hold phases of the task (first two regions marked within dashed lines). (b) Calculated extensor ratio
values for each time point of the 24 5-second isometric grip holds (12 grip holds with each device) using
data acquired from extensor carpi radialis (ECR) and flexor carpi ulnaris (ECU) with Delsys and Myoware
for each participant (denoted as 1 and 2). All distributions are centered around 0.5 in-dicating a similar
proportion of muscle activity recorded from both muscles with either device.

For offline analysis of both systems, EMG signals were preprocessed in Matlab with a
bandpass filter between 150 and 450 Hz, full-wave rectified, and smooth with a moving-average
rectangular window of 1s. This filter was chosen to match the preprocessing methods for the
patient case study (see below), although using a wider range of 15 and 450 Hz results in no
58
differences. Epochs of the hold phase were extracted to calculate values of normalized ER.
Finally, we used paired t-tests to evaluate whether distributions of ER values acquired with the
low-cost Myoware sensors were comparable to those acquired with the research-grade Delsys
system. We used linear regressions to evaluate the proportion of explained variance between the
muscle activity and the corresponding grip force during the ramp-up phase of the trials. We utilized
custom scripts in R (version 4.0.3, R Foundation for Statistical Computing, Vienna, Austria) to
perform all statistical analyses.

3.3.7 Case Study and Feasibility
We tested the feasibility and preliminary effects of using this system by delivering Tele-
REINVENT to a participant with stroke. We asked him to complete 40 sessions that spanned 10
weeks of training. For each session, the participant completed 5 blocks of 20 repetitions using the
SkeeBall game, described above, for 100 trials per session, which took approximately 1 h per
session. It is important to note that it was not possible to perform in-person behavioral
assessments before and after use, nor in-person familiarization with the system to avoid
unnecessary risks due to ongoing social distancing protocols during the COVID-19 pandemic.
Similarly, it was not possible to demonstrate proper sensor placement in-person. All familiarization
and verification of sensor placement was per-formed remotely via videoconferencing with the
clinician, who used the real-time EMG signal tracking to assess electrode placement. Therefore,
signal quality could still be variable across training sessions. In addition, the participant also
actively collaborated on providing user feedback and testing the device during the initial system
development for several weeks prior to beginning data collection. Because these sessions were
part of familiarization and configuration tunning, and not carefully structured, they are not included
in the current analysis.
For real-time online analysis, digitized EMG signals were bandpass filtered between 150 and
450 Hz, rectified, and normalized to the calculated activity during the grip at-tempt. Then, we
59
calculated ER values to provide positive reinforcement of individuated extension using the
SkeeBall game, as described above. Importantly, we selected this higher frequency range to
account for possible artifacts induced by motion, crosstalk, and increased susceptibility to
environmental noise. Larger contributions from these artifacts were expected due to the lack of
in-board filtering of the Myoware sensors and lack of in-person training. Previous research has
shown benefits of using high pass frequency filters at the 100 Hz range or above to account for
such artifacts and obtain better estimations of force, joint stability, motor unit recruitment,
intermuscular coherence, and corticomuscular coherence (T. W. Boonstra & Breakspear, 2012;
Brown et al., 2010; Laine et al., 2012; Laine & Valero-Cuevas, 2017, 2020; Potvin & Brown, 2004;
Riley et al., 2008; Staudenmann et al., 2007, 2010).
For offline analysis of the recorded sessions, trial data from the game and EMG signals were
processed with a custom script in Matlab. First, trial number and EMG signals were interpolated
and down-sampled, respectively, to 1000 Hz. Then, signals were processed with the same
pipeline as the online data; that is, bandpass filtered between 150 and 450 Hz, full-wave rectified,
and normalized to the session’s recorded calibration with values of maximum activity within a 250-
ms moving window. Then, values of averaged activity were calculated from the first 2 s of each
movement attempt trial. To account for the within-session variability and evaluate how these
activity patterns changed over time, we averaged the activity of the 100 trials of each session.
Finally, we used Pearson’s correlations to test for significant changes in EMG signals and game
performance across 28 training sessions. 12 sessions were excluded from the analysis due to
excessive noise in the recording or missing data. We examined session averages for normalized
extension activity, flexion activity, and ER with custom scripts in R.

60
3.4 Results
3.4.1 Signal Quality Validation Example
In this validation example, Myoware and Delsys systems showed no significant differences in
ER values during isometric grip holds. Paired t-tests, one for each participant, showed no
significant difference in ER values when comparing the two sensors (participant 1: t = −0.025, p
= 0.98; participant 2: t = −2.018, p = 0.068), as shown in Figure 3.4b. To ensure the systems do
not show a significant divergence within lower forces, we used linear regressions of the
normalized activity recorded from each muscle to the measured force during the ramp phase of
the trials, shown in Table 3.3. Across both participants, between 79 and 94% of the variance in
force was explained with the normalized EMG signals when these were acquired using the Delsys
system. The explained variance when recording with the Myoware dropped to within 19 and 78%.
Although ranges of explained variance suggest higher sensitivity to noise when using Myoware,
similar regression slopes during force build-up and the distributions of ER values during static
holds indicate that for our purposes, the two systems may provide similar measures of muscle
activity.
Table 3.3 Muscle activity correlation with grip force while building-up to a sustained grip.
Participant Muscle Sensor Slope R
2

1 Extensor Delsys 1.54 0.940
1 Extensor Myoware 1.77 0.784
1 Flexor Delsys 1.53 0.795
1 Flexor Myoware 1.11 0.543
2 Extensor Delsys 1.46 0.935
2 Extensor Myoware 1.64 0.339
2 Flexor Delsys 1.69 0.885
2 Flexor Myoware 1.46 0.193

61
3.4.2 Feasibility, Safety, and Adherence
The participant with stroke used the system for 40 sessions over 10 weeks, with 100%
adherence to the requested protocol. In terms of user experience, the participant reported no
perceived discomfort, pain, or fatigue, and there were no adverse events during any sessions
using the system. The average session duration was 40 min, including about 10 min for setup, 30
min of repeated practice (5 blocks of 20 trials per session), and 1–2 min of rest between blocks
of trials. Qualitatively, after 40 sessions, values of calculated individuated activity of the extensor
muscle showed a significant increase over time (rho = 0.59, p < 0.001), as shown in Figure 3.5d.
Normalized activity for the extensor muscle appeared to increase over time, while flexor activity
decreased, as seen in Figure 3.5b. However, these individual changes were not statistically
significant for either the extensor (rho = 0.27, p = 0.164) or the flexor (rho = −0.34, p = 0.071)
muscles. Similarly, while changes in game performance improved, this change did not reach
statistical significance (rho = 0.29, p = 0.06). Finally, the participant reported positive changes in
motor function, e.g., increased extension at the interphalangeal joints and decreased frequency
of muscle spasms, as well as improved overall quality of life, e.g., improved quality and duration
of sleep. The participant also reported enjoyment of using the system and requested to keep it
after the testing period concluded.
62

Figure 3.5 Muscle activity during REINVENT sessions.
Example of representative activity from the extensor (red) and flexor (blue) muscles during the calibration
of early (a) and late (c) trials. The calibration algorithm looks at the recorded activity during which a video
cued movement attempts and calculates the highest rectified muscle activity of a 250-ms moving window
average for each muscle. These values are later used to normalize the activity during each trial and
calculate extensor ratios (ER) to control the game. (b) Normalized muscle activity (EMG) averaged across
40 sessions for extensor and flexor muscles. Changes over time were not significant. (d) Calculated values
of ER averaged across 40 sessions. A significant increase in extension individuation appeared during
movement attempts (p < 0.01). Regression lines are shown for visualization in panels (b) and (d).

3.5 Discussion
In this paper, we report the development of a muscle-computer interface to train muscle
activity from wrist extensors while limiting unintended coactivation of wrist flexors for at-home
chronic stroke telerehabilitation. We first described the hardware and software components of the
system including acquisition, processing pipeline, and feedback in the form of game control. We
also provided a validation example in two healthy individuals, comparing the use of low-cost
sensors to calculate ratios of muscle activity versus research-grade sensors while performing the
63
same task. Finally, we showed the preliminary feasibility, safety, adherence, and efficacy in a
single case study of one person with a stroke using this system for 40 sessions over 10 weeks.
Overall, although the data are limited, our preliminary results suggest that our low-cost, portable
EMG biofeedback system may be used with telerehabilitation to contribute to the development of
accessible technology to improve post-stroke recovery.
In a previous study, we showed that it is feasible to use a research-grade EMG acquisition
system to train the activity of agonist muscles while avoiding the simultaneous activity of
antagonists via EMG biofeedback (Marin-Pardo et al., 2020). Here, we showed that these ratios
of muscle activity can be acquired in an individual with stroke using low-cost sensors. In line with
previous work (Fuentes del Toro et al., 2019; Lobov et al., 2018; Toro et al., 2019), we also show
that using low-cost sensors can result in similar results to research-grade sensors for low forces
during an isometric task in two healthy individuals, despite their differences regarding electrode
type, in-board filtering, and sensitivity. However, further investigation with a larger population is
required to ensure the validity and robustness of our system. Although a high proportion of the
measured force variability during the ramp phase of our task was explained by the recorded
muscle activity with either system, this proportion dropped when using the low-cost sensors. This
was expected due to the lack of in-board filtering, making the sensor more susceptible to noise
and adding variability to the measurement. The absence of in-board filtering could also increase
crosstalk, lower signal to noise ratios, and lower sensitivity, affecting the quality of EMG signals.
While these challenges can be addressed, to some extent, with proper skin preparation and
adequate sensor placement, it may also be beneficial to explore novel signal processing
techniques to account for the variability of these factors across sessions in future home-based
interventions. Previous research has evaluated the use of low-cost sensors and shown that they
can be used as an alternative to research-grade equipment in healthy populations (Fuentes del
Toro et al., 2019; Toro et al., 2019). In this pilot study, we explored the use of these sensors with
a stroke survivor, however, further research with an adequate sample size is required to evaluate
64
the reliability of these systems for additional users with movement impairments. We used an
isometric task as validation because most of our target population is incapable of performing
volitional movement. In addition, validation of non-isometric movement tasks with simultaneous
recordings using both low-cost technology and research-grade equipment could further improve
the performance and accessibility of the Tele-REINVENT system.
Finally, we showed that our system was feasible, safe, and enjoyable to for 40 1-hour sessions
over 10 weeks of training. Notably, the participant used the system at home without any in-person
instruction or guidance, relying only on telerehabilitation video calls and emails with the clinician
and research team. The participant reported high satisfaction with the system, demonstrated
100% adherence to the research protocol, and requested to keep the system afterward. The
participant also demonstrated significantly increased extensor activity while decreasing flexor
activity over time. We anticipate that improved results could be gained with an initial in-person
evaluation and familiarization with the system, as well as improved sensor placement at the first
visit with muscle palpation to identify optimal electrode placement. In addition to in-person
training, the Tele-REINVENT system could be further improved with enhanced calibration
algorithms that account for placement variability across sessions as well as processing pipelines
to help personalize the biofeedback for each individual. For example, including a higher number
of sensors would allow the use of algorithms to detect the best combination of sensors, similar to
the work by Lobov et al. (Lobov et al., 2018). Finally, although we showed moderate
improvements in coordinated muscle activity, that is, reduced antagonist activation during
attempted agonist activation, our reinforcement paradigm could be enhanced with improved in-
game feedback. For example, in the future, we may incorporate adaptive control thresholds to
encourage more individuated activity on a personalized basis, in a similar fashion to our previous
work (Marin-Pardo et al., 2020).

65
Limitations and Future Work
In terms of hardware, although we demonstrated the validity of using the proposed sensors,
future work may incorporate wireless communication between the computer and the sensors to
increase mobility and provide enhanced electrical insulation for the sensors. Additionally, we used
a laptop computer with specifications that allowed adequate re-al-time processing of the EMG
signals and simultaneous rendering of the feedback game, along with video conferencing.
However, future iterations may use lower-cost hardware, such as a tablet or cellular phone,
instead of a full laptop. In terms of software, even though we made significant efforts to develop
a user-friendly system, it still required a minimum level of computer literacy for adequate setup
and use. Therefore, to further increase the system’s portability and acceptability, it would be
crucial to optimize the acquisition and processing pipelines as well as the graphical interfaces to
provide feedback in mobile devices, such as cellular phones and tablets. Furthermore, the
variability of sensor lo-cation across sessions and the level of impairment of the participants make
it necessary to develop and test more robust calibration and processing algorithms that account
for such variability. For example, normalizing the recorded activity with the maximum amplitude
registered during any attempted movements, as opposed to specifically using an at-tempted grip,
would allow system usage to extend to participants without the ability to grasp an object.
Improvements in user interaction and calibration will ensure both an easier setup and increased
in-session robustness, even for sessions with inadequate sensor placement. In addition, an
expanded catalog of games could increase not only the acceptability of the system but the variety
of tasks to train. Finally, it is important to note that while the purpose of this work was to
demonstrate the feasibility of developing and deploying a portable muscle–computer interface for
telerehabilitation, the current investigation only provides a description of the system, a limited
comparison of the low-cost sensors with research grade sensors in two healthy individuals, and
a single case study with one person with stroke. Further research with large and diverse
populations is needed to examine the efficacy of this system for telerehabilitation services, along
66
with pre- and post-intervention behavioral assessments to further evaluate the effects of this
system in promoting functional recovery at both individual and group levels. In summary, this work
shows that using low-cost technology for individuated muscle training to reduce unintended
coactivation for supervised and unsupervised home-based telerehabilitation is feasible and may
widen the development of meaningful and personalized chronic stroke rehabilitation.

67
Chapter 4: Functional and Neuromuscular Changes Induced via a
Low-Cost, Muscle-Computer Interface for Telerehabilitation:
A Feasibility Study in Chronic Stroke
This section is adapted from:
Marin-Pardo, O., Donnelly, M. R., Phanord, C. S., Wong, K., Pan, J., and Liew, S.-L. (2022).
Functional and Neuromuscular Changes Induced via a Low-Cost, Muscle-Computer Interface for
Telerehabilitation: A Feasibility Study in Chronic Stroke. Frontiers in Neuroergonomics.

4.1 Abstract
Stroke is a leading cause of adult disability in the United States. High doses of repeated task-
specific practice have shown promising results in restoring upper limb function in chronic stroke.
However, it is currently challenging to provide such doses in clinical practice. At-home
telerehabilitation supervised by a clinician is a potential solution to provide higher-dose
interventions. However, telerehabilitation systems developed for repeated task-specific practice
typically require a minimum level of active movement. Therefore, severely impaired people
necessitate alternative therapeutic approaches. Measurement and feedback of electrical muscle
activity via electromyography (EMG) have been previously implemented in the presence of
minimal or no volitional movement to improve motor performance in people with stroke.
Specifically, muscle neurofeedback training to reduce unintended co-contractions of the impaired
hand may be a targeted intervention to improve motor control in severely impaired populations.
Here, we present the preliminary results of a low-cost, portable EMG biofeedback system (Tele-
REINVENT) for supervised and unsupervised upper limb telerehabilitation after stroke. We aimed
to explore the feasibility of providing higher doses of repeated task-specific practice during at-
home training. Therefore, we recruited 5 participants (age = 44–73 years) with chronic, severe
impairment due to stroke (Fugl-Meyer = 19–40/66). They completed a 6-week home-based
68
training program that reinforced activity of the wrist extensor muscles while avoiding coactivation
of flexor muscles via computer games. We used EMG signals to quantify the contribution of two
antagonistic muscles and provide biofeedback of individuated activity, defined as a ratio of
extensor and flexor activity during movement attempt. Our data suggest that 30 1-hour sessions
over six weeks of at-home training with our Tele-REINVENT system is feasible and may improve
individuated muscle activity as well as scores on standard clinical assessments (e.g., Fugl-Meyer
Assessment, Action Research Arm Test, active wrist range of motion) for some individuals.
Furthermore, tests of neuromuscular control suggest modest changes in the synchronization of
electroencephalography (EEG) and EMG signals within the beta band (12-30 Hz). Finally, all
participants showed high adherence to the training protocol and reported enjoying using the
system. These preliminary results suggest that using low-cost technology for home-based
telerehabilitation after severe chronic stroke is feasible and may be effective in improving motor
control via feedback of individuated muscle activity.

4.2 Introduction
Almost 800,000 people have a stroke each year in the United States (Virani et al., 2020),
making stroke a leading cause of long-term adult disability. Recent research has shown that high
doses of repeated task-specific practice may improve upper limb function in the chronic phase of
stroke (>6 months after onset) (Ballester et al., 2019; Ward et al., 2019; Winstein et al., 2016).
However, it is currently challenging to provide such doses in standard clinical practice due to time,
physical, and economic constraints. At-home telerehabilitation services supervised by a clinician
are a potential solution to provide higher-dose interventions.
Recent studies have shown that post-stroke telerehabilitation is feasible to induce positive
changes and may also be as effective as in-person interventions (Cramer et al., 2019; Dodakian
et al., 2017). However, repeated task-specific practice interventions, including those that use
telerehabilitation services, typically require a minimum level of active movement (Cramer et al.,
69
2019; Ward et al., 2019; Winstein et al., 2003). Therefore, severely impaired survivors necessitate
alternative therapeutic approaches. Biofeedback of relevant physiological activity (e.g., via
electroencephalography (EEG) or electromyography (EMG)) are promising approaches that have
been implemented in the presence of minimal or no volitional movement and could be
incorporated with telerehabilitation (Armagan et al., 2003; Remsik et al., 2016; Soekadar et al.,
2015; Wright et al., 2014). Specifically, training to reduce unintended co-contractions of the
impaired hand may be a targeted intervention to improve motor control in severely impaired
populations (Donoso Brown et al., 2014; Mugler et al., 2019).
Recently, we tested the feasibility of a multimodal biofeedback system (REINVENT) that can
interchangeably operate as a brain-computer interface (BCI) or a muscle-computer interface
(MCI), and that integrates immersive virtual reality (VR) for upper limb chronic stroke rehabilitation
(Marin-Pardo et al., 2020; Vourvopoulos, Pardo, et al., 2019). Both pilot studies with the
REINVENT system showed promising results in terms of user satisfaction and moderate
improvement in clinical assessments of stroke recovery. Ongoing limitations for in-person
research (e.g., sanitary precautions due to the COVID-19 pandemic) and the growing interest in
incorporating telerehabilitation services into standard clinical practice prompted us to develop a
low-cost, portable version of our system (Tele-REINVENT) (Marin-Pardo et al., 2021).
Additionally, using mobile devices for at-home telerehabilitation may further improve care access
for underserved populations limited, for example, by their proximity to rehabilitation centers,
mobility level, access to transportation services, and cost of care (Health Resources and Services
Administration, 2022; Marzolini et al., 2016).
Here, we explored the feasibility of using Tele-REINVENT with five stroke survivors in the
chronic stage of recovery that presented with severe hemiparesis (e.g., less than 20 degrees of
voluntary wrist motion) and unintended antagonist activation during attempted wrist movements.
We asked them to complete 30 remote training sessions, where our system reinforced extensor
activation without simultaneous activation of flexor muscles. To quantify changes before and after
70
training, we used a battery of standard clinical assessments in combination with a test of muscle
control to evaluate generalized functional improvements. Additionally, we evaluated changes over
time during the training intervention by quantifying the activation of agonist and antagonist
muscles and their overall contribution towards less unintended coactivation. Finally, we included
measurements of corticomuscular coherence (CMC) to investigate possible changes in brain
organization before and after training. Although variable across stroke survivors (e.g., in terms of
magnitude and localization), previous research has suggested that CMC could be interpreted as
a proxy of tract integrity and neural recovery (Krauth et al., 2019; J. Liu et al., 2019; Marin-Pardo
et al., 2020; Rossiter et al., 2013).
Overall, the purpose of this work is to explore the feasibility of Tele-REINVENT to safely
provide higher doses of repeated task-specific practice during at-home upper limb training.
Previous literature has identified that adherence to post-stroke home programs ranges from 50 to
100% (Cramer et al., 2019; Dodakian et al., 2017; Donoso Brown et al., 2014; Jurkiewicz et al.,
2011). We designed Tele-REINVENT to overcome some of the known limitations of technology-
based at-home training programs (e.g., portability and ease of use – further detailed in (Marin-
Pardo et al., 2021)). Therefore, we hypothesized that at-home reinforcement of EMG activity using
Tele-REINVENT in participants with severely impaired motor function of the upper extremity
would be safe, feasible, and enjoyable for participants to use with high adherence (e.g., at least
80%). Moreover, as previous research by us and others has shown, not all participants will
respond to training interventions (Kwakkel et al., 2016; Marin-Pardo et al., 2020; Vourvopoulos,
Pardo, et al., 2019; Ward et al., 2019). Therefore, we hypothesized that training with Tele-
REINVENT would produce improvements in clinical assessments for some but not all participants.
Finally, we expect to show evidence of improved neuromuscular control, measurable as increased
muscle recruitment individuation (i.e., group-specific activation with reduced unintended
coactivation) and improved CMC.
71
4.3 Materials and Methods
4.3.1 Participants
For this case series, we recruited five adult stroke survivors in the chronic phase of recovery
(>6 months since onset). Inclusion criteria required that participants presented with moderate to
severe upper extremity hemiparesis and residual hand function (e.g., less than 20 degrees of
active wrist or finger extension and enough muscle activity to measure with electromyography).
Furthermore, participants were not taking anti-spasticity medication and had no significant vision
loss (corrected vision was acceptable), receptive aphasia, hand contractures, or a secondary
neurological disease. Finally, none of the participants were receiving additional physical or
occupational therapy targeting wrist movements; however, regular exercise (e.g., training at the
gym) was allowed. Participants completed a screening session prior to enrollment to ensure their
compliance with these criteria. Additionally, we screened for cognitive impairment, as severe
impairment might impede participation in the study tasks. The experimental protocol was
approved by the Institutional Review Board of the University of Southern California (reference
number: HS-17-00916, approved on 7/20/2021) and all participants provided written informed
consent in accordance with the Declaration of Helsinki. A summary of the participant
demographics is presented in Table 4.1.

Table 4.1 Participant demographics and baseline evaluations.
Participant Age Onset
(months)
Paresis FMA ROM WE
(°)
MRS MOCA
1 61 156 Left 20 10 2 22
2 73 130 Left 19 10 2 22
3 58 14 Left 23 5 2 20
4 57 25 Left 40 22 1 21
5 44 61 Right 25 -15 2 21
Fugl–Meyer Assessment of the upper extremity (FMA), Range of Motion during active Wrist Extension
(ROM-WE), Modified Rankin Scale (MRS), Montreal Cognitive Assessment (MOCA).
72
4.3.2 Study Timeline
We tested the functional outcomes and feasibility of using the Tele-REINVENT system with
stroke survivors during at-home training sessions. An outline of the study timeline is shown in
Figure 4.1, and each component is detailed below. Potential participants were screened for
cognitive impairments using the Montreal Cognitive Assessment MOCA (Nasreddine, Phillips,
Bédirian, et al., 2005) and with a test of EMG amplitude during an in-person visit before enrolling
them in the study. Those who did not present a significant cognitive impairment likely to impair
their ability to use Tele-REINVENT (i.e., having a MOCA score below 20 points) and that could
maintain a minimum level of extensor EMG activity (i.e., hold 30% of a prerecorded maximum for
ten 4-second trials) were enrolled in the remote training protocol, where each participant was
asked to complete thirty 1-hour remote training sessions over six weeks. These sessions were
performed independently or remotely monitored by the research team (e.g., by an occupational
therapist or a research engineer), as described below. We evaluated participants’ improvements
with physiological and clinical assessments of upper limb function during pre- and post-training
in-person visits to our laboratory (sessions 1 and 32). Furthermore, we evaluated their muscle
control with a static hold test. In this test, participants used feedback of their wrist extension and
flexion EMG amplitude to follow a target level of activation (Figure 4.1). During this tracking test,
we simultaneously recorded EEG over the ipsilesional and contralesional motor cortices to
evaluate their CMC. We made every effort to perform in-person evaluations within less than three
hours per session (breaks included) to avoid unnecessary burden on our participants. In sessions
2–31 participants were asked to complete an hour of remotely monitored or independent at-home
training sessions of individuated wrist extension, as detailed below. We included an additional in-
person visit after completion of half of the remote sessions to ensure proper location of the EMG
sensors and no adverse effects for the remainder of the experiment. As mentioned above, since
we were interested in functional changes before and after 30 sessions of remote training, we
excluded this additional visit from our present analyses.
73

Figure 4.1 Experimental protocol.
(Top) Timeline of the 32-sessions program. Participants were asked to complete 30 remote sessions
(sessions 2–31) targeting training wrist extension activation of the more affected arm. We included pre-,
and post-intervention in-person assessments in sessions 1 and 32, respectively. These comprised a battery
of standard clinical assessments and a test of neuromuscular control (i.e., repeated static holds with
simultaneous electromyography (EMG) and electroencephalography (EEG) recordings). (Bottom, left)
Muscle control test (static hold). We tested control over wrist extensor and flexor muscles with a task where
participants were asked to maintain a constant level of 15% of a maximal voluntary contraction (MVC) for
twelve 4-second epochs while receiving feedback of their muscle EMG amplitude on a computer screen.
This test was used to quantify task performance and coherence between brain and muscle signals.
(Bottom, middle) Tele-REINVENT system (in a typical session), consisting of a laptop computer, a pair of
active bipolar EMG sensors, and an acquisition box to digitize the EMG signals. (Bottom, right)
Screenshots of two Tele-REINVENT training games used. These games were developed to actively
encourage wrist extension movements by providing feedback of adequate patterns of muscle activation
(i.e., avoiding unintended coactivation of antagonistic muscles). Successful trials (i.e., when EMG from the
extensor muscle was greater than the flexor muscle) were reinforced via different game mechanics (e.g.,
character movement or points awarded).

4.3.3 At-Home Training (Sessions 2 –31)
Our training paradigm builds on the protocol and telerehabilitation system that we described
in our previous work (Marin-Pardo et al., 2020, 2021). Briefly, our Tele-REINVENT system
consists of a laptop computer with all necessary programs preloaded, configured, and displayed
in an easy-to-use manner, a pair of low-cost EMG sensors (Figure 4.1), and a package of
disposable electrodes and alcohol wipes. To synchronize, record, and transfer data between the
different modules, we used the Labstreaming Layer (LSL) protocol and recorder (Swartz Center
74
for Computational Neuroscience, 2022). EMG signals were processed and analyzed in real time
with custom scripts in Matlab (R2021a, The Mathworks, Natick, USA) to quantify an activation
ratio of one extensor and one flexor muscle. Both the user interface and the games were
developed using the Unity game engine (v2020.1.11, Unity Technologies, San Francisco, USA)
and rendered on the laptop screen. These games were developed aiming to reward extensor
activation when it was produced without flexor coactivation. Further details regarding game
mechanics and system modules can be found in our previous work (Marin-Pardo et al., 2021).
Below, we present a summary of the games implemented in this study (Figure 4.1).
• SkeeBall. In this game, we use the activity from wrist muscles to move a ball to
different targets according to the ratio of activity between the extensors and flexors.
Calculated ratios are mapped to score values and trigger the corresponding animation
of a hand hitting the ball and the ball moving to the appropriate score ring.
• Blinko. Here we use the ratio of activity from the wrist extensors and flexors to move
a character holding a disc across the top of a vertical game board. For each trial, the
player attempts to move the character left and right with wrist extension and flexion
before time runs out. Then, the character drops the disc and receives the points that
correspond with the slot where the disc landed.
• Planet Jump. This is a side-scrolling game where a character moves along two
dimensions to avoid obstacles within a finite course. For time periods where muscle
activity is below an activation threshold (i.e., at rest), the character runs left to right
across the environment. Otherwise, the ratio of muscle activity is translated as jumping
commands for extension and stopping commands for flexion.
We used the proposed system to train individuated muscle control with the following protocol.
First, participants completed a detailed orientation in session 1, including a step-by-step
demonstration on how to place the EMG sensors and use the system. Additionally, we used a
surgical marker pen to mark adequate sensor locations. Then, the first week of remote sessions
75
(2–5) were monitored by the research team (e.g., an occupational therapist or a research
engineer) to ensure the system worked properly and that participants were confident using it on
their own. Subsequent sessions (6–31) were mostly independent, with the research team
monitoring two sessions of each week. This allows us to ensure the system worked properly,
make necessary adjustments to the game configurations (e.g., number of trials per game or
activation threshold), and interact with the participants. This was particularly important, as
research shows that when using home-based technology for rehabilitation, participant
engagement and adherence are highly dependent on the involvement of clinicians and
researchers (Chen et al., 2019; Cramer et al., 2019; Feldner et al., 2020). Additionally, the Tele-
REINVENT kit included a printed manual with detailed instructions on how to use and setup the
system, a surgical marker pen to mark adequate sensor positioning, and additional objects to
increase comfort and participation during the session (e.g., a low-stiffness ball, a pool noodle, and
a towel). Each remote session is described below.
For enhanced simplicity and consistency, the system interface included tutorial videos to
remind the participants how to start a telerehabilitation session, plug in the acquisition device, and
position the sensors over the targeted muscles. Prior to the start of all training sessions, a
calibration video guided the participants through wrist and hand movements, including gross
grasp, wrist extension, and wrist flexion. Upon completion of the recording, a Matlab script
calculated the mean maximum amplitude of each EMG signal during a 250-millisecond moving
window after processing the signals with a 15–450 Hz bandpass filter and full-wave rectification.
Then, the participants selected which game they would play. This launched another Matlab script
that would continuously calculate ratios of extensor activity as described in Equation 4.1:
ER =
EM G
e x t e n s o r
EM G
e x t e n s o r
+ EM G
f l e x o r

where each EMG signal corresponds to averaged activity of 250 ms of each muscle after
preprocessing and normalization to the calibration activity. In this equation, values closer to 1
76
would indicate that the extensor muscles were comparatively more active than flexor muscles.
Similarly, values closer to 0 would indicate higher flexor activity and values closer to 0.5 would
indicate similar recruitment from both muscles. Finally, online feedback was provided as the game
interactions described above. We asked participants to play their preferred combination of games
for at least one hour of training per session for a total of thirty sessions. To quantify functional
changes after training, we used the assessments described below.

4.3.4 Clinical Assessments (Sessions 1 and 32)
Two trained occupational therapists performed standard clinical assessments during the pre-
and post-training evaluations. Functional assessments were video recorded and independently
evaluated by both occupational therapists to ensure adequate inter-rater reliability. The complete
set of assessments used in this study included the following:
• Montreal Cognitive Assessment (MOCA). This is an assessment of mild cognitive
impairment that evaluates visuospatial abilities, memory, attention, concentration,
language, and orientation, and provides a score that ranges from 0 (greatest
impairment) to 30 (no impairment) (Nasreddine, Phillips, Bédirian, et al., 2005).
Typically, mild cognitive impairment is defined below 26 (Nasreddine, Phillips,
Bédirian, et al., 2005) or, more recently, 23 points (Carson et al., 2018). However,
other computer-based training paradigms have successfully recruited stroke
participants with mean scores as low as 20 points (Cramer et al., 2019; Ozen et al.,
2021; Yeh et al., 2019). Therefore, to broaden the pool of participants that could qualify
for our study, we used a 20-point threshold as inclusion criterion.
• Fugl–Meyer Assessment of the upper extremity (FMA). This scale measures
sensorimotor impairment of the upper limb following a hemiplegic stroke, including
movement, coordination, and reflexes, and provides a score that ranges from 0
(greatest impairment) to 66 (no impairment) (Fugl-Meyer et al., 1975).
77
• Action Research Arm Test (ARAT). This assessment measures functional
performance of the upper limb in terms of the ability to manipulate objects with different
sizes, weights, and shapes, and provides a score that ranges from 0 (greatest
impairment) to 57 (no impairment) (Lyle, 1981).
• Wrist range of motion (ROM). Using a goniometer, we recorded the maximum degrees
of active wrist extension and flexion. On average, activities of daily life usually require
40–60 degrees of wrist extension and 40–60 degrees of wrist flexion (Ryu et al., 1991;
van Andel et al., 2008).
• Modified Rankin Scale (MRS). This scale measures the disability of the stroke survivor
based on their independence to look after themselves in daily life, providing a range
from 0 (no symptoms of disability) to 5 (severe disability) (Swieten et al., 1988).
• Grip strength

4.3.5 Characterization of Muscle Control During EMG Amplitude Tracking (Sessions 1 and 32)
We sought to determine whether our task-specific training protocol induced generalized
functional changes in wrist muscle activity. To quantify changes in muscle control beyond task-
specific performance, participants completed two tracking tasks that were distinct from training
but required similar muscle activation. In the first task, participants were asked to maintain a
constant level of wrist extension of their more affected hand using the EMG amplitude from an
extensor muscle to follow a target of activity. As shown in Figure 4.1, this target corresponded to
a 4-second hold phase (plateau) of a trapezoid spanning six seconds and was set to 15% of the
tracked muscle’s maximal activity (established during a prerecorded power grip). Participants
completed twelve trials where EMG was rectified and smoothed with a 0.5-second moving window
to control the height of a cursor that moved left to right across the computer screen for ten seconds
before looping back. The second task followed the same protocol during wrist flexion. For each
78
tracking task, the muscle with the largest signal to noise ratio during voluntary activation was
chosen to provide EMG feedback, and practice trials were provided to ensure stable performance.
Importantly, this task is quasi-isometric for these participants as the required level of muscle
contraction resulted in little if any overt movement of the wrist. Participants were provided with a
low-stiffness ball to avoid uncomfortable curling of their fingers. We discouraged participants from
actively attempting to grip the ball, but unintended gripping was allowed. Participants rested their
more affected hand on a pillow for comfort for the duration of these tasks.
4.3.5.1 Data Acquisition
We measured surface EMG signals from four muscles at 2148 Hz using a Delsys Trigno
Wireless System (Delsys Incorporated, Natick, USA). EMG sensors were placed on the skin
above the extensor carpi radialis longus (ECR), extensor carpi ulnaris (ECU), flexor carpi radialis
(FCR), and flexor carpi ulnaris (FCU) of the more affected arm after cleaning the area with
isopropyl alcohol. Proper positioning was confirmed via muscle palpation and signal observation
during attempted wrist extension, flexion, radial and ulnar deviation, and light grip. EMG was
acquired using LabRecorder, an application designed to synchronize and record data using the
LSL protocol. Signals were down sampled to 1000 Hz for offline analysis.
Additionally, we concurrently recorded EEG at 500 Hz over the left and right motor cortices
using a 32-channel LiveAmp system (Brain Vision LLC, Morrisville, USA). Electrodes were
positioned over a subset of the 10-10 standard placement convention (Chatrian et al., 1985) using
the Brain Vision actiCAP. In this study, we only used EEG to assess corticospinal connectivity
changes via corticomuscular coherence after EMG-based training. Thus, we only analyzed the
electrodes corresponding to the frontal-central, central, and central parietal scalp locations (e.g.,
FC1, FC5, C3, CP1, CP5, C4, FC2, FC6, C4, CP2, and CP6), as these are on the areas
associated with motor control and correspond to the locations we used in our previous studies
(Marin-Pardo et al., 2020; Vourvopoulos, Pardo, et al., 2019). EEG signals were synchronized
and acquired using LSL and LabRecorder and then interpolated to 1000 Hz for offline analysis.
79
4.3.5.2 Muscle Group Individuation
We quantified the level of wrist muscle individuation with ratios of activation during the EMG
tracking task, i.e., the proportion of activity from the recorded extensor muscles to the total activity
from all recorded muscles using Equation 1. This allowed us to estimate the level of muscle group
recruitment during the extension and flexion tracking tasks. We used the last 3 seconds of each
hold phase to calculate ER values and evaluated changes in individuation before and after the
training sessions.
4.3.5.3 Corticomuscular Coherence
In addition to the calculations of muscle individuation and tracking error, we used the same
time epochs to assess the synchronization between brain and muscle signals. This measure,
known as corticomuscular coherence (CMC), is a frequency-domain correlation where a value of
0 indicates no correlation between signals and 1 indicates perfect correlation at a given frequency.
Specifically, we were interested in evaluating changes within the beta band (e.g., 12–30 Hz), as
this has been used to probe corticospinal communication, is frequently detected during static
muscle contractions, and has been suggested to improve during recovery (Krauth et al., 2019; J.
Liu et al., 2019; Marin-Pardo et al., 2020; Mima et al., 2001; Rossiter et al., 2013; von Carlowitz-
Ghori et al., 2014).
We calculated CMC in the same way as in our previous work (Marin-Pardo et al., 2020).
Briefly, EEG signals were first bandpass filtered between 5 and 100 Hz using a sixth order, zero-
phase Butterworth filter, and re-referenced to the common average after removing noisy or bad
channels identified using an artifact detection method within Matlab’s EEGLAB toolbox (Delorme
& Makeig, 2004; Kothe & Jung, 2016). We used channels C3 and C4 to calculate their coherence
with wrist EMG signals. If one of these channels had to be excluded, a neighboring electrode
corresponding to the respective sensorimotor hemisphere was used instead, as previous studies
have shown that CMC is not precisely localized during unimanual actions in this study population
(Krauth et al., 2019; Rossiter et al., 2013). EMG signals were first bandpass filtered between 15
80
and 450 Hz, then we used the Hilbert transform to obtain their amplitude envelope (T. W. Boonstra
& Breakspear, 2012; Farina et al., 2013). Finally, we normalized the resulting signals to have
zero-mean and unit variance (S. N. Baker, 2000; Halliday & Rosenberg, 2000). For each task,
pooled CMC (Amjad et al., 1997) was calculated between the pair of muscles involved in the task
(i.e., wrist flexors or wrist extensors) and the ipsilesional and contralesional hemispheres. All trial
epochs for a single subject were first concatenated, and then coherence was calculated using the
mscohere function in Matlab, using 512 ms Hann-windowed segments with 75% overlap. A 95%
confidence level for each coherence profile was calculated using Equation 2 (Rosenberg et al.,
1989):
CL = 1 − 0 .05
1
L − 1

where L is the number of segments used to calculate coherence, adjusted for tapering and
overlap (Kattla & Lowery, 2010). For a group-level analysis, we repeated the above procedure,
concatenating data from all five participants. This allowed us to use the same statistical methods
to evaluate group-level data as for individual coherence profiles (Amjad et al., 1997).

4.3.6 Statistical Analyses
We utilized custom scripts in Matlab (R2021a, The Mathworks, Natick, USA) and R (R
Foundation for Statistical Computing, Vienna, Austria) for offline signal processing and statistical
analyses. We anticipate that some people will improve across different metrics. However, we do
not expect that these changes will be consistent for the group, as previous literature from
ourselves and others have consistently shown that participants demonstrate variable
improvements after similar training interventions (Marin-Pardo et al., 2020; Mugler et al., 2019;
Ramos-Murguialday et al., 2013; Vourvopoulos, Pardo, et al., 2019). Therefore, we focus our
analyses on individual changes and report results at the group level for completeness.
81
4.3.6.1 Behavioral Changes in Clinical Assessments
Because the distribution of the data is not normally distributed, we used paired Wilcoxon
signed-rank tests to identify consistent changes across the group in FMA, ARAT, ROM, and grip
strength. As these assessments quantify impairments in different domains, we considered these
group-level tests as independent and thus significant at the p<0.05 level without correction.
Additionally, we used Pearson’s correlations to assess possible correlations between functional
improvements and time after stroke onset, setting a significance threshold at p<0.05.
4.3.6.2 Muscle Control Changes During EMG Amplitude Tracking
Using Equation 1, we calculated ratios of muscle-group activity to quantify the involvement
of both muscle groups while performing the extension and flexion tracking tasks. We calculated
mean ER values for each trial and evaluated differences between pre- versus post-training
assessments with signed-rank tests. Similarly, group-level effects were evaluated using paired
signed-rank tests of participants’ mean values for each task. The significance level for individual
and group-level differences was set at p<0.05.
4.3.6.3 Neuromuscular Changes Following Training
Similar to our previous work (Marin-Pardo et al., 2020), we evaluated changes in CMC at the
group level, with a Z-score difference of coherence (Laine et al., 2014; Rosenberg et al., 1989)
(pre versus post) at each frequency for each task (flexion and extension) and hemisphere
(ipsilesional and contralesional) using the formula in Equation 3:
Z
di f f
=
F z
po s t
− F z
pr e
√
1
2 L
po s t
+
1
2 L
pr e

where FZ is the Fisher-transformed coherence value (i.e., atanh(sqrt(coherence)) and L
represents the degrees of freedom, calculated as described for Equation 2. This provides a
standard Z score for the difference in coherence between sessions 1 and 32, for every frequency.
Then, we created a composite Z-score for the beta band using Stouffer’s Z-score method (Kilner
82
et al., 1999). Composite Z-scores with an absolute value above 1.96 are considered significant at
the 5% confidence level.
4.3.6.4 EMG Activity Across Training Sessions
To assess flexor and extensor amplitude changes over sessions, we took normalized EMG
signals and calculated the averaged amplitude for each muscle during game play. First, signals
were inspected manually (via visual inspection) and automatically (via examination of signal-to-
noise ratios) to remove files with poor signal quality. Records with acceptable signals were filtered,
rectified, and normalized as described for online processing, and trials containing activity periods
of at least 1 second for either muscle were concatenated. Then, we calculated each muscle’s
mean normalized amplitude for each session and used these values to calculate the respective
ER. Finally, we used Pearson correlations to evaluate changes in activity over time. We
considered correlations with p<0.05 as significant for individual and group tests.

4.4 Results
We evaluated feasibility, safety, and acceptability, as well as functional changes during pre-
vs post-training to characterize improvements induced by a 30-session training protocol.

4.4.1 Feasibility, Safety, and Acceptability
All participants completed more than 85% of the remote sessions (average: 90.7%  5.5%),
without adverse effects, including pain or general discomfort. All participants reported enjoying
using the system, that it was easy to use and that, having the chance, would like to continue using
it. On average, setup and calibration took less than ten minutes once the participants were familiar
with the system (e.g., after three sessions). Furthermore, all participants considered that having
two monitored sessions per week was sufficient (e.g., for social encouragement and technical
support), allowing them to successfully complete the independent sessions. Overall, all
83
participants reported having a positive experience. Some participants qualitatively noted modest
additional benefits in their arm function (e.g., more and faster movements), as well as reduced
muscle tone, increased hand sensation, increased attention to and use of the affected arm, and
better sleep quality. These qualitative observations were based on our interactions with the
participants and analysis of post-study semi-structured interview data. However, details of these
qualitative findings are beyond the scope of this manuscript and will be further discussed in future
work.

4.4.2 Behavioral Changes in Clinical Assessments
As seen in Table 4.2, all participants showed improvements in at least one of the five clinical
assessments measured. Figure 4.2 shows the most representative changes. At the individual
level, three participants showed improvement in their active extension (Participants 2, 4, and 5)
and active flexion (Participants 3, 4, and 5). Furthermore, three participants showed
improvements in the FMA and four in the ARAT. Notably, Participant 4 improved beyond the
minimum clinically important difference (MCID) in ARAT (difference of 16 points, MCID>5.7 (van
der Lee, de Groot, et al., 2001)) and FMA (difference of 5 points, MCID>4.25 (Page et al., 2012)).
However, no changes were statistically significant at the group level. Finally, we evaluated
whether participants with less time since stroke onset showed greater improvements (i.e., greater
change between pre- and post-intervention measurements). However, we did not find significant
correlations between time since onset and clinical measurements, which may be due to the small
sample size.

84
Table 4.2 Clinical assessments before and after training.
ARAT FMA ROM AE ROM AF Grip (kg)
Participant Pre Post Pre Post Pre Post Pre Post Pre Post
1 4 6 20 18 10 10 15 15 5.8 4.7
2 7 6 19 20 10 35 30 25 5.4 8.9
3 8 13 23 25 5 0 35 45 2.5 8.3
4 18 34 40 45 22 41 35 43 7.7 7.5
5 18 20 25 24 -15 10 0 10 7.3 5.3
Clinical assessment scores of 5 chronic stroke survivors before and after training with Tele-REINVENT.
Action research arm test (ARAT), Fugl–Meyer Assessment of the upper extremity (FMA), ranges of motion
(ROM) for active extension (AE), and active flexion (AF).

Figure 4.2 Clinical assessments before and after training.
Markers represent the score for each participant as evaluated in sessions 1 and 32. Action research arm
test (ARAT), Fugl-Meyer assessment of the upper extremity (FMA), active range of motion during wrist
extension. Additional results can be seen in Table 4.2.

4.4.3 Muscle Control Changes During EMG Amplitude Tracking
As seen in Figure 4.3, participants showed trends of improved motor control, measured by
greater muscle individuation for the extension and flexion tracking tasks (i.e., ER values closer to
1 for the extension task and closer to 0 for the flexion task). At the individual level, two participants
had a significant improvement in the extension task (p=0.004 for Participant 4, and p=0.033 for
85
participant 5). Similarly, four participants showed significant improvements in the flexion task
(p<0.01 for Participants 1, 2, 3, and 4). However, this change was not significant at the group
level for either task. Notably, all participants showed reduced variability (i.e., decreased standard
deviations) in both tracking tasks. Results of individual tests are shown in Table 4.3.

Figure 4.3 Muscle group individuation during EMG amplitude control.
Each panel represents the individual changes (Participants 1–5, left to right) of muscle group individuation
(quantified as the ratio of extensor to total activity (ER)) at constant levels of extension (top) and flexion
(bottom) tracking before and after 30 EMG training sessions. Notably, remote EMG training sessions did
not require a constant level of EMG activation and did not provide explicit error-based feedback. ER values
closer to 1 indicate higher extensor activity whereas values closer to 0 indicate higher flexor activity.
Improvements were seen in individuated recruitment for both tracking tasks (e.g., higher individuation and
lower variability). * indicates a significance of p<0.05, ** indicates significance of p<0.001. Statistical results
can be seen in Table 3.

86
Table 4.3 Muscle individuation during extension and flexion tracking.
Extension Flexion
Participant p ER Pre ER Post p ER Pre ER Post
1 0.243 0.47  0.04 0.49  0.03 <0.001* 0.55  0.01 0.12  0.01
2 0.291 0.44  0.22 0.37  0.03 <0.001* 0.61  0.07 0.47  0.04
3 0.946 0.42  0.11 0.46  0.04 0.479 0.27  0.13 0.28  0.06
4 0.004* 0.77  0.01 0.79  0.02 <0.001* 0.38  0.07 0.24  0.01
5 0.033* 0.53  0.07 0.59  0.05 <0.001* 0.32  0.15 0.56  0.05
Signed-rank tests for muscle group individuation of five chronic stroke survivors during an EMG extension
tracking task before and after training with Tele-REINVENT. Mean individuation (ER)  standard deviations
included for pre- and post-intervention evaluations. * and bold fonts indicate significance of p<0.05. ER
values closer to 1 indicate higher activation of extensor muscles while values closer to 0 indicate higher
activation of flexors. Values closer to 0.5 indicate a similar level of activation for both muscle groups.

4.4.4 Changes in Corticomuscular Coherence Following Training
Individual coherence profiles showed high variability across participants, frequencies, and
conditions. As seen in the top right panel of Figure 4.4, Participants 2, 4 and 5 showed significant
coherence peaks in the ipsilesional hemisphere before and after training. Similarly, significant
peaks after training are observed for Participants 1, 2, and 3 in the contralesional hemisphere.
Our analysis of pooled coherence across participants suggests that during wrist extension, the
only frequency band that had a significant increase of coherence after training was the beta band
of the contralesional cortex (Figure 4.4, top left panel, Z-score pre-post difference=2.35,
p=0.018). Importantly, this does not imply lack of activity from the ipsilesional cortex, as the pooled
coherence profiles show significant coherence before and after training. For completeness, we
also analyzed changes in the alpha and gamma frequency bands, where a non-significant trend
of increased coherence in the gamma band is seen for the ipsilesional cortex (Z-score
difference=1.68, p=0.093). As seen in the bottom right panel of Figure 4.4, we also identified
peaks of coherence after training in the beta band for the ipsilesional (Participants 2 and 5) and
87
contralesional hemispheres (Participants 2, 3, and 5) during the flexion task, but these changes
were not consistent according to the pooled coherence analysis. We detected a non-significant
trend in the beta band in the contralesional hemisphere (Figure 4.4, bottom left panel, Z-score
pre-post difference=1.55, p=0.121).

Figure 4.4 Corticomuscular coherence (CMC) during static extension and flexion.
Left panels show the group-level pooled coherence for extension (top) and flexion (bottom) tasks, with
activity pre-REINVENT in lighter colors and post-REINVENT in darker colors. Note that significant
coherence is present during extension (top) before and after training in the beta band at the ipsilesional
hemisphere, and that there is a significant increase in the contralesional hemisphere. Similarly, beta band
coherence during flexion (botom) seems to shift from the ipsilesional towards the contralesional
hemisphere. Middle panels show bar plots of pooled-coherence spectra and represent the composite
group difference in coherence before versus after training within alpha (8–12 Hz), beta (12–30 Hz), and
gamma (30–50 Hz) frequency bands. Asterisks denote significant changes in a frequency band before
versus after training. Right panels show individual profiles (Participants 1–5, left to right) of coherence for
extension (top) and flexion (bottom) tasks. Asterisks in individual plots note participants that showed
significant CMC in the beta band before or after training. Top rows of each panel show plots for ipsilesional
electrodes and bottom rows show contralesional electrodes. Coherence spectra within 0 and 60 Hz are
shown in all plots, including vertical dashed lines indicating boundaries of the beta band and a solid
horizontal line to indicate significant coherence.
88
4.4.5 Increased EMG Individuation Across Sessions
As shown in Figure 4.5, measures of muscle activity suggest higher individuation over time
(top row). Overall, muscle individuation appears to increase during training for Participants 3, 4,
and 5. These changes were significant for Participants 4 (p=0.018) and 5 (p=0.013). These
changes were accompanied by a significant decrease of flexor activity for Participants 4 (p=0.009)
and 5 (0.016). At the group level, we observed a significant decrease of overall flexor activity
(p=0.045). Table 4.4 shows correlation results for all five participants.

Figure 4.5 Muscle activity during remote training sessions.
Each column (1–5) represents per-session averages (markers) of muscle activity (rows) during training for
each participant. (Top) Muscle individuation (ER) where values closer to 1 indicate higher extension activity
and values closer to 0 indicate higher flexion activity. (Middle) Normalized extensor electromyography
(EMG) activity. (Bottom) Normalized flexor EMG. Overall, these plots suggest modest trends of improved
individuation over time (top row). Changes of ER over time showed a significant increase for Participants 4
(p=0.018) and 5 (p=0.013). This change was accompanied by a decrease of flexor activity over time
(p=0.009 for Participant 4 and p=0.016 for Participant 5). Best fit lines are included to visualize trends across
sessions in gray.
89
Table 4.4 Changes of muscle activity over time during remote training sessions.
ER Extensor Flexor
Participant rho p rho p rho p
1 0.02 0.937 0.12 0.607 0.12 0.587
2 -0.15 0.536 -0.38 0.100 -0.18 0.448
3 0.19 0.502 0.23 0.417 -0.13 0.642
4 0.49 0.018* -0.19 0.389 -0.53 0.009*
5 0.53 0.013* 0.06 0.807 -0.52 0.016*
Pearson correlations of five chronic stroke survivors during EMG biofeedback training with Tele-
REINVENT. Bold fonts and * indicate significance of p<0.05.

4.5 Discussion
We explored the use of an EMG-based telerehabilitation program that attempts to improve
severe motor deficits after chronic stroke via training of muscle activation without unintended
activation of antagonistic muscles. We found that 30 1-hr sessions of combined supervised and
unsupervised remote training can induce positive outcomes in motor function in a pilot study with
five stroke survivors. Although variable across participants, similar to our previous studies (Marin-
Pardo et al., 2020; Vourvopoulos, Pardo, et al., 2019), these results suggest that our
telerehabilitation system can elicit functional changes in severely impaired individuals. These
changes were measurable with standard clinical assessments, the proportion of activation of
agonistic and antagonistic muscles, and modest improvements in corticomuscular connectivity.
Additionally, all participants reported an overall positive experience and showed high adherence
to the proposed training protocol.

4.5.1 Clinical Assessments
We hypothesized that we would observe improvements in clinical assessments after training.
Much effort has been invested in developing novel and effective non-invasive rehabilitation
90
treatments. However, previous research has shown that, even for treatments with proven efficacy
and narrow inclusion criteria, not all participants will positively respond to interventions (Kwakkel
et al., 2016; Ward et al., 2019). Accordingly, more investigation is needed to better understand
which elements (e.g., population characteristics and neural mechanisms engaged in repeated
practice) could better predict which participants might benefit from targeted rehabilitation
approaches (Bernhardt et al., 2017; Winstein et al., 2016). Similarly, we expected that some, but
not all, of our participants would show improvements in clinical assessments. Subsequently, due
to our limited sample size, we expected that inconsistent changes (i.e., not seen across all
participants) would result in non-significant changes at the group level. Furthermore, previous
research has suggested that individuals presenting with less time after stroke onset might show
greater improvement after training interventions (Ballester et al., 2019). While that could also be
the case for our proposed training, we did not make a specific hypothesis about it due to the small
sample of this feasibility study.
In previous work, we showed that few sessions (n=7) of wrist training to encourage extension
movements without unintended flexor recruitment produced moderate improvements in several
clinical assessments (Marin-Pardo et al., 2020). Our results from that experiment and this new
iteration had similar trends, i.e., improvement across different assessments and some participants
showing changes beyond established MCIDs. Qualitatively, our current results suggest that
higher training dosage might have induced greater functional changes. This can be observed in
greater differences between pre- and post-training measurements, compared across participants
that initially had similar levels of impairment. On average, ARAT and wrist extension changes
were greater for those who participated in 30 remote training sessions, compared to those who
participated in 7 laboratory-based sessions. However, we did not find a statistically significant
difference between the two groups (i.e., a significant correlation between level of improvement
and completed training sessions). As we lack the statistical power to explore such relationships,
91
further research with larger samples is necessary to better quantify the effect that training dosage
might have induced on functional outcomes.
Literature suggests that relatively high dosages of repeated task-specific practice can induce
positive outcomes when using EMG-based technologies after chronic stroke (Jian et al., 2021;
Mugler et al., 2019). In line with our results, Mugler et al. and Jian et al. showed that training to
reduce abnormal muscle coactivation with similar dosage and intensity could reduce impairment
after chronic stroke with comparable results to ours. Importantly, this level of exercise is already
higher than what patients receive during standard clinical practice, as sessions are often limited
by the time a clinician can spend on a specific limb or joint (Lang et al., 2009). However, further
research with larger samples is required to allow stratification of different levels of dosage and
intensity to evaluate the effect that these parameters have on the level of improvement that EMG
biofeedback can induce.
It is important to note that although our training paradigm sought to specifically train wrist
movements, we saw functional changes beyond the activation of the muscles we used for training.
As expected, most participants improved their active ranges of motion during wrist extension and
flexion. However, changes in FMA and ARAT scores are not fully explained only by improved
recruitment of hand muscles, as some participants showed improvement on assessment items
that required coordination of the whole upper limb. Results from our tracking task suggest that
wrist extension training also allowed for more individuated control during the flexion task (see
below). Therefore, we theorize that strategies our participants learned to produce isolated
movement of the wrist could have also been applied to improve control of other muscle pairs,
improving their overall upper limb control. However, further research is required to evaluate
whether learning to better control a specific muscle group correlates with improved control of
untrained muscles.
Finally, it is widely accepted that most spontaneous biological recovery plateaus within the
first six months after stroke onset and that improvement in motor outcomes beyond this time
92
window might be mostly driven by learning-dependent processes or compensation strategies
(Bernhardt et al., 2017; Kwakkel et al., 2004). Therefore, although we did not specifically quantify
spontaneous improvement with multiple baseline measurements, it is reasonable to assume that
changes in motor function could be attributed to our proposed training. Future work should include
multiple baseline measurements to confirm that, at enrollment, participants’ recovery has indeed
plateaued. Furthermore, the time after stroke for when no additional recovery can be expected is
currently unknown. As noted above, increasing evidence shows that it is possible to induce
positive motor outcomes even years after the cerebrovascular accident. However, it is likely that
the efficacy of such approaches is correlated with the time after onset, allowing for greater
improvements for people presenting in the more acute and subacute phases (Ballester et al.,
2019). Similarly, it is possible that such populations might have benefitted more from our proposed
program, and it would be valuable to explore this in future work. We did not, however, have a
specific hypothesis regarding possible correlations between stroke onset and functional
improvement in this study, as our sample limits such analysis.

4.5.2 Improvements in Muscle Group Individuation
Our next hypothesis was that our training program would improve neuromuscular control,
quantified as increased muscle recruitment individuation. Accordingly, we showed that most
participants improved their muscle individuation (measured as the extension ratio, ER) during in-
person evaluations and remote training. First, most participants showed trends of increased
individuation during the tracking task. This was seen not only as increased ER values during the
extension tracking and decreased ER values during flexion tracking but also as overall decreased
ER variability during both tracking tasks. Importantly, our training paradigm was designed to
encourage extension movements via reinforcement of extension-like attempts that did not present
significant flexion activation. That is, ER values that corresponded to those expected during
attempted individuated extension (closer to 1) showed positive feedback that corresponded with
93
the game goal (e.g., correct control commands or higher scores). Motor control literature refers
to this as reinforcement learning, i.e., learning to produce successful motor commands by
attempting to increase positive rewards (Krakauer, 2006; Maier et al., 2019). However, during the
EMG amplitude tracking assessments, we provided an instantaneous difference between
expected and produced values of EMG. This required a different control strategy that used error
cues to correct behavior (i.e., error-based learning (Krakauer, 2006; Maier et al., 2019)).
Furthermore, although we did not provide feedback of the concurrent activation of the antagonist
muscles during the tracking task, participants showed increased individuation in post-training
sessions for both extension and flexion tasks. Overall, this suggests that we may have improved
generalized behavior as the training and tracking tasks required similar activation patterns but
were distinct in nature. However, further research is required to disambiguate changes in error-
based and reinforcement-based strategies.
Additionally, some participants showed increased individuation over time during remote
training sessions and this change was statistically significant for two participants. Importantly, this
change was accompanied by a significant decrease of flexor activity over time and not increased
activity of extensor muscles. Ellis and others (Ellis et al., 2017) showed that the flexion synergy,
commonly seen in impaired stroke populations, is detrimental for reaching movements and that
specifically targeting this impairment might be beneficial for arm function. Similarly, our results
suggest that reducing activity of antagonist muscles (e.g., flexor muscles during our extension
training protocol) may have a higher influence in increased muscle group individuation than
increasing activity of agonist muscles. However, further research is needed to confirm this trend.

4.5.3 Changes in Corticomuscular Coherence
In addition to seeing improvements in clinical assessments and muscle individuation during
movement attempts, we also hypothesized that these changes would be accompanied by
neuronal reorganization. We used CMC, a frequency-domain quantification of the synchrony
94
between EEG and EMG signals, to probe such neuroplastic changes. CMC is typically interpreted
as an indication of functional connectivity between neurons in the motor cortex and the motor
neurons in the spinal cord (T. Boonstra, 2013; J. Liu et al., 2019; Mima & Hallett, 1999).
Importantly, previous literature has identified inter-subject variability in this measurement in
impaired and neurotypical populations (Chwodhury et al., 2015; Rossiter et al., 2013). However,
CMC in the beta and gamma bands (12–30 Hz and 30–50 Hz, respectively) has been shown to
increase due to spontaneous and rehabilitation-induced recovery after stroke (Krauth et al., 2019;
von Carlowitz-Ghori et al., 2014; Zheng et al., 2018). Furthermore, although CMC is typically
located in the primary motor area contralateral to the muscle used to record EMG in neurotypical
populations (Kilner et al., 1999; J. Liu et al., 2019; Rossiter et al., 2013), studies with post-stroke
populations have shown variable localization sources, including the motor and premotor areas of
both the ipsilesional and contralesional hemispheres (Krauth et al., 2019; Rossiter et al., 2013).
Moreover, the inherent variability of lesion location among stroke survivors might introduce
additional variability in the source localization of CMC (Rossiter et al., 2013). In this work, we
analyzed CMC using electrodes over our participants’ ipsilesional and contralesional motor
cortices. Given our small sample and because we do not have data regarding the specific lesion
locations of our participants, we chose such electrodes to better compare our current and previous
findings (as discussed below). However, further research with larger samples and lesion location
data (e.g., including anatomical magnetic resonance imaging) would be necessary to better
understand possible correlations between CMC and lesion location.
Overall, our results showed variable changes in CMC across individuals during static holds of
wrist flexion and extension. We observed significant beta band coherence in the ipsilesional
hemisphere before and after training during the extension task, shown in individual and group
coherence plots. Furthermore, we observed a significant increase in beta band coherence with
the contralesional cortex, also shown in individual and group plots. Finally, we observed what
appears to be a shift in coherence from the ipsilesional to the contralesional cortex during the
95
flexion task; however, this change was not statistically significant. In line with our previous work
(Marin-Pardo et al., 2020), we found a significant group increase in the beta band in the
contralesional hemisphere after training. However, this change was not accompanied with the
increase in the ipsilesional hemisphere we observed in our previous work. Differences between
our previous and current results could be driven by several factors (e.g., demographics or training
dosage). Although we aimed to include a similar population, the participants in our previous study
presented with shorter times after stroke onset and, on average, more severe impairment. This
could partially explain why our current participants showed significant coherence in the
ipsilesional hemisphere before training and only showed increased coherence in the
contralesional hemisphere. Additionally, our previous work explored the effects of relatively few
training sessions, whereas here we tripled the number of sessions, which may have induced
greater plasticity in CMC of the contralesional hemisphere. However, more research is needed to
better understand the observed changes. Together, our results suggest that CMC may be used
to quantify changes induced by EMG biofeedback training, and such changes may be mediated
by the contralesional cortex both at early and late stages of training with accompanied ipsilesional
contribution at early stages. However, further research with a larger sample is necessary to
investigate possible correlations between CMC changes and other factors, such as, baseline
impairment, training dosages, induced recovery, and training tasks.

4.5.4 Limitations and Future Directions
A main limitation of this feasibility case series study is the small sample size. Future work
employing remote assessments (Amano et al., 2018; Palsbo et al., 2007), in addition to remote
training, would likely increase our sample size as the accessibility for people with mobility
limitations may be improved. Future work could also engage a larger and more diverse sample
that includes broader levels of impairment to fully characterize the recovery process potentially
induced by Tele-REINVENT. A larger sample would also allow disambiguating potential
96
correlations between improved outcomes, key elements of the intervention (e.g., dosage,
intensity, and learning strategy), and demographics (e.g., baseline impairment, lesion location,
lesion size, time after onset, and spasticity).

97
Chapter 5: Discussion
Measurement and feedback of electrical muscle activity via electromyography (EMG) may
improve upper limb movement after severe chronic stroke when used to train to avoid unintended
coactivation of antagonistic muscle pairs. However, the efficacy of this approach is not fully
understood. Specifically, it is unknown how key elements such as feedback environment and
dosage might affect the efficacy of the intervention. The main goal of this dissertation is to assess
EMG biofeedback delivered in virtual environments with varying feedback immersion (e.g., virtual
reality (VR)-based and screen-based environments) and dosage levels (e.g., after three and six
weeks of training) to explore how they may affect the functional improvement of severely impaired
chronic stroke survivors. A deeper understanding of how these factors affect functional outcomes
would allow us to develop personalized and effective interventions for underserved populations.

5.1 Summary of Key Findings
5.1.1 Chapter 2 Summary – Aim One
In Chapter 2, I aimed to develop and evaluate the feasibility of an EMG-based VR biofeedback
rehabilitation system to increase volitional wrist muscle activity while reducing unintended
antagonistic co-contractions. We recruited four severely impaired participants in the chronic stage
of stroke recovery and asked them to complete seven training sessions where our biofeedback
system reinforced individuated muscle activity via a computer game in VR. We quantified
functional improvements with behavioral assessments (e.g., Fugl-Meyer Assessment of the upper
extremity (FMA), Action Research Arm Test (ARAT), and wrist Ranges of Motion (ROM)), tests
of muscle control (using EMG amplitude from the more-affected arm to follow a target of muscle
activity), and a proxy of brain-muscle connectivity (corticomuscular coherence during movement
attempt). Although changes varied across our pilot sample, our results suggest that this training
98
paradigm may improve upper limb movement in severely impaired stroke survivors more than six
months after stroke onset.

5.1.2 Chapter 3 Summary – Aim Two
Our findings in Aim 1, i.e., that few sessions of reinforced individuated muscle activity may
elicit positive changes in motor control, prompted us to examine the effects of increased training
dosages. However, as in-person research was limited due to sanitary precautions during the
onset of the COVID-19 pandemic, in Chapter 3, I aimed to develop a low-cost system for home-
based EMG biofeedback and test the feasibility of using it in a training program for stroke
telerehabilitation. We built the necessary programs to process EMG signals acquired with low-
cost sensors and provide feedback of muscle activity via computer video games. We provided
examples of using these sensors with neurotypical participants in a laboratory setting during
isometric movement trials. We showed that, for our research purposes, we could obtain
measurements comparable to those acquired with research-grade sensors. Furthermore, as this
work aimed to implement the proposed system for stroke telerehabilitation, we tested the
feasibility of using it during remotely supervised and unsupervised home-based training sessions
with one stroke survivor. We found that our system was safe and easy to use during several
weeks of training. Together, our results suggest that it is feasible to use low-cost technology to
estimate and encourage individuated muscle activity in in-person and remote environments.

5.1.3 Chapter 4 Summary – Aim Three
We learned in Aims 1 and 2 that using muscle activity feedback in laboratory-based training
may induce positive changes in functional outcomes and that higher dosages of this paradigm
could be implemented from participants’ homes. Therefore, in Chapter 4, I aimed to determine
the efficacy of using a home-based EMG biofeedback system to improve upper limb function via
telerehabilitation training encouraging wrist movements executed without unintended muscle
99
coactivation. We recruited five stroke survivors in the chronic phase of recovery and asked them
to complete thirty sessions of home training using the system we developed in Aim 2. We used
clinical assessments (e.g., FMA, ARAT, and wrist ROM), tests of muscle control (EMG amplitude
from the more-affected arm to follow a target of muscle activity), and a proxy of brain-muscle
connectivity (corticomuscular coherence) to quantify their motor function before and after training.
We confirmed that our system was safe and easy to use across several weeks of training.
Furthermore, although changes were variable across our sample, participants showed functional
improvements across different assessments, and one had improvements beyond minimum
clinically important differences. Therefore, our results suggest that training to avoid unintended
coactivation of antagonistic muscles via EMG telerehabilitation might be an effective intervention
to improve upper limb function for some chronic stroke survivors.

5.2 Implications and Significance
5.2.1 Clinical Implications
Most of the currently available and effective non-invasive therapeutic treatments to improve
upper limb function after stroke focus on moderately and mildly impaired populations in the acute
and subacute stages of recovery. Therefore, there is an urgent need for targeted interventions for
chronic stroke survivors, especially for those with minimal overt movement. As shown above,
EMG biofeedback may be a feasible alternative for this underserved population. However, gaining
a better understanding of how to better personalize the intervention is of utmost importance since
this might affect its efficacy.
In this work, we have shown that our training paradigm can induce modest positive changes
in functional assessments of upper limb movement (e.g., FMA, ARAT, and wrist ROM). At the
group level, that is, taking together the results of the pilot experiments in Aim1 and Aim 3, we
observed trends of increased active wrist extension, FMA, and ARAT. However, the nature of
these changes may depend on specific components of the intervention, for example, feedback
100
environment and dosage. In Aim 1 we explored the use of relatively few sessions of EMG
biofeedback in an immersive VR environment (i.e., using a head-mounted display), whereas, in
Aim 3 we increased the number of sessions by a factor of four and provided feedback in a non-
immersive environment (i.e., using a computer screen). This might have influenced the specific
benefits each group showed, as the group that experienced a VR environment showed greater
changes in FMA, whereas the group that had more sessions showed greater improvements in
ROM. However, more research with greater samples is necessary to disambiguate how these
and other factors may interact in this intervention and how to make it more efficacious.
It is important to note that while few participants improved beyond the accepted levels of
minimum clinically important differences in FMA and ARAT, as a group, the average wrist
extension increased from 6.5 to 15.3 degrees. This is of great relevance because less than 10
degrees of wrist extension is a common inclusion criterion for other non-invasive treatments (e.g.,
constrained-induced movement therapy (Winstein et al., 2003; Wolf et al., 2010)). Therefore,
some participants that we recruited for the experiments in Aim 1 and Aim 3, that would have been
initially disqualified from participating in such treatments are now in a state where they could
benefit from more traditional and effective interventions.

5.2.2 Research and Technological Implications
The ability to acquire data and monitor participation in therapeutic interventions and research
experiments remotely has received widespread attention over the last few years. This was more
evident with the onset of the COVID-19 pandemic, which highlighted the need for systems and
procedures to safely provide uninterrupted healthcare services and continued experimentation
beyond hospitals and research laboratories. As much of the world came to a halt due to safety
measures that hindered in-person activities, our research group worked to develop a mobile
platform that could bring REINVENT from our laboratory to our participant’s homes. This effort
101
allowed us to continue our research activities and file a patent application currently under review
(Liew et al., n.d.).
From a technological standpoint, as computers and their peripheral devices continue to get
more power and sophistication at an exponential rate, it is of great interest to have the ability to
incorporate novel acquisition systems, processing pipelines, and rendering devices quickly and
efficiently. Accordingly, the modular nature of our system allows us to update only necessary
components to tailor them to our current and future needs. Since our acquisition, processing, and
feedback algorithms are connected but work independently, incorporating novel devices,
calculations, and environments would only require developing the appropriate interfaces without
necessitating rewriting or adapting the rest of the working code. This would be of great benefit
should we explore feedback modalities that require devices and tasks that are not currently
integrated into the mobile REINVENT platform. For example, reincorporating immersive VR
environments, expanding our current collection of tasks and games, and reincorporating
research-grade sensors to investigate motor learning mechanisms that may interact in our training
paradigm.

5.3 Limitations and Future Directions
The main limitation of this dissertation is the sample size in all pilot experiments. Our results
show that stroke survivors had positive changes in functional outcomes after participating in our
protocols. We attribute this to the use of our training system since it is generally assumed that
such changes are unlikely to happen spontaneously in our targeted population (Langhorne et al.,
2011). However, our sample size limits our ability to generalize these findings to broader
populations. Therefore, further research with greater samples is necessary to confirm whether
similar improvement patterns are seen in stroke survivors with varying impairment levels and at
earlier recovery stages. Additionally, increased samples would allow for a more robust
characterization of unexplored covariates that may affect the efficacy of the intervention. Of
102
particular interest would be to investigate whether personal factors such as demographics (e.g.,
age, computer literacy, baseline impairment, and time since stroke onset), lesion characteristics
(e.g., location and size), or psychological variables (e.g., sense of embodiment, agency, and
usability) interact with this training paradigm to induce greater functional benefits. At the time of
writing this dissertation, we are continuing data collection and we expect that Tele-REINVENT will
continue to be used in future large-scale studies.
Another area that would be beneficial to explore is how different neural mechanisms and
learning processes could interact to make our gamified EMG biofeedback training more
efficacious. For example, previous research has shown that some people can learn skills in both
screen-based and immersive environments and that, for some, it is easier to acquire a new skill
in the non-immersive option (Juliano et al., 2020). It has been suggested that cognitive load (i.e.,
how much information is retained in working memory) may interfere with acquiring new skills in
immersive environments (Juliano et al., 2022). This should be further explored in stroke survivors
as it could hinder their ability to regain motor control if feedback is provided in such a fashion.
In this work, we showed that even though our training task was developed with an approach
based on reinforcement learning mechanisms (i.e., encouraging behavior that provided increased
rewards), participants also showed improvements in a task that depended on supervised learning
(i.e., attempting behavior that decreased performance errors). Our findings suggest that strategies
our participants used in the training paradigm (via reinforcement learning) might have been
transferred to the supervised context (error-based task). Therefore, it would be valuable to explore
the possible interaction between these two learning mechanisms and develop training
experiences that exploit the benefits that both paradigms can provide.
Finally, as discussed above, the modularity of the REINVENT platform favors the exploitation
of novel technologies, such as acquisition devices and processing algorithms. Our findings
suggest that using EMG biofeedback to avoid unintended coactivation might be beneficial for our
target population. However, developing additional training tasks with distinct therapeutic goals
103
(e.g., aiming to increase strength, reduce tremor, or target different muscle groups) might make
our proposed system a feasible option for a wide variety of neurological conditions (e.g.,
Parkinson’s disease or Alzheimer’s disease).
Overall, the work presented in this dissertation shows that human-computer interfaces may
be a feasible and promising alternative for personalized upper limb rehabilitation after stroke.
Moreover, it capitalizes on the added functionality of having modular systems to enhance
development and adaptation to different environments and rehabilitation goals. Finally, it further
supports that new and exciting opportunities can be exploited when we shift the question “does
this novel treatment work?” towards “whom this new treatment might benefit most?”. It is our hope
that this seminal work will allow the expansion of using inexpensive and customizable technology
to aid clinicians in improving people’s quality of life after stroke.

104
References
Alkoby, O., Abu-Rmileh, A., Shriki, O., & Todder, D. (2018). Can We Predict Who Will Respond
to Neurofeedback? A Review of the Inefficacy Problem and Existing Predictors for
Successful EEG Neurofeedback Learning. Neuroscience, 378, 155–164.
https://doi.org/10.1016/j.neuroscience.2016.12.050

Amano, S., Umeji, A., Uchita, A., Hashimoto, Y., Takebayashi, T., Kanata, Y., Uchiyama, Y., &
Domen, K. (2018). Reliability of remote evaluation for the Fugl–Meyer assessment and
the action research arm test in hemiparetic patients after stroke. Topics in Stroke
Rehabilitation, 1–6. https://doi.org/10.1080/10749357.2018.1481569

Amjad, A. M., Halliday, D. M., Rosenberg, J. R., & Conway, B. A. (1997). An extended difference
of coherence test for comparing and combining several independent coherence estimates:
Theory and application to the study of motor units and physiological tremor. Journal of
Neuroscience Methods, 73(1), 69–79. https://doi.org/10.1016/S0165-0270(96)02214-5

Andrade, A. O., & Andrade, C. I. (2012). On the relationship between features extracted from
EMG and force for constant and dynamic protocols. 2012 Annual International Conference
of the IEEE Engineering in Medicine and Biology Society, 3392–3395.
https://doi.org/10.1109/EMBC.2012.6346693

Armagan, O., Tascioglu, F., & Oner, C. (2003). Electromyographic Biofeedback in the Treatment
of the Hemiplegic Hand: A Placebo-Controlled Study. American Journal of Physical
Medicine & Rehabilitation, 82(11), 856–861.
https://doi.org/10.1097/01.PHM.0000091984.72486.E0

Arpa, S., & Ozcakir, S. (2019). Does electromyographic biofeedback improve exercise effects in
hemiplegic patients? A pilot randomized controlled trial. Journal of Rehabilitation
Medicine, 51(2), 109–112. https://doi.org/10.2340/16501977-2513

Bai, Z., Fong, K. N. K., Zhang, J. J., Chan, J., & Ting, K. H. (2020). Immediate and long-term
effects of BCI-based rehabilitation of the upper extremity after stroke: A systematic review
and meta-analysis. Journal of NeuroEngineering and Rehabilitation, 17(1), 57.
https://doi.org/10.1186/s12984-020-00686-2

Baker, M. R., & Baker, S. N. (2003). The effect of diazepam on motor cortical oscillations and
corticomuscular coherence studied in man. The Journal of Physiology, 546(3), 931–942.
https://doi.org/10.1113/jphysiol.2002.029553

Baker, S. N. (2000). “Pooled coherence” can overestimate the significance of coupling in the
presence of inter-experiment variability. Journal of Neuroscience Methods, 96(2), 171–
172. https://doi.org/10.1016/s0165-0270(00)00165-5

Ballester, B. R., Maier, M., Duff, A., Cameirão, M., Bermúdez, S., Duarte, E., Cuxart, A.,
Rodríguez, S., San Segundo Mozo, R. M., & Verschure, P. F. M. J. (2019). A critical time
window for recovery extends beyond one-year post-stroke. Journal of Neurophysiology,
122(1), 350–357. https://doi.org/10.1152/jn.00762.2018

105
Beer, R. F., Dewald, J. P. A., & Rymer, W. Z. (2000). Deficits in the coordination of multijoint arm
movements in patients with hemiparesis: Evidence for disturbed control of limb dynamics.
Experimental Brain Research, 131(3), 305–319. https://doi.org/10.1007/s002219900275

Belardinelli, P., Laer, L., Ortiz, E., Braun, C., & Gharabaghi, A. (2017). Plasticity of premotor
cortico-muscular coherence in severely impaired stroke patients with hand paralysis.
NeuroImage: Clinical, 14, 726–733. https://doi.org/10.1016/j.nicl.2017.03.005

Bernhardt, J., Hayward, K. S., Kwakkel, G., Ward, N. S., Wolf, S. L., Borschmann, K., Krakauer,
J. W., Boyd, L. A., Carmichael, S. T., Corbett, D., & Cramer, S. C. (2017). Agreed
definitions and a shared vision for new standards in stroke recovery research: The Stroke
Recovery and Rehabilitation Roundtable taskforce. International Journal of Stroke, 12(5),
444–450. https://doi.org/10.1177/1747493017711816

Boonstra, T. (2013). The potential of corticomuscular and intermuscular coherence for research
on human motor control. Frontiers in Human Neuroscience, 7.
https://www.frontiersin.org/article/10.3389/fnhum.2013.00855

Boonstra, T. W., & Breakspear, M. (2012). Neural mechanisms of intermuscular coherence:
Implications for the rectification of surface electromyography. Journal of Neurophysiology,
107(3), 796–807. https://doi.org/10.1152/jn.00066.2011

Borrego, A., Latorre, J., Alcañiz, M., & Llorens, R. (2019). Embodiment and Presence in Virtual
Reality After Stroke. A Comparative Study With Healthy Subjects. Frontiers in Neurology,
10. https://www.frontiersin.org/article/10.3389/fneur.2019.01061

Braun, C., Staudt, M., Schmitt, C., Preissl, H., Birbaumer, N., & Gerloff, C. (2007). Crossed
cortico-spinal motor control after capsular stroke. European Journal of Neuroscience,
25(9), 2935–2945. https://doi.org/10.1111/j.1460-9568.2007.05526.x

Brown, S. H. M., Brookham, R. L., & Dickerson, C. R. (2010). High-pass filtering surface EMG in
an attempt to better represent the signals detected at the intramuscular level: High-Pass
Filtering Surface EMG. Muscle & Nerve, 41(2), 234–239.
https://doi.org/10.1002/mus.21470

Bundy, D. T., Souders, L., Baranyai, K., Leonard, L., Schalk, G., Coker, R., Moran, D. W., Huskey,
T., & Leuthardt, E. C. (2017). Contralesional Brain–Computer Interface Control of a
Powered Exoskeleton for Motor Recovery in Chronic Stroke Survivors. Stroke, 48(7),
1908–1915. https://doi.org/10.1161/STROKEAHA.116.016304

Carson, N., Leach, L., & Murphy, K. J. (2018). A re-examination of Montreal Cognitive
Assessment (MoCA) cutoff scores: Re-examination of MoCA cutoff scores. International
Journal of Geriatric Psychiatry, 33(2), 379–388. https://doi.org/10.1002/gps.4756

Carter, G. C. (1987). Coherence and time delay estimation. Proceedings of the IEEE, 75(2), 236–
255. https://doi.org/10.1109/PROC.1987.13723

Celnik, P., Webster, B., Glasser, D. M., & Cohen, L. G. (2008). Effects of Action Observation on
Physical Training After Stroke. Stroke, 39(6), 1814–1820.
https://doi.org/10.1161/STROKEAHA.107.508184

106
Cervera, M. A., Soekadar, S. R., Ushiba, J., Millán, J. del R., Liu, M., Birbaumer, N., & Garipelli,
G. (2018). Brain-computer interfaces for post-stroke motor rehabilitation: A meta-analysis.
Annals of Clinical and Translational Neurology, 5(5), 651–663.
https://doi.org/10.1002/acn3.544

Chatrian, G. E., Lettich, E., & Nelson, P. L. (1985). Ten Percent Electrode System for Topographic
Studies of Spontaneous and Evoked EEG Activities. American Journal of EEG
Technology, 25(2), 83–92. https://doi.org/10.1080/00029238.1985.11080163

Chen, Y., Abel, K. T., Janecek, J. T., Chen, Y., Zheng, K., & Cramer, S. C. (2019). Home-based
technologies for stroke rehabilitation: A systematic review. International Journal of Medical
Informatics, 123, 11–22. https://doi.org/10.1016/j.ijmedinf.2018.12.001

Chwodhury, A., Raza, H., Dutta, A., Nishad, S. S., Saxena, A., & Prasad, G. (2015). A study on
cortico-muscular coupling in finger motions for exoskeleton assisted neuro-rehabilitation.
2015 37th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), 4610–4614. https://doi.org/10.1109/EMBC.2015.7319421

Cortes, J. C., Goldsmith, J., Harran, M. D., Xu, J., Kim, N., Schambra, H. M., Luft, A. R., Celnik,
P., Krakauer, J. W., & Kitago, T. (2017). A Short and Distinct Time Window for Recovery
of Arm Motor Control Early After Stroke Revealed With a Global Measure of Trajectory
Kinematics. Neurorehabilitation and Neural Repair, 31(6), 552–560.
https://doi.org/10.1177/1545968317697034

Cramer, S. C., Dodakian, L., Le, V., See, J., Augsburger, R., McKenzie, A., Zhou, R. J., Chiu, N.
L., Heckhausen, J., Cassidy, J. M., Scacchi, W., Smith, M. T., Barrett, A. M., Knutson, J.,
Edwards, D., Putrino, D., Agrawal, K., Ngo, K., Roth, E. J. (2019). Efficacy of Home-Based
Telerehabilitation vs In-Clinic Therapy for Adults After Stroke: A Randomized Clinical Trial.
JAMA Neurology, 76(9), 1079. https://doi.org/10.1001/jamaneurol.2019.1604

de Noordhout, A. M., Rapisarda, G., Bogacz, D., Gérard, P., De Pasqua, V., Pennisi, G., &
Delwaide, P. J. (1999). Corticomotoneuronal synaptic connections in normal man: An
electrophysiological study. Brain, 122(7), 1327–1340.
https://doi.org/10.1093/brain/122.7.1327

Delorme, A., & Makeig, S. (2004). EEGLAB: An open source toolbox for analysis of single-trial
EEG dynamics including independent component analysis. Journal of Neuroscience
Methods, 134(1), 9–21. https://doi.org/10.1016/j.jneumeth.2003.10.009

Dewald, J. P. A., Pope, P. S., Given, J. D., Buchanan, T. S., & Rymer, W. Z. (1995). Abnormal
muscle coactivation patterns during isometric torque generation at the elbow and shoulder
in hemiparetic subjects. 16.

Dodakian, L., McKenzie, A. L., Le, V., See, J., Pearson-Fuhrhop, K., Burke Quinlan, E., Zhou, R.
J., Augsberger, R., Tran, X. A., Friedman, N., Reinkensmeyer, D. J., & Cramer, S. C.
(2017). A Home-Based Telerehabilitation Program for Patients With Stroke.
Neurorehabilitation and Neural Repair, 31(10–11), 923–933.
https://doi.org/10.1177/1545968317733818

107
Doğan-Aslan, M., Nakipoğlu-Yüzer, G. F., Doğan, A., Karabay, İ., & Özgirgin, N. (2012). The
Effect of Electromyographic Biofeedback Treatment in Improving Upper Extremity
Functioning of Patients with Hemiplegic Stroke. Journal of Stroke and Cerebrovascular
Diseases, 21(3), 187–192. https://doi.org/10.1016/j.jstrokecerebrovasdis.2010.06.006

Donoso Brown, E. V., McCoy, S. W., Fechko, A. S., Price, R., Gilbertson, T., & Moritz, C. T.
(2014). Preliminary Investigation of an Electromyography-Controlled Video Game as a
Home Program for Persons in the Chronic Phase of Stroke Recovery. Archives of Physical
Medicine and Rehabilitation, 95(8), 1461–1469.
https://doi.org/10.1016/j.apmr.2014.02.025

Doya, K. (2000). Complementary roles of basal ganglia and cerebellum in learning and motor
control. Current Opinion in Neurobiology, 10(6), 732–739. https://doi.org/10.1016/S0959-
4388(00)00153-7

Duncan, P. W., Lai, S. M., Bode, R. K., Perera, S., & DeRosa, J. (2003). Stroke impact scale-16:
A brief assessment of physical function. Neurology, 8.

Ellis, M. D., Schut, I., & Dewald, J. P. A. (2017). Flexion synergy overshadows flexor spasticity
during reaching in chronic moderate to severe hemiparetic stroke. Clinical
Neurophysiology, 128(7), 1308–1314. https://doi.org/10.1016/j.clinph.2017.04.028

Fang, Y., Daly, J. J., Sun, J., Hvorat, K., Fredrickson, E., Pundik, S., Sahgal, V., & Yue, G. H.
(2009). Functional corticomuscular connection during reaching is weakened following
stroke. Clinical Neurophysiology, 120(5), 994–1002.
https://doi.org/10.1016/j.clinph.2009.02.173

Farina, D., Merletti, R., & Enoka, R. M. (2014). The extraction of neural strategies from the surface
EMG: An update. Journal of Applied Physiology, 117(11), 1215–1230.
https://doi.org/10.1152/japplphysiol.00162.2014

Farina, D., Negro, F., & Jiang, N. (2013). Identification of common synaptic inputs to motor
neurons from the rectified electromyogram: EMG rectification. The Journal of Physiology,
591(10), 2403–2418. https://doi.org/10.1113/jphysiol.2012.246082

Farmer, S. F. (1998). Rhythmicity, synchronization and binding in human and primate motor
systems. The Journal of Physiology, 509(1), 3–14. https://doi.org/10.1111/j.1469-
7793.1998.003bo.x

Feigin, V. L., Stark, B. A., Johnson, C. O., Roth, G. A., Bisignano, C., Abady, G. G., Abbasifard,
M., Abbasi-Kangevari, M., Abd-Allah, F., Abedi, V., Abualhasan, A., Abu-Rmeileh, N. M.,
Abushouk, A. I., Adebayo, O. M., Agarwal, G., Agasthi, P., Ahinkorah, B. O., Ahmad, S.,
Ahmadi, S., … Murray, C. J. L. (2021). Global, regional, and national burden of stroke and
its risk factors, 1990–2019: A systematic analysis for the Global Burden of Disease Study
2019. The Lancet Neurology, 20(10), 795–820. https://doi.org/10.1016/S1474-
4422(21)00252-0

Feldner, H. A., Howell, D., Kelly, V. E., McCoy, S. W., & Steele, K. M. (2019). “Look, Your Muscles
Are Firing!”: A Qualitative Study of Clinician Perspectives on the Use of Surface
Electromyography in Neurorehabilitation. Archives of Physical Medicine and
Rehabilitation, 100(4), 663–675. https://doi.org/10.1016/j.apmr.2018.09.120
108
Feldner, H. A., Papazian, C., Peters, K., & Steele, K. M. (2020). “It’s All Sort of Cool and
Interesting…but What Do I Do With It?” A Qualitative Study of Stroke Survivors’
Perceptions of Surface Electromyography. Frontiers in Neurology, 11, 1037.
https://doi.org/10.3389/fneur.2020.01037

Fisher, K. M., Zaaimi, B., Williams, T. L., Baker, S. N., & Baker, M. R. (2012). Beta-band
intermuscular coherence: A novel biomarker of upper motor neuron dysfunction in motor
neuron disease. Brain, 135(9), 2849–2864. https://doi.org/10.1093/brain/aws150

Fong, K. N. K., Ting, K. H., Zhang, J. J. Q., Yau, C. S. F., & Li, L. S. W. (2021). Event-Related
Desynchronization During Mirror Visual Feedback: A Comparison of Older Adults and
People After Stroke. Frontiers in Human Neuroscience, 15, 629592.
https://doi.org/10.3389/fnhum.2021.629592

Franceschini, M., Ceravolo, M. G., Agosti, M., Cavallini, P., Bonassi, S., Dall’Armi, V., Massucci,
M., Schifini, F., & Sale, P. (2012). Clinical Relevance of Action Observation in Upper-Limb
Stroke Rehabilitation: A Possible Role in Recovery of Functional Dexterity. A Randomized
Clinical Trial. Neurorehabilitation and Neural Repair, 26(5), 456–462.
https://doi.org/10.1177/1545968311427406

Fuentes del Toro, S., Wei, Y., Olmeda, E., Ren, L., Guowu, W., & Díaz, V. (2019). Validation of a
Low-Cost Electromyography (EMG) System via a Commercial and Accurate EMG Device:
Pilot Study. Sensors, 19(23). https://doi.org/10.3390/s19235214

Fugl-Meyer, A. R., Jaasko, L., Leyman, I., Olsson, S., & Steglind, S. (1975). The post-stroke
hemiplegic patient. 1. A method for evaluation of physical performance. Scand J Rehabil
Med, 7(1), 13–31.

Gadidi, V., Katz-Leurer, M., Carmeli, E., & Bornstein, N. M. (2011). Long-Term Outcome
Poststroke: Predictors of Activity Limitation and Participation Restriction. Archives of
Physical Medicine and Rehabilitation, 92(11), 1802–1808.
https://doi.org/10.1016/j.apmr.2011.06.014

Garrison, K. A., Aziz-Zadeh, L., Wong, S. W., Liew, S.-L., & Winstein, C. J. (2013). Modulating
the Motor System by Action Observation After Stroke. Stroke, 44(8), 2247–2253.
https://doi.org/10.1161/STROKEAHA.113.001105

Giggins, O. M., Persson, U. M., & Caulfield, B. (2013). Biofeedback in rehabilitation. Journal of
NeuroEngineering and Rehabilitation 10:60. https://doi:10.1186/1743-0003-10-60

Gittler, M., & Davis, A. M. (2018). Guidelines for Adult Stroke Rehabilitation and Recovery. JAMA
319(8):820–821. doi:10.1001/jama.2017.22036

Gladstone, D. J., Danells, C. J., & Black, S. E. (2002). The Fugl-Meyer Assessment of Motor
Recovery after Stroke: A Critical Review of Its Measurement Properties.
Neurorehabilitation and Neural Repair, 16(3), 232–240.
https://doi.org/10.1177/154596802401105171

109
Glanz, M., Klawansky, S., Stason, W., Berkey, C., Shah, N., Phan, H., & Chalmers, T. C. (1995).
Biofeedback therapy in poststroke rehabilitation: A meta-analysis of the randomized
controlled trials. Archives of Physical Medicine and Rehabilitation, 76(6), 508–515.
https://doi.org/10.1016/S0003-9993(95)80503-6

Guo, Z., Qian, Q., Wong, K., Zhu, H., Huang, Y., Hu, X., & Zheng, Y. (2020). Altered
Corticomuscular Coherence (CMCoh) Pattern in the Upper Limb During Finger
Movements After Stroke. Frontiers in Neurology, 11.
https://www.frontiersin.org/article/10.3389/fneur.2020.00410

Halliday, D. M., & Rosenberg, J. R. (2000). On the application, estimation and interpretation of
coherence and pooled coherence. Journal of Neuroscience Methods, 100(1–2), 173–174.
https://doi.org/10.1016/S0165-0270(00)00267-3

Hao, J., Xie, H., Harp, K., Chen, Z., & Siu, K.-C. (2022). Effects of Virtual Reality Intervention on
Neural Plasticity in Stroke Rehabilitation: A Systematic Review. Archives of Physical
Medicine and Rehabilitation, 103(3), 523–541. https://doi.org/10.1016/j.apmr.2021.06.024

Health Resources and Services Administration. (2022, June 3). Health equity in telehealth.
https://telehealth.hhs.gov/providers/health-equity-in-telehealth/

Hochstenbach-Waelen, A., & Seelen, H. A. (2012). Embracing change: Practical and theoretical
considerations for successful implementation of technology assisting upper limb training
in stroke. Journal of NeuroEngineering and Rehabilitation, 9(1), 52.
https://doi.org/10.1186/1743-0003-9-52

Holper, L., Muehlemann, T., Scholkmann, F., Eng, K., Kiper, D., & Wolf, M. (2010). Testing the
potential of a virtual reality neurorehabilitation system during performance of observation,
imagery and imitation of motor actions recorded by wireless functional near-infrared
spectroscopy (fNIRS). Journal of NeuroEngineering and Rehabilitation, 7(1), 57.
https://doi.org/10.1186/1743-0003-7-57

Hsieh, Y.-W., Lin, Y.-H., Zhu, J.-D., Wu, C.-Y., Lin, Y.-P., & Chen, C.-C. (2020). Treatment Effects
of Upper Limb Action Observation Therapy and Mirror Therapy on Rehabilitation
Outcomes after Subacute Stroke: A Pilot Study. Behavioural Neurology, 2020, 1–9.
https://doi.org/10.1155/2020/6250524

Huang, H., Wolf, S. L., & He, J. (2006). Recent developments in biofeedback for neuromotor
rehabilitation. Journal of NeuroEngineering and Rehabilitation, 3(1), 11.
https://doi.org/10.1186/1743-0003-3-11

Jian, C., Deng, L., Liu, H., Yan, T., Wang, X., & Song, R. (2021). Modulating and restoring inter-
muscular coordination in stroke patients using two-dimensional myoelectric computer
interface: A cross-sectional and longitudinal study. Journal of Neural Engineering, 18(3),
036005. https://doi.org/10.1088/1741-2552/abc29a

Juliano, J. M., Schweighofer, N., & Liew, S.-L. (2022). Increased cognitive load in immersive
virtual reality during visuomotor adaptation is associated with decreased long-term
retention and context transfer. Journal of NeuroEngineering and Rehabilitation, 19(1),
106. https://doi.org/10.1186/s12984-022-01084-6

110

Juliano, J. M., Spicer, R. P., Vourvopoulos, A., Lefebvre, S., Jann, K., Ard, T., Santarnecchi, E.,
Krum, D. M., & Liew, S.-L. (2020). Embodiment Is Related to Better Performance on a
Brain–Computer Interface in Immersive Virtual Reality: A Pilot Study. Sensors, 20(4),
1204. https://doi.org/10.3390/s20041204

Jurkiewicz, M. T., Marzolini, S., & Oh, P. (2011). Adherence to a Home-Based Exercise Program
for Individuals After Stroke. Topics in Stroke Rehabilitation, 18(3), 277–284.
https://doi.org/10.1310/tsr1803-277

Kattla, S., & Lowery, M. M. (2010). Fatigue related changes in electromyographic coherence
between synergistic hand muscles. Experimental Brain Research, 202(1), 89–99.
https://doi.org/10.1007/s00221-009-2110-0

Kawase, T., Sakurada, T., Koike, Y., & Kansaku, K. (2017). A hybrid BMI-based exoskeleton for
paresis: EMG control for assisting arm movements. Journal of Neural Engineering, 14(1),
016015. https://doi.org/10.1088/1741-2552/aa525f

Kellis, E., Arabatzi, F., & Papadopoulos, C. (2003). Muscle co-activation around the knee in drop
jumping using the co-contraction index. Journal of Electromyography and Kinesiology,
13(3), 229–238. https://doi.org/10.1016/S1050-6411(03)00020-8

Kennedy, R. S., Lane, N. E., Berbaum, K. S., & Lilienthal, M. G. (1993). Simulator Sickness
Questionnaire: An Enhanced Method for Quantifying Simulator Sickness. The
International Journal of Aviation Psychology, 3(3), 203–220.
https://doi.org/10.1207/s15327108ijap0303_3

Kilner, J. M., Baker, S. N., Salenius, S., Jousmäki, V., Hari, R., & Lemon, R. N. (1999). Task-
dependent modulation of 15-30 Hz coherence between rectified EMGs from human hand
and forearm muscles. The Journal of Physiology, 516(2), 559–570.
https://doi.org/10.1111/j.1469-7793.1999.0559v.x

Kim, J.-H. (2017). The effects of training using EMG biofeedback on stroke patients upper
extremity functions. Journal of Physical Therapy Science, 29(6), 1085–1088.
https://doi.org/10.1589/jpts.29.1085

Kothe, C. A. E., & Jung, T.-P. (2016). Artifact removal techniques with signal reconstruction.
Google Patents.

Krakauer, J. W. (2006). Motor learning: Its relevance to stroke recovery and neurorehabilitation:
Current Opinion in Neurology, 19(1), 84–90.
https://doi.org/10.1097/01.wco.0000200544.29915.cc

Krakauer, J. W., Hadjiosif, A. M., Xu, J., Wong, A. L., & Haith, A. M. (2019). Motor Learning. In
Comprehensive Physiology (pp. 613–663). American Cancer Society.
https://doi.org/10.1002/cphy.c170043

111
Krauth, R., Schwertner, J., Vogt, S., Lindquist, S., Sailer, M., Sickert, A., Lamprecht, J., Perdikis,
S., Corbet, T., Millán, J. del R., Hinrichs, H., Heinze, H.-J., & Sweeney-Reed, C. M. (2019).
Cortico-Muscular Coherence Is Reduced Acutely Post-stroke and Increases Bilaterally
During Motor Recovery: A Pilot Study. Frontiers in Neurology, 10, 126.
https://doi.org/10.3389/fneur.2019.00126

Kristeva, R., Patino, L., & Omlor, W. (2007). Beta-range cortical motor spectral power and
corticomuscular coherence as a mechanism for effective corticospinal interaction during
steady-state motor output. NeuroImage, 36(3), 785–792.
https://doi.org/10.1016/j.neuroimage.2007.03.025

Kwakkel, G., Kollen, B., & Lindeman, E. (2004). Understanding the pattern of functional recovery
after stroke: Facts and theories. Restorative Neurology and Neuroscience 22(3-5), 281-
299.

Kwakkel, G., Winters, C., van Wegen, E. E. H., Nijland, R. H. M., van Kuijk, A. A. A., Visser-Meily,
A., de Groot, J., de Vlugt, E., Arendzen, J. H., Geurts, A. C. H., Meskers, C. G. M. (2016).
Effects of Unilateral Upper Limb Training in Two Distinct Prognostic Groups Early After
Stroke: The EXPLICIT-Stroke Randomized Clinical Trial. Neurorehabilitation and Neural
Repair, 30(9), 804–816. https://doi.org/10.1177/1545968315624784

Laine, C. M., Nickerson, L. A., & Bailey, E. F. (2012). Cortical entrainment of human hypoglossal
motor unit activities. Journal of Neurophysiology, 107(1), 493–499.
https://doi.org/10.1152/jn.00769.2011

Laine, C. M., & Valero-Cuevas, F. J. (2017). Intermuscular coherence reflects functional
coordination. Journal of Neurophysiology, 118(3), 1775–1783.
https://doi.org/10.1152/jn.00204.2017

Laine, C. M., & Valero-Cuevas, F. J. (2020). Parkinson’s Disease Exhibits Amplified Intermuscular
Coherence During Dynamic Voluntary Action. Frontiers in Neurology, 11, 204.
https://doi.org/10.3389/fneur.2020.00204

Laine, C. M., Yavuz, Ş. U., & Farina, D. (2014). Task-related changes in sensorimotor integration
influence the common synaptic input to motor neurones. Acta Physiologica, 211(1), 229–
239. https://doi.org/10.1111/apha.12255

Lang, C. E., MacDonald, J. R., Reisman, D. S., Boyd, L., Jacobson Kimberley, T., Schindler-
Ivens, S. M., Hornby, T. G., Ross, S. A., & Scheets, P. L. (2009). Observation of Amounts
of Movement Practice Provided During Stroke Rehabilitation. Archives of Physical
Medicine and Rehabilitation, 90(10), 1692–1698.
https://doi.org/10.1016/j.apmr.2009.04.005

Langhorne, P., Bernhardt, J., & Kwakkel, G. (2011). Stroke rehabilitation. The Lancet, 377(9778),
1693–1702. https://doi.org/10.1016/S0140-6736(11)60325-5

Larsen, L. H., Zibrandtsen, I. C., Wienecke, T., Kjaer, T. W., Christensen, M. S., Nielsen, J. B., &
Langberg, H. (2017). Corticomuscular coherence in the acute and subacute phase after
stroke. Clinical Neurophysiology, 128(11), 2217–2226.
https://doi.org/10.1016/j.clinph.2017.08.033

112
Latash, M. L. (2018). Muscle coactivation: Definitions, mechanisms, and functions. Journal of
Neurophysiology, 120(1), 88–104. https://doi.org/10.1152/jn.00084.2018

Laver, K. E., Lange, B., George, S., Deutsch, J. E., Saposnik, G., & Crotty, M. (2017). Virtual
reality for stroke rehabilitation. Cochrane Database of Systematic Reviews.
https://doi.org/10.1002/14651858.CD008349.pub4

Leeb, R., Sagha, H., Chavarriaga, R., & Millán, J. del R. (2011). A hybrid brain–computer interface
based on the fusion of electroencephalographic and electromyographic activities. Journal
of Neural Engineering, 8(2), 025011. https://doi.org/10.1088/1741-2560/8/2/025011

Li, S., Chen, Y.-T., Francisco, G. E., Zhou, P., & Rymer, W. Z. (2019). A Unifying
Pathophysiological Account for Post-stroke Spasticity and Disordered Motor Control.
Frontiers in Neurology, 10, 468. https://doi.org/10.3389/fneur.2019.00468

Liew, S.-L., Marin-Pardo, O., & Phanord, C. (n.d.). Neurofeedback Rehabilitation System
(Patent No. WO2022183009A1).
https://patentscope.wipo.int/search/en/detail.jsf?docId=WO2022183009

Liew, S.-L., Rana, M., Cornelsen, S., Fortunato de Barros Filho, M., Birbaumer, N., Sitaram, R.,
Cohen, L. G., & Soekadar, S. R. (2016). Improving Motor Corticothalamic Communication
After Stroke Using Real-Time fMRI Connectivity-Based Neurofeedback.
Neurorehabilitation and Neural Repair, 30(7), 671–675.
https://doi.org/10.1177/1545968315619699

Lindenberg, R., Renga, V., Zhu, L. L., Betzler, F., Alsop, D., & Schlaug, G. (2010). Structural
integrity of corticospinal motor fibers predicts motor impairment in chronic stroke.
Neurology, 74(4), 280–287. https://doi.org/10.1212/WNL.0b013e3181ccc6d9

Liu, H., Gao, Y., Huang, W., Li, R., Houston, M., Benoit, J. S., Roh, J., Zhang, Y., (2022). Inter-
muscular coherence and functional coordination in the human upper extremity after stroke.
Mathematical Biosciences and Engineering, 19(5), 4506–4525.
https://doi.org/10.3934/mbe.2022208

Liu, J., Sheng, Y., & Liu, H. (2019). Corticomuscular Coherence and Its Applications: A Review.
Frontiers in Human Neuroscience, 13, 100. https://doi.org/10.3389/fnhum.2019.00100

Lobov, S., Krilova, N., Kastalskiy, I., Kazantsev, V., & Makarov, V. (2018). Latent Factors Limiting
the Performance of sEMG-Interfaces. Sensors, 18(4), 1122.
https://doi.org/10.3390/s18041122

Lohse, K. R., Lang, C. E., & Boyd, L. A. (2014). Is More Better? Using Metadata to Explore Dose–
Response Relationships in Stroke Rehabilitation. Stroke, 45(7), 2053–2058.
https://doi.org/10.1161/STROKEAHA.114.004695

Luo, J., Sun, W., Wu, Y., Liu, H., Wang, X., Yan, T., & Song, R. (2018). Characterization of the
coordination of agonist and antagonist muscles among stroke patients, healthy late
middle-aged and young controls using a myoelectric-controlled interface. Journal of
Neural Engineering, 15(5), 056015. https://doi.org/10.1088/1741-2552/aad387

113
Lyle, R. C. (1981). A performance test for assessment of upper limb function in physical
rehabilitation treatment and research. International Journal of Rehabilitation Research,
4(4), 483–492.

Maier, M., Ballester, B. R., & Verschure, P. F. M. J. (2019). Principles of Neurorehabilitation After
Stroke Based on Motor Learning and Brain Plasticity Mechanisms. Frontiers in Systems
Neuroscience, 13, 74. https://doi.org/10.3389/fnsys.2019.00074

Marin-Pardo, O., Donnelly, M., Phanord, C., Wong, K., Pan, J., & Liew, S.-L. (2022). Functional
and Neuromuscular Changes Induced via a Low-Cost, Muscle-Computer Interface for
Telerehabilitation: A Feasibility Study in Chronic Stroke. Frontiers in Neuroergonomics.
3:1046695. https://doi.org/10.3389/fnrgo.2022.1046695

Marin-Pardo, O., Laine, C. M., Rennie, M., Ito, K. L., Finley, J., & Liew, S.-L. (2020). A Virtual
Reality Muscle–Computer Interface for Neurorehabilitation in Chronic Stroke: A Pilot
Study. Sensors, 20(13), 3754. https://doi.org/10.3390/s20133754

Marin-Pardo, O., Phanord, C., Donnelly, M. R., Laine, C. M., & Liew, S.-L. (2021). Development
of a Low-Cost, Modular Muscle–Computer Interface for At-Home Telerehabilitation for
Chronic Stroke. Sensors, 21(5), 1806. https://doi.org/10.3390/s21051806

Marin-Pardo, O., Vourvopoulos, A., Neureither, M., Saldana, D., Jahng, E., & Liew, S.-L. (2019).
Electromyography as a Suitable Input for Virtual Reality-Based Biofeedback in Stroke
Rehabilitation. In C. Stephanidis (Ed.), HCI International 2019 —Posters (Vol. 1032, pp.
274–281). Springer International Publishing. https://doi.org/10.1007/978-3-030-23522-
2_35

Marzolini, S., Balitsky, A., Jagroop, D., Corbett, D., Brooks, D., Grace, S. L., Lawrence, D., & Oh,
P. I. (2016). Factors Affecting Attendance at an Adapted Cardiac Rehabilitation Exercise
Program for Individuals with Mobility Deficits Poststroke. Journal of Stroke and
Cerebrovascular Diseases, 25(1), 87–94.
https://doi.org/10.1016/j.jstrokecerebrovasdis.2015.08.039

Matarasso, A. K., Rieke, J. D., White, K., Yusufali, M. M., & Daly, J. J. (2021). Combined real-
time fMRI and real time fNIRS brain computer interface (BCI): Training of volitional wrist
extension after stroke, a case series pilot study. PLOS ONE, 16(5), e0250431.
https://doi.org/10.1371/journal.pone.0250431

Mathiowetz, V., Volland, G., Kashman, N., & Weber, K. (1985). Adult Norms for the Box and Block
Test of Manual Dexterity. The American Journal of Occupational Therapy, 39(6), 386–
391. https://doi.org/10.5014/ajot.39.6.386

Matsuya, R., Ushiyama, J., & Ushiba, J. (2017). Inhibitory interneuron circuits at cortical and
spinal levels are associated with individual differences in corticomuscular coherence
during isometric voluntary contraction. Scientific Reports, 7(1), Article 1.
https://doi.org/10.1038/srep44417

114
McCabe, J., Monkiewicz, M., Holcomb, J., Pundik, S., & Daly, J. J. (2015). Comparison of
Robotics, Functional Electrical Stimulation, and Motor Learning Methods for Treatment of
Persistent Upper Extremity Dysfunction After Stroke: A Randomized Controlled Trial.
Archives of Physical Medicine and Rehabilitation, 96(6), 981–990.
https://doi.org/10.1016/j.apmr.2014.10.022

McFarland, D. J., & Vaughan, T. M. (2016). BCI in practice. In Progress in Brain Research (Vol.
228, pp. 389–404). Elsevier. https://doi.org/10.1016/bs.pbr.2016.06.005

McPherson, J. G., Chen, A., Ellis, M. D., Yao, J., Heckman, C. J., & Dewald, J. P. A. (2018).
Progressive recruitment of contralesional cortico-reticulospinal pathways drives motor
impairment post stroke. The Journal of Physiology, 596(7), 1211–1225.
https://doi.org/10.1113/JP274968

Mehrkanoon, S., Breakspear, M., & Boonstra, T. W. (2014). The reorganization of corticomuscular
coherence during a transition between sensorimotor states. NeuroImage, 100, 692–702.
https://doi.org/10.1016/j.neuroimage.2014.06.050

Mima, T., & Hallett, M. (1999). Corticomuscular Coherence: A Review. Journal of Clinical
Neurophysiology, 16(6), 501.

Mima, T., Toma, K., Koshy, B., & Hallett, M. (2001). Coherence Between Cortical and Muscular
Activities After Subcortical Stroke. Stroke, 32(11), 2597–2601.
https://doi.org/10.1161/hs1101.098764

Mugler, E. M., Tomic, G., Singh, A., Hameed, S., Lindberg, E. W., Gaide, J., Alqadi, M., Robinson,
E., Dalzotto, K., Limoli, C., Jacobson, T., Lee, J., & Slutzky, M. W. (2019). Myoelectric
Computer Interface Training for Reducing Co-Activation and Enhancing Arm Movement
in Chronic Stroke Survivors: A Randomized Trial. Neurorehabilitation and Neural Repair,
33(4), 284–295. https://doi.org/10.1177/1545968319834903

Mulas, M., Folgheraiter, M., & Gini, G. (2005). An EMG-Controlled Exoskeleton for Hand
Rehabilitation. 9th International Conference on Rehabilitation Robotics, 2005. ICORR
2005., 371–374. https://doi.org/10.1109/ICORR.2005.1501122

Nasreddine, Z. S., Phillips, N. A., Bédirian, V., Charbonneau, S., Whitehead, V., Collin, I.,
Cummings, J. L., & Chertkow, H. (2005). The Montreal Cognitive Assessment, MoCA: A
Brief Screening Tool For Mild Cognitive Impairment. Journal of the American Geriatrics
Society, 53(4), 695–699. https://doi.org/10.1111/j.1532-5415.2005.53221.x

Ngeo, J., Tamei, T., Shibata, T., Orlando, M. F. F., Behera, L., Saxena, A., & Dutta, A. (2013).
Control of an optimal finger exoskeleton based on continuous joint angle estimation from
EMG signals. 2013 35th Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), 338–341.
https://doi.org/10.1109/EMBC.2013.6609506

Norton, J. A., & Gorassini, M. A. (2006). Changes in Cortically Related Intermuscular Coherence
Accompanying Improvements in Locomotor Skills in Incomplete Spinal Cord Injury.
Journal of Neurophysiology, 95(4), 2580–2589. https://doi.org/10.1152/jn.01289.2005

115
Owen, M., Ingo, C., & Dewald, J. P. A. (2017). Upper Extremity Motor Impairments and
Microstructural Changes in Bulbospinal Pathways in Chronic Hemiparetic Stroke.
Frontiers in Neurology, 8. https://www.frontiersin.org/article/10.3389/fneur.2017.00257

Ozen, S., Senlikci, H. B., Guzel, S., & Yemisci, O. U. (2021). Computer Game Assisted Task
Specific Exercises in the Treatment of Motor and Cognitive Function and Quality of Life in
Stroke: A Randomized Control Study. Journal of Stroke and Cerebrovascular Diseases,
30(9), 105991. https://doi.org/10.1016/j.jstrokecerebrovasdis.2021.105991

Page, S. J., Fulk, G. D., & Boyne, P. (2012). Clinically Important Differences for the Upper-
Extremity Fugl-Meyer Scale in People With Minimal to Moderate Impairment Due to
Chronic Stroke. Physical Therapy, 92(6), 791–798. https://doi.org/10.2522/ptj.20110009

Palmer, E., & Ashby, P. (1992). Corticospinal projections to upper limb motoneurones in humans.
The Journal of Physiology, 448(1), 397–412.
https://doi.org/10.1113/jphysiol.1992.sp019048

Palsbo, S. E., Dawson, S. J., Savard, L., Goldstein, M., & Heuser, A. (2007). Televideo
assessment using Functional Reach Test and European Stroke Scale. The Journal of
Rehabilitation Research and Development, 44(5), 659.
https://doi.org/10.1682/JRRD.2006.11.0144

Pan, L.-L. H., Yang, W.-W., Kao, C.-L., Tsai, M.-W., Wei, S.-H., Fregni, F., Chen, V. C.-F., &
Chou, L.-W. (2018). Effects of 8-week sensory electrical stimulation combined with motor
training on EEG-EMG coherence and motor function in individuals with stroke. Scientific
Reports, 8(1), 9217. https://doi.org/10.1038/s41598-018-27553-4

Perez-Marcos, D. (2018). Virtual reality experiences, embodiment, videogames and their
dimensions in neurorehabilitation. Journal of NeuroEngineering and Rehabilitation, 15(1),
113. https://doi.org/10.1186/s12984-018-0461-0

Peters, B., Bieker, G., Heckman, S. M., Huggins, J. E., Wolf, C., Zeitlin, D., & Fried-Oken, M.
(2015). Brain-Computer Interface Users Speak Up: The Virtual Users’ Forum at the 2013
International Brain-Computer Interface Meeting. Archives of Physical Medicine and
Rehabilitation, 96(3), S33–S37. https://doi.org/10.1016/j.apmr.2014.03.037

Pfurtscheller, G., Müller, G. R., Pfurtscheller, J., Gerner, H. J., & Rupp, R. (2003). ‘Thought’ –
control of functional electrical stimulation to restore hand grasp in a patient with
tetraplegia. Neuroscience Letters, 351(1), 33–36. https://doi.org/10.1016/S0304-
3940(03)00947-9

Platz, T., Pinkowski, C., van Wijck, F., Kim, I.-H., di Bella, P., & Johnson, G. (2005). Reliability
and validity of arm function assessment with standardized guidelines for the Fugl-Meyer
Test, Action Research Arm Test and Box and Block Test: A multicentre study. Clinical
Rehabilitation, 19(4), 404–411. https://doi.org/10.1191/0269215505cr832oa

Potvin, J. R., & Brown, S. H. M. (2004). Less is more: High pass filtering, to remove up to 99% of
the surface EMG signal power, improves EMG-based biceps brachii muscle force
estimates. Journal of Electromyography and Kinesiology, 14(3), 389–399.
https://doi.org/10.1016/j.jelekin.2003.10.005

116
Power, H. A., Norton, J. A., Porter, C. L., Doyle, Z., Hui, I., & Chan, K. M. (2006). Transcranial
direct current stimulation of the primary motor cortex affects cortical drive to human
musculature as assessed by intermuscular coherence. The Journal of Physiology, 577(3),
795–803. https://doi.org/10.1113/jphysiol.2006.116939

Ramos-Murguialday, A., Broetz, D., Rea, M., Läer, L., Yilmaz, Ö., Brasil, F. L., Liberati, G.,
Curado, M. R., Garcia-Cossio, E., Vyziotis, A., Cho, W., Agostini, M., Soares, E.,
Soekadar, S., Caria, A., Cohen, L. G., & Birbaumer, N. (2013). Brain-machine interface in
chronic stroke rehabilitation: A controlled study: BMI in Chronic Stroke. Annals of
Neurology, 74(1), 100–108. https://doi.org/10.1002/ana.23879

Rathore, S. S., Hinn, A. R., Cooper, L. S., Tyroler, H. A., & Rosamond, W. D. (2002).
Characterization of Incident Stroke Signs and Symptoms: Findings From the
Atherosclerosis Risk in Communities Study. Stroke, 33(11), 2718–2721.
https://doi.org/10.1161/01.STR.0000035286.87503.31

Remsik, A., Young, B., Vermilyea, R., Kiekhoefer, L., Abrams, J., Evander Elmore, S., Schultz,
P., Nair, V., Edwards, D., Williams, J., & Prabhakaran, V. (2016). A review of the
progression and future implications of brain-computer interface therapies for restoration
of distal upper extremity motor function after stroke. Expert Review of Medical Devices,
13(5), 445–454. https://doi.org/10.1080/17434440.2016.1174572

Riley, Z. A., Terry, M. E., Mendez-Villanueva, A., Litsey, J. C., & Enoka, R. M. (2008). Motor unit
recruitment and bursts of activity in the surface electromyogram during a sustained
contraction. Muscle & Nerve, 37(6), 745–753. https://doi.org/10.1002/mus.20978

Rosenberg, J. R., Amjad, A. M., Breeze, P., Brillinger, D. R., & Halliday, D. M. (1989). The Fourier
approach to the identification of functional coupling between neuronal spike trains.
Progress in Biophysics and Molecular Biology, 53(1), 1–31. https://doi.org/10.1016/0079-
6107(89)90004-7

Rossiter, H. E., Eaves, C., Davis, E., Boudrias, M.-H., Park, C., Farmer, S., Barnes, G., Litvak,
V., & Ward, N. S. (2013). Changes in the location of cortico-muscular coherence following
stroke. NeuroImage: Clinical, 2, 50–55. https://doi.org/10.1016/j.nicl.2012.11.002

Ryu, J., Cooney, W. P., Askew, L. J., An, K.-N., & Chao, E. Y. S. (1991). Functional ranges of
motion of the wrist joint. The Journal of Hand Surgery, 16(3), 409–419.
https://doi.org/10.1016/0363-5023(91)90006-W

Saes, M., Meskers, C. G. M., Daffertshofer, A., de Munck, J. C., Kwakkel, G., & van Wegen, E.
E. H. (2019). How does upper extremity Fugl-Meyer motor score relate to resting-state
EEG in chronic stroke? A power spectral density analysis. Clinical Neurophysiology,
130(5), 856–862. https://doi.org/10.1016/j.clinph.2019.01.007

Schleenbaker, R. E., & Mainous, A. G. (1993). Electromyographic biofeedback for neuromuscular
reeducation in the hemiplegic stroke patient: A meta-analysis. Archives of Physical
Medicine and Rehabilitation, 74(12), 1301–1304. https://doi.org/10.1016/0003-
9993(93)90083-M

117
Sebastián-Romagosa, M., Udina, E., Ortner, R., Dinarès-Ferran, J., Cho, W., Murovec, N.,
Matencio-Peralba, C., Sieghartsleitner, S., Allison, B. Z., & Guger, C. (2020). EEG
Biomarkers Related With the Functional State of Stroke Patients. Frontiers in
Neuroscience, 14, 582. https://doi.org/10.3389/fnins.2020.00582

Soekadar, S. R., Birbaumer, N., Slutzky, M. W., & Cohen, L. G. (2015). Brain–machine interfaces
in neurorehabilitation of stroke. Neurobiology of Disease, 83, 172–179.
https://doi.org/10.1016/j.nbd.2014.11.025

Spampinato, D., & Celnik, P. (2020). Multiple Motor Learning Processes in Humans: Defining
Their Neurophysiological Bases. The Neuroscientist, 107385842093955.
https://doi.org/10.1177/1073858420939552

Spicer, R., Anglin, J., Krum, D. M., & Liew, S. (2017). REINVENT: A low-cost, virtual reality brain-
computer interface for severe stroke upper limb motor recovery. 2017 IEEE Virtual Reality
(VR), 385–386. https://doi.org/10.1109/VR.2017.7892338

Staudenmann, D., Potvin, J. R., Kingma, I., Stegeman, D. F., & van Dieën, J. H. (2007). Effects
of EMG processing on biomechanical models of muscle joint systems: Sensitivity of trunk
muscle moments, spinal forces, and stability. Journal of Biomechanics, 40(4), 900–909.
https://doi.org/10.1016/j.jbiomech.2006.03.021

Staudenmann, D., Roeleveld, K., Stegeman, D. F., & van Dieën, J. H. (2010). Methodological
aspects of SEMG recordings for force estimation – A tutorial and review. Journal of
Electromyography and Kinesiology, 20(3), 375–387.
https://doi.org/10.1016/j.jelekin.2009.08.005

Stinear, C. M. (2017). Prediction of motor recovery after stroke: Advances in biomarkers. The
Lancet Neurology, 16(10), 826–836. https://doi.org/10.1016/S1474-4422(17)30283-1

Stinear, C. M., Lang, C. E., Zeiler, S., & Byblow, W. D. (2020). Advances and challenges in stroke
rehabilitation. The Lancet Neurology, 19(4), 348–360. https://doi.org/10.1016/S1474-
4422(19)30415-6

Sugg, K., Müller, S., Winstein, C., Hathorn, D., & Dempsey, A. (2015). Does Action Observation
Training With Immediate Physical Practice Improve Hemiparetic Upper-Limb Function in
Chronic Stroke? Neurorehabilitation and Neural Repair, 29(9), 807–817.
https://doi.org/10.1177/1545968314565512

Suresh, N. L., Concepcion, N. S., Madoff, J., & Rymer, W. Z. (2015). Anomalous EMG–force
relations during low-force isometric tasks in hemiparetic stroke survivors. Experimental
Brain Research, 233(1), 15–25. https://doi.org/10.1007/s00221-014-4061-3

Swartz Center for Computational Neuroscience. (2022, September 23). Labstreaming Layer
[Repository]. https://github.com/sccn/labstreaminglayer

Swieten, J. C. van, Koudstaal, P. J., Visser, M. C., Schouten, H. J., & Gijn, J. van. (1988).
Interobserver agreement for the assessment of handicap in stroke patients. Stroke, 19(5),
604–607. https://doi.org/10.1161/01.STR.19.5.604

118
Thieme, H., Morkisch, N., Mehrholz, J., Pohl, M., Behrens, J., Borgetto, B., & Dohle, C. (2018).
Mirror therapy for improving motor function after stroke. Cochrane Database of Systematic
Reviews. https://doi.org/10.1002/14651858.CD008449.pub3

Toro, S. F. del, Santos-Cuadros, S., Olmeda, E., Álvarez-Caldas, C., Díaz, V., & San Román, J.
L. (2019). Is the Use of a Low-Cost sEMG Sensor Valid to Measure Muscle Fatigue?
Sensors, 19(14), 3204. https://doi.org/10.3390/s19143204

Tsao, C. W., Aday, A. W., Almarzooq, Z. I., Alonso, A., Beaton, A. Z., Bittencourt, M. S., Boehme,
A. K., Buxton, A. E., Carson, A. P., Commodore-Mensah, Y., Elkind, M. S. V., Evenson,
K. R., Eze-Nliam, C., Ferguson, J. F., Generoso, G., Ho, J. E., Kalani, R., Khan, S. S.,
Kissela, B. M. (2022). Heart Disease and Stroke Statistics—2022 Update: A Report From
the American Heart Association. Circulation, 145(8).
https://doi.org/10.1161/CIR.0000000000001052

van Andel, C. J., Wolterbeek, N., Doorenbosch, C. A. M., Veeger, D. (H. E. J. ), & Harlaar, J.
(2008). Complete 3D kinematics of upper extremity functional tasks. Gait & Posture, 27(1),
120–127. https://doi.org/10.1016/j.gaitpost.2007.03.002

van der Lee, J. H., Beckerman, H., Lankhorst, G. J., & Bouter, L. M. (2001). The Responsiveness
of the Action Research Arm Test and the Fugl-Meyer Assessment Scale in Chronic Stroke
Patients. Journal of Rehabilitation Medicine, 33(3), 110–113.
https://doi.org/10.1080/165019701750165916

van der Lee, J. H., de Groot, V., Beckerman, H., Wagenaar, R. C., Lankhorst, G. J., & Bouter, L.
M. (2001). The intra- and interrater reliability of the action research arm test: A practical
test of upper extremity function in patients with stroke. Archives of Physical Medicine and
Rehabilitation, 82(1), 14–19. https://doi.org/10.1053/apmr.2001.18668

Veerbeek, J. M., van Wegen, E., van Peppen, R., van der Wees, P. J., Hendriks, E., Rietberg,
M., & Kwakkel, G. (2014). What Is the Evidence for Physical Therapy Poststroke? A
Systematic Review and Meta-Analysis. PLoS ONE, 9(2), e87987.
https://doi.org/10.1371/journal.pone.0087987

Virani, S. S., Alonso, A., Benjamin, E. J., Bittencourt, M. S., Callaway, C. W., Carson, A. P.,
Chamberlain, A. M., Chang, A. R., Cheng, S., Delling, F. N., Djousse, L., Elkind, M. S. V.,
Ferguson, J. F., Fornage, M., Khan, S. S., Kissela, B. M., Knutson, K. L., Kwan, T. W.,
Lackland, D. T., … On behalf of the American Heart Association Council on Epidemiology
and Prevention Statistics Committee and Stroke Statistics Subcommittee. (2020). Heart
Disease and Stroke Statistics—2020 Update: A Report From the American Heart
Association. Circulation, 141(9). https://doi.org/10.1161/CIR.0000000000000757

von Carlowitz-Ghori, K., Bayraktaroglu, Z., Hohlefeld, F. U., Losch, F., Curio, G., & Nikulin, V. V.
(2014). Corticomuscular coherence in acute and chronic stroke. Clinical Neurophysiology,
125(6), 1182–1191. https://doi.org/10.1016/j.clinph.2013.11.006

Vourvopoulos, A., & Bermúdez i Badia, S. (2016). Motor priming in virtual reality can augment
motor-imagery training efficacy in restorative brain-computer interaction: A within-subject
analysis. Journal of NeuroEngineering and Rehabilitation, 13(1), 69.
https://doi.org/10.1186/s12984-016-0173-2

119
Vourvopoulos, A., Marin-Pardo, O., Neureither, M., Saldana, D., Jahng, E., & Liew, S.-L. (2019).
Multimodal Head-Mounted Virtual-Reality Brain-Computer Interface for Stroke
Rehabilitation: A Clinical Case Study with REINVENT. In J. Y. C. Chen & G. Fragomeni
(Eds.), Virtual, Augmented and Mixed Reality. Multimodal Interaction (Vol. 11574, pp.
165–179). Springer International Publishing. https://doi.org/10.1007/978-3-030-21607-
8_13

Vourvopoulos, A., Pardo, O. M., Lefebvre, S., Neureither, M., Saldana, D., Jahng, E., & Liew, S.-
L. (2019). Effects of a Brain-Computer Interface With Virtual Reality (VR) Neurofeedback:
A Pilot Study in Chronic Stroke Patients. Frontiers in Human Neuroscience, 13, 210.
https://doi.org/10.3389/fnhum.2019.00210

Ward, N. S., Brander, F., & Kelly, K. (2019). Intensive upper limb neurorehabilitation in chronic
stroke: Outcomes from the Queen Square programme. Journal of Neurology,
Neurosurgery & Psychiatry, 90(5), 498–506. https://doi.org/10.1136/jnnp-2018-319954

Weiskopf, N., Veit, R., Erb, M., Mathiak, K., Grodd, W., Goebel, R., & Birbaumer, N. (2003).
Physiological self-regulation of regional brain activity using real-time functional magnetic
resonance imaging (fMRI): Methodology and exemplary data. NeuroImage, 19(3), 577–
586. https://doi.org/10.1016/S1053-8119(03)00145-9

Winstein, C. J., Miller, J. P., Blanton, S., Taub, E., Uswatte, G., Morris, D., Nichols, D., & Wolf, S.
(2003). Methods for a Multisite Randomized Trial to Investigate the Effect of Constraint-
Induced Movement Therapy in Improving Upper Extremity Function among Adults
Recovering from a Cerebrovascular Stroke. Neurorehabilitation and Neural Repair, 17(3),
137–152. https://doi.org/10.1177/0888439003255511

Winstein, C. J., Stein, J., Arena, R., Bates, B., Cherney, L. R., Cramer, S. C., Deruyter, F., Eng,
J. J., Fisher, B., Harvey, R. L., Lang, C. E., MacKay-Lyons, M., Ottenbacher, K. J., Pugh,
S., Reeves, M. J., Richards, L. G., Stiers, W., & Zorowitz, R. D. (2016). Guidelines for
Adult Stroke Rehabilitation and Recovery: A Guideline for Healthcare Professionals From
the American Heart Association/American Stroke Association. Stroke, 47(6).
https://doi.org/10.1161/STR.0000000000000098

Winstein, C. J., Wolf, S. L., & Schweighofer, N. (2008). Task-oriented training to promote upper
extremity recovery. Stroke Recovery and Rehabilitation, 2, 320–343.

Wolf, S. L. (1983). Electromyographic Biofeedback Applications to Stroke Patients. Physical
Therapy, 63(9), 1448–1459. https://doi.org/10.1093/ptj/63.9.1448

Wolf, S. L., Thompson, P. A., Winstein, C. J., Miller, J. P., Blanton, S. R., Nichols-Larsen, D. S.,
Morris, D. M., Uswatte, G., Taub, E., Light, K. E., & Sawaki, L. (2010). The EXCITE Stroke
Trial: Comparing Early and Delayed Constraint-Induced Movement Therapy. Stroke,
41(10), 2309–2315. https://doi.org/10.1161/STROKEAHA.110.588723

Woodford, H. J., & Price, C. I. (2007). EMG biofeedback for the recovery of motor function after
stroke. Cochrane Database of Systematic Reviews.
https://doi.org/10.1002/14651858.CD004585.pub2

120
Wright, Z. A., Rymer, W. Z., & Slutzky, M. W. (2014). Reducing Abnormal Muscle Coactivation
After Stroke Using a Myoelectric-Computer Interface: A Pilot Study. Neurorehabilitation
and Neural Repair, 28(5), 443–451. https://doi.org/10.1177/1545968313517751

Yeh, T., Chang, K., & Wu, C. (2019). The Active Ingredient of Cognitive Restoration: A Multicenter
Randomized Controlled Trial of Sequential Combination of Aerobic Exercise and
Computer-Based Cognitive Training in Stroke Survivors With Cognitive Decline. Archives
of Physical Medicine and Rehabilitation, 100(5), 821–827.
https://doi.org/10.1016/j.apmr.2018.12.020

Zackowski, K. M. (2004). How do strength, sensation, spasticity and joint individuation relate to
the reaching deficits of people with chronic hemiparesis? Brain, 127(5), 1035–1046.
https://doi.org/10.1093/brain/awh116

Zhang, J. J. Q., Fong, K. N. K., Welage, N., & Liu, K. P. Y. (2018). The Activation of the Mirror
Neuron System during Action Observation and Action Execution with Mirror Visual
Feedback in Stroke: A Systematic Review. Neural Plasticity, 2018, 1–14.
https://doi.org/10.1155/2018/2321045

Zheng, Y., Peng, Y., Xu, G., Li, L., & Wang, J. (2018). Using Corticomuscular Coherence to
Reflect Function Recovery of Paretic Upper Limb after Stroke: A Case Study. Frontiers in
Neurology, 8, 728. https://doi.org/10.3389/fneur.2017.00728

121
Appendix A: List of Relevant Publications

Peer-reviewed Articles:
Marin-Pardo, O., Donnelly, M., Phanord, C., Wong, K., Pan, J., & Liew, S.-L. (2022). Functional
and Neuromuscular Changes Induced via a Low-Cost, Muscle-Computer Interface for
Telerehabilitation: A Feasibility Study in Chronic Stroke. Frontiers in Neuroergonomics.
3:1046695. https://doi.org/10.3389/fnrgo.2022.1046695

Marin-Pardo, O., Phanord, C., Donnelly, M. R., Laine, C. M., & Liew, S.-L. (2021). Development
of a Low-Cost, Modular Muscle–Computer Interface for At-Home Telerehabilitation for Chronic
Stroke. Sensors, 21(5), 1806. https://doi.org/10.3390/s21051806

Marin-Pardo, O., Laine, C. M., Rennie, M., Ito, K. L., Finley, J., & Liew, S.-L. (2020). A Virtual
Reality Muscle–Computer Interface for Neurorehabilitation in Chronic Stroke: A Pilot Study.
Sensors, 20(13), 3754. https://doi.org/10.3390/s20133754

Vourvopoulos, A., Pardo, O. M., Lefebvre, S., Neureither, M., Saldana, D., Jahng, E., & Liew, S.-
L. (2019). Effects of a Brain-Computer Interface with Virtual Reality (VR) Neurofeedback: A Pilot
Study in Chronic Stroke Patients. Frontiers in Human Neuroscience, 13, 210.
https://doi.org/10.3389/fnhum.2019.00210

Patents Under Review:
Liew, S.-L., Marin-Pardo, O., & Phanord, C. (2022) Neurofeedback Rehabilitation System (Patent
Application WO2022183009A1).

Abstract (if available)

Abstract Stroke is a leading cause of long-term adult disability. People with severe motor impairment (e.g., little to no residual hand or wrist movement) in the chronic phase (more than six months after stroke onset) have limited rehabilitation options. High doses of repeated movement can improve upper limb function. However, it remains a challenge to provide such doses in standard clinical practice. Therefore, there is an urgent need for alternative therapeutic interventions for chronic stroke survivors with severe motor impairment. In addition, engaging and goal-directed exercise can be implemented via gamified training programs. Specifically, providing feedback of muscle activity to avoid unintended coactivation of antagonistic muscles has been shown to improve motor function. Therefore, in this dissertation I present and discuss the development and test of REINVENT, a low-cost, portable, modular, mixed-reality, biofeedback system for in-person and remote neurorehabilitation. The main goal is to evaluate the feasibility of using REINVENT with different feedback modalities (e.g., immersive and screen-based) in various training environments (e.g., in-person and at-home). Accordingly, we asked ten severely impaired chronic stroke survivors to use REINVENT to practice individuated use of wrist muscles during three separate case studies. Our combined results suggest that this approach is feasible, safe, and capable of inducing positive outcomes in stroke rehabilitation across varied functional domains. Furthermore, these improvements were presented along indications of neuroplastic changes and improved muscle control. Together, this work supports the expansion of using inexpensive and customizable technology to aid clinicians in improving people’s quality of life after stroke.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

The effects of fast walking, biofeedback, and cognitive impairment on post-stroke gait

PDF

Virtual surgeries as a tool for studying motor learning

PDF

Relationship between brain structure and motor behavior in chronic stroke survivors

PDF

A multi-site neuroimaging approach to studying hippocampal damage in chronic stroke

PDF

Feasibility, acceptability, and implementation context of a complex telerehabilitation intervention for post-stroke upper extremity recovery

PDF

Design of adaptive automated robotic task presentation system for stroke rehabilitation

PDF

Mechanisms of motor learning in immersive virtual reality and their influences on retention and context transfer

PDF

Computational model of stroke therapy and long term recovery

PDF

Deficits and rehabilitation of upper-extremity multi-joint movements in individuals with chronic stroke

PDF

Hemisphere-specific deficits in the control of bimanual movements after stroke

PDF

Understanding reactive balance control strategies in non-disabled and post-stroke gait

PDF

Experimental and computational explorations of different forms of plasticity in motor learning and stroke recovery

PDF

Locomotor skill learning in virtual reality in healthy adults and people with Parkinson disease

PDF

Arm choice post-stroke

PDF

Engineering 2D & 3D microphysiological systems for interrogating skeletal muscle tissues

PDF

Engineering scalable two- and three-dimensional striated muscle microtissues for human disease modeling

PDF

Interaction between Artificial Intelligence Systems and Primate Brains

PDF

Motor and visual principles in a human-computer interface training system for rehabilitation

PDF

Model-based approaches to objective inference during steady-state and adaptive locomotor control

PDF

New approaches for precisely engineering heterotypic muscle tissues by naturally and synthetically controlling cell fate

Asset Metadata

Creator Marin-Pardo, Octavio (author)

Core Title Development and implementation of a modular muscle-computer interface for personalized motor rehabilitation after stroke

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Degree Conferral Date 2023-05

Publication Date 05/04/2023

Defense Date 02/08/2023

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag biofeedback,electromyography,human-computer interface,OAI-PMH Harvest,Rehabilitation,stroke

Format theses (aat)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Liew, Sook-Lei (committee chair), Finley, James M. (committee member), Leech, Kristan (committee member), McCain, Megan (committee member), Schweighofer, Nicolas (committee member)

Creator Email marinpar@usc.edu,octaviomp92@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC113099724

Unique identifier UC113099724

Identifier etd-MarinPardo-11774.pdf (filename)

Legacy Identifier etd-MarinPardo-11774

Document Type Dissertation

Format theses (aat)

Rights Marin-Pardo, Octavio

Internet Media Type application/pdf

Type texts

Source 20230505-usctheses-batch-1037 (batch), University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu