University of Southern California Dissertations and Theses

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

(USC Thesis Other)

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

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

LOCOMOTOR SKILL LEARNING IN VIRTUAL REALITY IN HEALTHY ADULTS AND
PEOPLE WITH PARKINSON DISEASE

by

Aram Kim

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
(BIOKINESIOLOGY)

August 2021

Copyright 2021 Aram Kim
ii
DEDICATION
To my parents, Kyung-Han and Hyang, for their unconditional support and love.
iii
ACKNOWLEDGEMENTS
I wish to express my deepest gratitude to my MS and Ph.D. advisor, Dr. James Finley. He
gave me opportunities for many different research projects and invaluable guidance to grow as a
scientist. He has taught me how to think critically like a good scientist, write clearly, and better
present my research. I instantly fell in love with research because I observed his initiative,
creativity, and problem-solving towards research. Without his endless support for learning new
skillsets and exploring new areas, none of this was possible. Finally, his enthusiasm, vision,
genuineness, and mindfulness in diversity and equity inspired me and shaped me into what kind
of scientist and leader I want to be.
Next, I would like to thank my committee members.
I want to thank Dr. Fisher for our remarkable meetings, reading papers together, feedback
on my writing to think about my research more logically and clearly by asking critical questions.
I learned many ways to look at Parkinson disease and motor learning thanks to her expertise. I
also appreciate her writing a letter of recommendation for the Link Foundation fellowship. Her
generosity goes not only professionally by always sharing her time, whether at school or a
conference, but also personally by inviting me into her amazing Thanksgiving dinners.
Thank you, Dr. Kaplan, for your countless advice and feedback about fMRI processing
and analysis. Neuroimaging using MRI and fMRI was a completely new field for me, but he has
taught me everything I know about how to collect, process, and analyze data. I appreciate him
always available when I ran into problems and together trying to figure out these issues. I learned
to do rigorous imaging analysis thanks to him.
iv
I want to thank Dr. Liew for our discussions on understanding functional connectivity of
the brain, study designs for motor learning, and quality checks on imaging results. She gave me a
skillset to utilize open-source, big data, which was extremely helpful as a beginner who just
started learning neuroimaging. Moreover, writing is always hard, but it is especially harder when
I write for a new field. Writing a manuscript for neuroimaging was one of them, but she
generously provided me suggestions and feedback. I also appreciate her writing an amazing letter
of recommendation for the Link Foundation fellowship. Beyond academically, our joint-lab
outdoor days, parties, and happy hours are the memories that I will forever cherish.
Thank you, Dr. Schweighofer, for the encouragement and inspiration to reach the highest.
He has given me a skillset of statistical learning, computational modeling, and the most
importantly, Bayesian statistics. His non-stop search for better analysis tools is inspiring, and
without his initiative, my dissertation would not have been possible. He really cared about my
future career and equipped me to become a better scientist. His comments and feedback in any
setting, whether for writing, presentations, or during a journal club, were so stimulating that I
spent many hours reading about additional materials.
Next, I want to thank the Locomotor Control Lab members. Dr. Natalia Sanchez joined
the lab when I started the second year of my Master’s degree. She provided me invaluable
insights into my research and one of the most reliable people who reviewed various writing from
my personal statement to research manuscripts. Chang Liu joined the lab with me as a new Ph.D.
student, and we quickly became a ‘coffee buddy’ every morning. She was one of the most
dependable people to have my back whenever I have a problem running experiments,
mathematical equations, and my writing. Both Natalia and Chang became one of my closest
friends in the course of my time at USC. We shared our time not only professionally but also
v
personally exploring California. Our friendship was one of the reasons that this long journey was
a bit easier. Dr. Sungwoo Park was a former member of LCL and joined the lab with me from
the start. Sung was the person I would ask for all the processes that I needed to go through in the
division and during my Ph.D. He was always very willing to share his writings and codes, which
helped me in numerous ways. I also specifically thank Dr. Russell Johnson. Russell was the to-
go person whenever I had a question about Bayesian statistics. He was always available for my
questions and happy to share his time to improve myself. He was very helpful in going through
the process of finding a post-doctoral position. Shreya Jain and Isaiah Lachica helped me to
collect data, especially during this uncertain time. Pouria Nozari, Dr. Kristan Leech, and Sarah
Kettlety from the Gait Rehabilitation and Motor Learning lab provided me invaluable feedback
on my presentations to think about my research more thoroughly.
I would also like to thank researchers and students at USC who inspired me and assisted
me and my dissertation. Dr. Giselle Petzinger generously offered me to be part of her research
project, which helped part of my dissertation. Her insightful feedback during our discussions was
extremely helpful, and her enthusiasm and energy have been very inspiring. Vy Bui has been an
integral part of the DoD project. She coordinated all different labs and collected data. Whenever
I was not available, she was a reliable person who I can ask to substitute me. Julia Juliano, Kaori
Ito, Octavio Marin Pardo, and Dr. Christopher Laine never hesitated to help me whenever I have
questions regarding neuroimaging, post-doc positions, writing, and general concerns regarding
the Ph.D. process.
I want to thank the division and funding organizations. The Division of Biokinesiology
and Physical Therapy provided excellent opportunities to perform high-quality research and held
a platform to present my work to an interdisciplinary crowd. They also awarded me a student
vi
travel grant to attend the OHBM that immensely helped me improve my neuroimaging analysis.
Moreover, the Link Foundation, the National Institute of Health, and the Department of Defense
funded either my stipend or research. Without their financial support, this would have been
possible. I would also like to express great appreciation to all my participants for their time and
support to participate in my studies, even during COVID.
Finally, I want to thank the people who impacted me personally. I owe my parents for
everything. My journey was not possible without their support and love. It was difficult for them
to send their little girl so far away from home, but they have been the best and my number one
cheerleaders. Last but not least, I cannot thank my partner in crime, Aaron Thompson, enough
who’s extraordinary support and love made getting through this long journey a little bit easier.
All the dreams, wishes, and laughter we shared were my fuel to complete this marathon.

vii
Table of Contents
DEDICATION ............................................................................................................................................ ii
ACKNOWLEDGEMENTS ..................................................................................................................... iii
LIST OF TABLES ..................................................................................................................................... x
LIST OF FIGURES .................................................................................................................................. xi
ABSTRACT .............................................................................................................................................. xxi
CHAPTER 1 OVERVIEW ............................................................................................................................ 1
1.2. Significance ..................................................................................................................... 3
1.3. Specific Aims ................................................................................................................... 4
CHAPTER 2 BACKGROUND AND SIGNIFICANCE .................................................................................... 9
1.1. Motor Skill Learning ..................................................................................................... 9
1.1.1. Motor sequence learning ........................................................................................ 9
1.1.2. Motor adaptation .................................................................................................. 10
1.1.3. Phases of Motor Skill Learning ........................................................................... 12
1.1.4. Neural representations of Motor Skill Learning................................................ 14
1.2. Obstacle Negotiation as a Locomotor Skill ................................................................ 18
1.2.1. Visual information during obstacle negotiation ................................................. 19
1.3. Motor Skill Learning in Virtual Reality .................................................................... 20
1.4. Motor Skill Learning in Parkinson Disease ............................................................... 21
1.4.1. Challenges with Obstacle Negotiation ................................................................. 21
1.4.2. Motor and Locomotor Skill Learning ................................................................. 23
1.4.3. Context-Dependent Learning .............................................................................. 24
1.5. Explaining Individual Differences of Motor Learning in People with Parkinson
Disease ...................................................................................................................................... 25
1.5.1. Altered Resting-state Functional Connectivity in People with Parkinson
Disease 26
1.5.2. Associations between Resting-state Functional Connectivity and motor skill
acquisition and learning ...................................................................................................... 27
CHAPTER 3 The quality of visual information about the lower extremities influences
visuomotor coordination during virtual obstacle negotiation ............................................................. 29
Abstract .................................................................................................................................... 29
Introduction ............................................................................................................................. 30
Methods .................................................................................................................................... 33
viii
Results ...................................................................................................................................... 40
Discussion ................................................................................................................................. 45

CHAPTER 4 Locomotor skill acquisition in virtual reality shows sustained transfer to the real
world .......................................................................................................................................................... 52
Abstract .................................................................................................................................... 52
Background .............................................................................................................................. 53
Methods .................................................................................................................................... 57
Results ...................................................................................................................................... 65
Discussion ................................................................................................................................. 70
Conclusion ................................................................................................................................ 75
CHAPTER 5 People with Parkinson disease acquire a locomotor skill faster but retain less than
age-matched adults ................................................................................................................................... 76
Abstract .................................................................................................................................... 76
Introduction ............................................................................................................................. 77
Methods .................................................................................................................................... 80
Results ...................................................................................................................................... 92
Discussion ................................................................................................................................. 98
Limitations ............................................................................................................................. 103
Conclusions ............................................................................................................................ 104
CHAPTER 6 Corticostriatal resting-state functional connectivity is associated with locomotor
skill learning in people with Parkinson disease.................................................................................. 107
Abstract .................................................................................................................................................. 107
Introduction ........................................................................................................................................... 108
Methods .................................................................................................................................................. 112
Results ..................................................................................................................................................... 119
Discussion ............................................................................................................................................... 123
Conclusions ............................................................................................................................................ 127
CHAPTER 7 DISCUSSION ................................................................................................................. 128
Findings from our work ........................................................................................................ 129
Impact of Dissertation ........................................................................................................... 130
Limitations and future work ................................................................................................ 132
ix
APPENDIX A VALIDATION OF THE HIERARCHICAL BAYESIAN STATE-SPACE
MODEL: SIMULATION STUDY ...................................................................................................... 135
Results .................................................................................................................................... 145
Discussion ............................................................................................................................... 150
JAGS script in R for hierarchical Bayesian state-space model ........................................ 151
APPENDIX B DEVELOPMENT OF A NOVEL LOCOMOTOR LEARNING TASK FOR
REINFORCEMENT LEARNING ....................................................................................................... 153
Introduction ........................................................................................................................... 153
Methods .................................................................................................................................. 155
Results .................................................................................................................................... 163
Discussion ............................................................................................................................... 169
REFERENCES ....................................................................................................................................... 173

x
LIST OF TABLES
Table 3-1. Average and standard deviation of foot placement and clearance. Values are means
(M) ± standard deviation (SD). ..................................................................................................... 41
Table 4-1. Dependent variables derived from the NLME model. ................................................ 63
Table 5-1. Demographics and clinical gait and balance assessments. Results are presented as the
mean and standard deviation except for sex and Hohen and Yarh (H&Y). Sex and H&Y are
presented as the number of participants. Overground speed was calculated using 10 Meter Walk
Test (10MWT). ............................................................................................................................. 81
Table 5-2. Final regression result to explain the amount of locomotor learning with cognition. 98
Table 6-1. Participant characteristics. MDS-UPDRS: Movement Disorder Society – Unified
Parkinson’s Disease Rating Scale. H&Y: Hoehn and Yahr scale. LEDD: Levodopa equivalent
daily dosage in mg/day. MoCA: Montreal Cognitive Assessment. ............................................ 112
Table 6-2. Significant functional connectivity with learning parameters. ................................. 122
Table A-1. Description of computational models tested. p: participant number, g: group number
..................................................................................................................................................... 144
Table A-2. Summary of model comparisons. ............................................................................ 146
Table B-1. Summary statistics. Mean and standard deviation of performance error in meters and
success rate in a percentage during reinforcement learning trials for each participant. Success rate
was calculated only for practice trials. Bold texts indicate average performance error within the
reward zone. ................................................................................................................................ 164

xi
LIST OF FIGURES
Figure 1-1. Summary of significance. This work would help to create a virtuous cycle. When
patients visit the clinic for fall prevention programs or training after falls, clinicians would
deliver patient-specific gait training based on the patient characteristics targeting specific neural
correlates to maximize carry-over of the locomotor skills to the community. This would
ultimately reduce the number of falls and the number of visits to the clinic. ................................. 4
Figure 1-2. Summary of specific aims. Aim 1 assesses how the quality of visual information
about the body impacts obstacle negotiation performance in VR. Aim 2 determines how
individual differences in locomotor skill learning in VR influence retention and transfer of
learned skills to the real world. Aim 3-1 determines how PD influences the acquisition and
retention of skilled locomotor behavior in VR compared to healthy older adults. Aim 3-2 assesses
how changes in incidental context affect the retention of the recently learned locomotor skill in
individuals with PD and healthy older adults. Aim 4 investigates whether resting-state
corticostriatal FC is associated with inter-individual differences in motor learning and CDL in
individuals with PD......................................................................................................................... 8
Figure 2-1. A schematic of the classic view of motor learning. Motor learning occurs in three
transient phases: Cognitive, associative, and autonomic phases. During the cognitive phase,
performance error rapidly reduces. During the associative phase, performance error slowly
reduces, and the skill is further refined. During the autonomous phase, performance error no
longer reduces and performance plateaus. Conventionally in motor learning research, skill
acquisition is referred to as the cognitive and associative phases, and skill retention is referred to
as the autonomic phase. During skill acquisition, studies investigate how fast the skill is acquired
xii
(learning rate) and how much the skill is acquired (amount of improvement). During skill
retention, studies find how much the skill is retained (amount of retention). .............................. 13
Figure 2-2. Theoretical framework of corcito-striato-cerebellar circuitry involved in different
phases of motor skill learning based on changes in brain activities. Both the basal ganglia and
cerebellum contribute to skill acquisition regardless of motor skill learning processes. During
early learning, neural substrates known for ‘associative’ area such as the caudate nucleus,
anterior putamen, posterior cerebellar cortices, and frontal associative cortices are activated.
During late learning, neural substrates known for ‘sensorimotor’ area such as the putamen, deep
cerebellar nuclei, and sensorimotor and parietal cortices are activated. When the skill is
consolidated and retained, areas of neural activations differ between the type of motor skill
learning. During motor sequence learning, the putamen and sensorimotor cortices are activated.
Whereas, during motor adaptation, the cerebellum and sensorimotor cortices are activated. ...... 18
Figure 3-1. A) Virtual corridor with obstacles. Participants were able to see their score and one
point was deducted for each collision. B) Visual feedback conditions from a third person
viewpoint. Spheres represent the position of markers placed on the lower extremities. Segments
connecting the spheres were used to provide a visual representation of limb segment length. .... 35
Figure 3-2. A) Experimental protocol. Participants were counterbalanced to begin the
experiment with either obstacle-free trials or obstacle negotiation trials. The box above each trial
represent three feedback conditions: 1) no model as represented by a horizontal line at the floor
level, 2) end-point feedback as represented by two dots and 3) link-segment feedback as
represented by dots at the foot, knee and hip and sticks connecting the dots. B) Schematic figure
showing dependent variables. The solid and dashed lines represent the leading and trailing limbs
xiii
during obstacle crossing, respectively. Each step number was defined as the time period between
two consecutive heel strikes (HS). C) Representative time series of head pitch angle during an
obstacle negotiation trial. The solid black line refers to the head angle. The vertical grey lines
refer to point along the corridor where the center of the obstacle was located. D) Representative
time series data for heel trajectory. The dashed grey line refers to the left heel trajectory and the
solid grey line refers to the right heel trajectory. Black boxes are scaled to the height and width
of the obstacles along the path. ..................................................................................................... 39
Figure 3-3. Box and whisker plot illustrating leading foot placement variability. Horizontal lines
within each box indicate median values, and the bottom and top boundaries of the box indicate
the 25th and 75th percentiles. Points outside each box indicate outliers. The asterisk denotes a
significant difference at the Bonferroni corrected p < 0.05 level. ................................................ 40
Figure 3-4. Average head pitch angle as a function of step number and feedback condition. Error
bars indicate the standard error. The black line refers to the condition when no visual information
about lower extremities was provided, the dark grey line refers to the condition when an end-
point foot representation was provided, and the light grey line refers to the condition when the
link-segment leg representation was provided. ***: Bonferroni corrected p<0.005, ****:
Bonferroni corrected p<0.001 ....................................................................................................... 43
Figure 3-5. Estimated regression coefficients for the dependent variables from linear models
describing the relationship between (A) leading foot placement after the obstacle and head angle
during the approach step, (B) leading foot placement after the obstacle and head angle during
lead foot crossing, and (C) trailing foot clearance and head angle during trailing foot crossing in
each feedback condition. Negative coefficients indicate that greater downward head angle was
xiv
associated with larger values for each crossing variable. The illustrations in the left column
represent the period analyzed for head angle and the corresponding crossing variable. The solid
black lines indicate the analyzed step for head angle (thin black arrows) and the respective
obstacle crossing performance variable (solid grey arrows). Error bars represent standard errors.
HS: Heel Strike. **: Bonferroni corrected p<0.01, ***: Bonferroni corrected p<0.005. ............. 45
Figure 4-1. Experimental setup and protocol. (A) Virtual corridor with obstacles and an eye-
level display of participants' current score. (B) Visual feedback of the lower extremities viewed
from a third-person perspective. Spheres represent the position of markers placed on the lower
extremities. Line segments connecting the spheres were used to provide a visual representation
of limb segment length. During the study, participants viewed the representation of the lower
extremities from a first-person viewpoint. (C) Schematic diagram of the mapping between the
participant’s performance and the auditory feedback they received. (D) Over-ground obstacle
negotiation setup. (E) Experimental protocol illustrating the day of the study, the trial type,
number of obstacles per trial, and whether auditory performance feedback was provided. ......... 58
Figure 4-2. Individual foot clearance data and fit from the NLME model. Gray points represent
foot clearance during each obstacle crossing in VR on Day 1, the black curve represents the
participant-specific fit from the NLME, and the dashed curve represents the group level fit of the
NLME. The black points and error bars after the gray dashed vertical line represent average and
standard deviation foot clearance in VR on Day 2, respectively. ................................................. 67
Figure 4-3. Over-ground transfer on Day 1 and Day 2. (A) Transfer to over-ground walking on
Day 1. Here, reductions in foot clearance indicate improvements in skill. (B) Over-ground
retention on Day 2. All data are reported as boxplots with the horizontal lines within each box
xv
indicating median values and the bottom and top boundaries of each box indicating the 25th and
75th percentiles. Dark gray points represent individual data points and the gray lines connecting
the represent the change in foot clearance across trials. (C) Trials during over-ground obstacle
negotiation. Black points represent average foot clearance across all participants and gray
vertical lines represent standard deviations. BASE: baseline block for over-ground on Day 1, TF:
transfer block for over-ground on Day 1, and RET_OG: retention block for over-ground on Day
2. The asterisks (***) indicate statistically significant differences from zero at p<0.001............ 68
Figure 4-4. Associations between performance on Day 1 and retention on Day 2. (A)
Relationship between foot clearance during retention in VR on Day 2 (RET_VR) and the final
foot clearance in VR on Day 1. (B) Relationship between over-ground foot clearance during
retention on Day 2 (RET_OG) and over-ground foot clearance during transfer (TF) on Day 1.
Each participant is represented by a single data point, the solid black line represents the
regression fit, and the dashed gray lines are 95% confidence intervals. ....................................... 69
Figure 5-1. Experimental setup and protocol. (A) Schematic of the experimental setup. (B) The
virtual environment. (C) Model of the lower extremities. Participants viewed the model from a
first-person perspective. (D) Experimental protocol. BASE: baseline block, RET: retention
block. (E) An example of foot trajectories for two obstacles with prescribed success ranges in
yellow. (F) Collision (empty sound icon) and task performance (filled sound icon) sound
feedback. ....................................................................................................................................... 85
Figure 5-2. Foot clearance during BASE for (A) low obstacles and (B) high obstacles. Each dot
represents one participant. ............................................................................................................ 93
xvi
Figure 5-3. Results of Bayesian modeling. (A) Observed and estimated foot clearance in the
course of trials for an example participant. Each point represents the observed foot clearance on a
given obstacle and the blue line represents estimated foot clearance from the state-space model.
(B) The posterior probability of the estimated learning rate for an example participant. (C) The
posterior probability of the estimated inference for an example participant. (D) Median of the
subject-specific posterior probabilities for learning rate for each group. (E) The posterior
probability for the group-level learning rate. (F) The posterior probability of the effect size of
group differences on learning rate. (G) Subject-specific median interference for each group. (H)
The posterior probability of the group-level inference. (I) The posterior probability of the effect
size for the group-level interference. Red and blue represent people with PD and age-matched
controls, respectively. ................................................................................................................... 94
Figure 5-4. (A) Initial performance error in PD and Control group. Red dots represent
participants with PD and blue dots represent control participants. The black solid line represents
a median for each group. (B) A correlation between performance error during initial practice and
learning rate. Red dots represent participants with PD and blue dots represent control
participants. The black solid line and gray shaded area indicate a fit and 95% confidence
intervals from the correlation regardless of the group. ................................................................. 96
Figure 5-5. Performance error during retention compared to initial practice as a measure of
retention in the SAME and SWITCH context in people with PD and controls. SAME context
illustrates retention and SWITCH context illustrates context-dependent learning. Red dots
represent the PD group and the blue dots represent the control group. Solid lines represent
medians for each condition for each group. Dotted lines represent zero ...................................... 97
xvii
Figure 5-6. Adjusted regression plots for the amount of learning as a function of memory. The
gray dots represent participants with PD. The solid red line represents the regression fit, and the
dashed lines are 95% confidence intervals. .................................................................................. 98
Supplementary Figure 5-1. Observed and estimated foot clearance. Each panel represents a
participant. Gray dots represent observed foot clearance in the course of locomotor learning. Red
and blue solid lines represent foot clearance estimated by the Bayesian modeling for people with
PD and the controls, respectively................................................................................................ 105
Supplementary Figure 5-2. Group-level posterior probabilities for learning rate with the subset
of participants. (A) The posterior probability for the group-level learning rate. (B) The posterior
probability of the effect size of group differences on learning rate. ........................................... 106
Figure 6-1. Experimental protocol. BASE: Baseline obstacle crossing. MDS-UPDRS:
Movement Disorder Society-Unified Parkinson’s Disease Rating Scale. RET: Retention. ....... 114
Figure 6-2. Flowchart of data processing and analyses. (A) Structural and functional image pre-
processing. (B) Seed extraction. (C) Partial linear regression for each run for each participant to
regress out nuisance regressors. (D) Seed-to-voxel analysis using general linear modeling to
obtain a seed-to-voxel spatial map across all runs for each participant. (E) Extraction of
parameter estimates with the cortical seed masks....................................................................... 119
Figure 6-3. (A) The correlation between the functional connectivity with the DLPFC and
learning rate estimated from the state-space model. Each data point represents a median of the
posterior distribution from the hierarchical Bayesian estimation for each participant. The error
bars represent 89% CIs. The solid line and shaded area indicate the fit and CIs of the correlation,
xviii
respectively. (B) The correlation between rsFC between the posterior putamen and M1 and
retention. Each data point represents each participant. The solid line and shaded area indicate the
fit and CIs of the correlation, respectively. (C-D) p-value and r
2
from the Bootstrap results for the
correlation between rsFC between the anterior putamen and the DLPFC and learning rate. The
dotted line indicates the median of the bootstrap result. ............................................................. 121
Figure 6-4. Regions whose functional connectivity was associated with learning rate for (A) left
and (B) right anterior putamen. DLPFC: dorsolateral prefrontal cortex. ................................... 123
Figure A-1. A flowchart describing artificial dataset generation and Bayesian estimation using
Markov Chain Monte Carlo (MCMC) using Gibbs sampling. ................................................... 139
Figure A-2. Example convergence results. (A) Trace plot of all chains that visualize 𝑹𝑹 for
learning rate. Each color represents a chain. The overlap across the chains and the lack of a
systematic change in in learning rate as a function of iterations are indicative of 𝑹𝑹 close to 1. (B)
Rank plots that visualize rank-normalized 𝑹𝑹 for learning rate. Each color represents a chain. All
chains show uniform distributions across the ranks, indicating convergence. ........................... 143
Figure A-3. Posterior distributions of learning rate. Learning rate density plots for (A) Group 1
and (B) Group 2. The red dashed line represents true learning rate for each participant, and the
gray dashed line represents estimated learning rate using the least squares algorithms (fmincon).
The black curve represents the posterior distribution from the MCMC. (C) Difference between
true and estimated learning rate by MCMC (Gray dots) and least squares (empty black dots)
algorithms. Each dot represents one participant. For MCMC, 89% highest density intervals
(HDI) are plotted as error bars. ................................................................................................... 147
xix
Figure A-4. Posterior distribution of interference. Interference density plots for (A) Group 1 and
(B) Group 2. (C) Difference between true and estimated interference. The description of the
figure is the same as Figure A-3. ................................................................................................ 149
Figure B-1. Experimental protocol and setup. (A) Experimental protocol and design. On Day 1,
participants performed baseline trials (BASE) with the instruction to step over virtual obstacles
as naturally as possible. Following BASE, participants received an instruction that there was an
invisible “reward zone” that they needed to find based on a binary sound feedback. The lower
threshold of the reward zone was 8 cm higher than mean foot clearance during BASE. The upper
threshold was 4 cm higher than the lower threshold. Participants performed a total of 6 practice
blocks. The second last block on Day 1 was no feedback trials (NFB) to test a short-term
retention. During NFB, there was no sound feedback provided. After 24 hours (Day 2),
participants revisited the lab and performed retention trials (RET). Participants did not receive
sound feedback during RET. (B) Virtual environment. (C) Lower extremity representation. (D)
Experimental setup. Participants walked on a treadmill while wearing a head-mounted display
and holding the handrails lightly................................................................................................. 158
Figure B-2. Experimental protocol and raw data on Day 1 for an example participant. (A) RL
group protocol. (B) Control group protocol. (C) An example foot clearance data of a single
participant with the adaptive reward zone on Day 1. .................................................................. 163
Figure B-3. Performance error as a function of trials. The dotted horizontal lines indicate the
reward zone. Blue dots indicate trials within the reward zone, and gray dots indicate trials that
were not within the reward zone. Each subplot represents one participant. The order of the trials
xx
was as follows: Baseline (BASE), reinforcement learning practice (REWARD), and retention
(RET). RET was performed after 24 hours. ................................................................................ 165
Figure B-4. (A) Success rate for each trials. Each data point represents each participant. (B)
Mean performance error for the last five trials during BASE and for the first five trials during
RET. Each color represents one participant. The dotted horizontal lines represent the lower and
upper threshold of the reward zone. ............................................................................................ 166
Figure B-5. Individual foot clearance during reinforcement learning and control trials. Trials for
the RL group were baseline (BASE), pre-training (PRE), and reward trials (REWARD) on Day 1
and PRE and REWARD on Day 2. Trials for the Control group were BASE, PRE, and trials
without reward feedback (Control) on Day 1 and PRE and Control on Day 2. The gray dotted
line in PRE trials indicate a change in trials from PRE-FB to PRE-NFB. In the REWARD group
figure, the thick black line indicates the changes in reward zones. Blue dots represent successful
trials. Each subplot containing Day 1 and Day 2 represents one participant. ............................. 167
Figure B-6. (A) Mean foot clearance for each block. (B) A number of trials required to adapt to
the desired lower threshold of the reward zone. Each data point represents each participant.
Dotted lines indicate more trials to achieved the desired lower threshold on Day 2 than Day 1. C)
Success rate in each block of reward trials. Each data point represents each participant........... 169

xxi
ABSTRACT
Individuals have an incredible capacity to learn new motor skills, and can retain the
learned skill in a long-term period and flexibly transfer the skill to other environments. This
capacity in motor learning is the foundation of neurorehabilitation. Particularly, learning
dynamic walking skills relevant to community walking is a vital element of rehabilitation as
walking in dynamical environments such as negotiating obstacles is critical in independent
community participation.
Chapter 1 provides the overview, significance, and specific aims of the study. Chapter 2
reviews previous literature regarding a general overview of motor skill learning and obstacle
negotiation as a locomotor skill. We also summarize previous literature about motor skill
learning in virtual reality. Moreover, we review how PD influences motor skill learning and
context-dependent learning during locomotion. Lastly, we review how PD alters resting-state
functional connectivity and previous literature demonstrating associations between resting-state
functional connectivity and motor skill acquisition and learning.
Chapter 3 aimed to assess how the quality of visual information about the body impacts
obstacle negotiation performance in VR. Participants stepped over virtual obstacles with varying
levels of visual information about the lower extremity: 1) no model, 2) end-point model, and 3)
link-segment model. We found that absence of visual information about the lower extremities led
to an increase in the variability of leading foot placement after crossing. Moreover, the presence
of visual information about the lower extremities promoted greater downward head pitch angle
during the approach to and subsequent crossing of an obstacle. In addition, having greater
downward head pitch was associated with closer placement of the trailing foot to the obstacle,
further placement of the leading foot after the obstacle, and higher trailing foot clearance. These
xxii
results demonstrate that the quality of visual information about the lower extremities influences
both feed-forward and feedback aspects of visuomotor coordination during obstacle negotiation.
Chapter 4 aimed to demonstrate how individual differences in locomotor skill learning in
VR influence retention and transfer of learned skills to the real world. Participants practiced
stepping over virtual obstacles on Day 1 with instruction to minimize the vertical distance
between the foot height and obstacle based on auditory feedback. They also performed transfer
trials over-ground before and after practicing obstacle negotiation in VR. On Day 2, participants
performed retention trials both in VR and over-ground. On Day 1, participants systematically
reduced foot clearance throughout practice and transferred this reduction to over-ground
walking. The reduction in foot clearance was also retained after 24 hours in VR and over-ground.
Overall, our results support the use of VR for locomotor training as skills learned in a virtual
environment readily transfer to real-world locomotion.
Chapter 5 aimed to determine how PD influences the acquisition and retention of skilled
locomotor behavior in VR compared to age-matched adults. Moreover, we aimed to assess how
changes in incidental context affect the retention of the recently learned locomotor skill in people
with PD and age-matched adults. Finally, we aimed to determine the associations between
cognition and locomotor skill retention in people with PD. On Day 1, participants with PD
performed a comprehensive battery of cognitive assessments. On Day 2, participants practiced
stepping over virtual obstacles with instruction to maintain their foot clearance within prescribed
success range based on auditory feedback. On Day 3, participants performed retention trials in
either the same environment as the practice environment or a different environment. We found
that all participants could acquire the locomotor skill, but people with PD had less retention.
However, there was no effect of environmental context on retention performance. Moreover, the
xxiii
clinical assessment of memory was significantly associated with locomotor skill retention in
people with PD. These results demonstrate that PD negatively influences locomotor skill
learning, but people with PD do not exhibit context-dependency during locomotor skill retention.
Further, memory may be an important clinical predictor to pinpoint patients who require
additional gait training to facilitate retention.
Chapter 6 aims to determine the associations between resting-state corticostriatal
functional connectivity and locomotor skill learning in people with PD. Prior to the locomotor
skill acquisition task, we obtained resting-state functional MRI from participants with PD. We
found that the cognitive circuit of the basal ganglia, identified by rsFC between the anterior
putamen and the DLPFC, was significantly associated with learning rate. Our result suggests that
the cognitive circuit of the basal ganglia may be a useful marker in designing targeted gait
intervention for people with PD.
Together, we found that the quality of visual information about lower extremities
influences consistency of obstacle negotiation performance. Moreover, the results of transfer and
retention of locomotor skills. Together, these results provide important evidence for designs of
VR applications and potential efficacy of VR-based locomotor training interventions. Further,
PD negatively influenced locomotor skill retention, but people with PD retained the locomotor
skill similarly regardless of the environmental context. Inter-individual variability in locomotor
skill learning in people with PD was explained by the clinical assessment of memory and the
cognitive circuit of the basal ganglia. Overall, a measure of cognition and its associated
functional connectivity of the brain may be a critical source of individual variability in locomotor
skill learning and can be potential ingredients for patient-specific physical interventions.

1
CHAPTER 1
OVERVIEW
1.1. Background
Every year, 50-68% of people with Parkinson disease (PD) experience one or more falls
(Bloem et al. 2001; Gray and Hildebrand 2000; Wood et al. 2002) related to walking (Stack and
Ashburn 1999). As one of the most common reasons for falls in people with PD is during
obstacle negotiation, gait training emphasizes the importance of skilled walking intervention
such as obstacle negotiation (Morris 2006; Morris et al. 2010). However, the damage to the basal
ganglia deteriorates abilities to learn motor skills and retain them. These abilities refer to as
motor learning, which is a foundation of gait training (Kleim and Jones 2008). Consequently, it
is important to understand the effect of PD on motor learning. Moreover, people with PD often
have difficulty executing learned skills in different contexts (Lee et al. 2015; Nieuwboer et al.
2001, 2009; Onla-or and Winstein 2008) even when the different contextual features, such as the
background scene, are not directly associated with the tasks, also known as incidental contexts
(Lee et al. 2015; Lee and Fisher 2017; Nieuwboer et al. 2001). This phenomenon is referred to as
context-dependent learning (CDL) (Wright and Shea 1991). CDL may arise because people with
PD overly rely on incidental contexts when they learn a new skill as a compensatory strategy to
overcome impairments in neural networks involved in skill learning (Lee and Fisher 2017;
Nakamura et al. 2001). Although deficits in motor learning and aberrant CDL has been observed
during upper extremity tasks in people with PD, limited evidence during locomotion exists.
Due to the heterogeneity of the disease, the level of deficits in motor learning and CDL
are differently present across people with PD. This individual variability in motor learning and
CDL may be due partly to individual differences in neural networks in people with PD. Recent

2
studies suggest that intrinsic functional organization of the brain, which can be measured using
resting-state functional connectivity (rsFC), may reflect a readiness to perform or learn future
motor skills. Specifically, consistent with the tight interconnectedness between the basal ganglia
and motor skill learning, basal ganglia rsFC has been associated with motor skill acquisition and
retention in healthy young adults (Hamann et al. 2014; Mary et al. 2017; Stillman et al. 2013).
Given that the motor symptoms of individuals with PD are expressed heterogeneously, possibly
due to differences in the extent of impairment in the basal ganglia and related cortical areas
(Albin and Dauer 2012; Holiga et al. 2013), rsFC could explain individual variability of motor
learning and CDL during locomotion in people with PD.
Here, we will use a novel locomotor learning paradigm in virtual reality (VR) to
investigate the effects of PD for locomotor skill learning and CDL and if rsFC explains
individual differences in locomotor learning. The use of VR is rapidly growing in the field of
motor learning and rehabilitation, but its efficacy in retention and transfer of motor skills is not
well-understood. Moreover, the immersiveness of VR eliminates visual information about the
user’s body, which can influence motor control during learning. Therefore, our preliminary
works will address how the quality of visual information about the body impacts locomotor
control during obstacle negotiation, and how young adults retain a locomotor skill and transfer
the skill to the real world. The results of these works will lead to a better understanding of
neurological factors influencing the acquisition and retention of locomotor skills in people with
PD and older adults. The behavioral and neuroimaging evidence we establish may help the
clinical community improve current interventions to enhance the generalization of locomotor
skill training to the community.

3
1.2. Significance
This dissertation has theoretical and translational significance. This work contributes to
an understanding of the effect of PD on locomotor skill learning. Despite the important of gait
training in people with PD, motor learning has been primarily investigated with upper extremity
tasks. As some symptoms of gait and balance are resistant to dopamine replacement therapy, this
work will determine how PD influences motor learning in the context of locomotion. Also, this
work will determine if the intrinsic functional organization of the brain, particularly with the
basal ganglia, explains the heterogeneity of locomotor skill learning in people with PD. The
basal ganglia circuits serve an important link to motor skill learning. However, a direct
association between the basal ganglia circuits and locomotor skill learning remains unclear.
Therefore, this work determines how the intrinsic functional connectivity of the basal ganglia
explains individual differences in locomotor skill learning in people with PD.
This work also contributes to evidence-based practice by investigating the fundamentals
of motor learning using VR. VR is a promising tool to provide dynamic, engaging, and more
ecological gait training. However, as VR is a relatively new tool, it remains to be seen which
sources of visual feedback about the body and the environment are necessary to facilitate skillful
obstacle negotiation. It is also not well-understood how locomotor learning occurs in VR and
how individuals transfer locomotor learning outside of VR. Therefore, our works address the
efficacy of VR in motor learning by understanding how the quality of visual information about
the body affects visuomotor control during obstacle negotiation, and if individuals carry over
locomotor skills to the real world or after a time delay. In addition, flexible generalization and
carry-over of motor skills across different contexts is the ultimate goal of neurorehabilitation.
The findings of this work will provide clinical community potential factors that impact motor

4
learning in varying contextual features. Further, our neuroimaging approach allows us to probe
the factors that contribute to individual differences in locomotor learning in people with PD,
providing valuable markers to consider in precision rehabilitation. Overall, understanding the
behavioral significance of the altered brain networks offers potential ingredients of more
effective, patient-specific clinical interventions for locomotor learning based on each
individual’s neural correlates and their impairments.

Figure 1-1. Summary of significance. This work would help to create a virtuous cycle. When
patients visit the clinic for fall prevention programs or training after falls, clinicians would
deliver patient-specific gait training based on the patient characteristics targeting specific
neural correlates to maximize carry-over of the locomotor skills to the community. This would
ultimately reduce the number of falls and the number of visits to the clinic.

1.3. Specific Aims
Aim 1: Assess how the quality of visual information about the body impacts obstacle
negotiation performance in VR. Given the importance of visual information about the body for
improving the perception of distance and estimating performance capability, it is plausible that

5
the quality of visual information about the body influences coordination during virtual obstacle
negotiation. Healthy young adults will practice stepping over obstacles in VR on a treadmill to
minimize foot clearance, and they will perform a transfer test to the real world. Participants will
step over virtual obstacles while walking on a treadmill with one of three types of visual
feedback about the lower extremities: no feedback, end-point feedback, or a link-segment model.
We hypothesize that 1) restriction of visual information about the body will lead to increased
variability of foot placement during obstacle negotiation and 2) increased downward head pitch
angle, a proxy of gaze, when approaching obstacles.
Aim 2: Determine how individual differences in locomotor skill learning in VR
influence retention and transfer of learned skills to the real world. Given the growing use of
VR for motor skill learning in research and rehabilitation, it is important to understand the
efficacy of VR in long-term retention as well as transfer to the real world. On Day 1, healthy
young adults will practice stepping over obstacles in VR on a treadmill to minimize foot
clearance, and they will perform a transfer test to the real world. On Day 2, participants will
perform a retention test in VR and the real world. We hypothesize that 1) participants will reduce
foot clearance in VR during practice, 2) the reduced foot clearance in VR will transfer to the real
world obstacle negotiation, and 3) the reduced foot clearance will be retained in VR and the real
world after 24 hours.
Aim 3-1: Determine how PD influences the acquisition and retention of skilled
locomotor behavior in VR compared to age-matched adults. Despite the importance of
locomotor training for people with PD, it is yet to be determined how PD affects acquisition and
long-term retention of skilled locomotion. On Day 1, individuals with PD and older adults will
step over virtual obstacles of two different heights associated with specific success ranges. The

6
goal is to step over an obstacle while maintaining their foot within the success range during the
crossing. Auditory feedback will be provided based on their performance of foot clearance. On
Day 2, half of the participants will perform a retention test in the same virtual environment. We
will calculate composite scores for cognitive domains of memory, attention, visuospatial
processing, and executive functions from a comprehensive neuropsychological battery only in
people with PD to investigate if cognition explains individual variability in motor learning in
people with PD. We hypothesize that 1) people with PD will acquire the locomotor skill more
slowly and exhibit a greater performance error compared to controls 2) people with PD will have
greater performance error during retention compared to controls, and 3) composite scores of
cognition will be positively associated with learning rate and the amount of learning.
Aim 3-2: Assess how changes in incidental context affect the retention of the
recently learned locomotor skill in people with PD and age-matched adults. An important
aspect of motor learning for people with PD is to flexibly perform learned motor skills in varying
contexts. Therefore, understanding the level of interference from incidental contexts during
motor learning is critical. On Day 2 in Aim 3-1, the other half of the participants will perform a
retention test in the incidentally different virtual environment. The floor of the virtual
environment will be changed to a different texture, and the color of the walls and ceiling will be
changed. We hypothesize that people with PD will have a greater performance error in the
switched context compared to controls due to their tendency for CDL.
Aim 4: Determine the relationship between corticostriatal rsFC and inter-individual
differences in motor learning in people with PD. The heterogeneity of the disease results in
large variability in motor learning in people with PD. This variability may be explained by
intrinsic functional organization of the brain. Therefore, we will measure corticostriatal rsFC

7
using resting-state functional magnetic resonance imaging (fMRI) before locomotor skill
acquisition. The FC will be correlated with inter-individual differences in skill acquisition and
learning (Aim 3-1). We hypothesize that (1) the strength of rsFC in the cognitive corticostriatal
circuit will be positively correlated with the inter-individual differences in locomotor skill
acquisition, and (2) the strength of rsFC in the sensorimotor circuit will be positively correlated
with individual differences in skill retention.

8

Figure 1-2. Summary of specific aims. Aim 1 assesses how the quality of visual information
about the body impacts obstacle negotiation performance in VR. Aim 2 determines how
individual differences in locomotor skill learning in VR influence retention and transfer of
learned skills to the real world. Aim 3-1 determines how PD influences the acquisition and
retention of skilled locomotor behavior in VR compared to healthy older adults. Aim 3-2
assesses how changes in incidental context affect the retention of the recently learned
locomotor skill in individuals with PD and healthy older adults. Aim 4 investigates whether
resting-state corticostriatal FC is associated with inter-individual differences in locomotor
learning in individuals with PD.

9
CHAPTER 2
BACKGROUND AND SIGNIFICANCE
1.1. Motor Skill Learning
Motor skill learning by trial-and-error is an essential foundation of our daily lives.
Particularly, adapting the motor skills during walking to various environmental constraints is
critical to navigating predictable and unpredictable environments in the community. Typically,
motor skills are acquired rapidly early in learning, and then the rate of improvement gradually
slows down over multiple training sessions until performance reaches nearly plateau. The
mastery of motor skills leads to sustained performance for an extended time as well as in a
changing environment. The improvement of the skill through repetition and retention of the
improved skill is an integral part of motor learning (Schmidt and Lee 2011). These skills are
acquired through various processes of motor skill learning. Two important processes are motor
sequence learning and motor adaptation (Doyon and Benali 2005; Seidler 2010).
1.1.1. Motor sequence learning
Motor sequence learning is a process that combines a set of motor components or sub-
movements, which can be either discrete or continuous, into a predetermined sequence with
increasing spatial and temporal accuracy (Krakauer et al. 2019; Seidler 2010). For instance,
when an individual kicks a ball, he places his standing foot at the right location near the ball, lifts
his other leg backwards, then swing this leg with a proper amount of ankle plantarflexion to kick
the ball. These discrete actions are accompanied by a series of continuous muscle activations to
allow these discrete movements. The goal of motor sequence learning is to make fast and
accurate movement by integrating compartmentalized sub-movements, which is a critical part in
motor skill learning.

10
Motor sequence learning is primarily studied using a paradigm that imposes learning of
discrete sequential orders. These tasks provide a set of discrete sequences that individuals must
follow a pre-defined order by key pressing (Abrahamse et al. 2013; Karni et al. 1998; Robertson
2007; Willingham 1998), arm reaching (Ghilardi et al. 2003, 2009), saccadic eye movements
(Gersch et al. 2004; Hayhoe and Ballard 2005; McSorley et al. 2019; Sternberg et al. 1978), and
stepping (Du and Clark 2017, 2018). Learning is assessed by how quickly and accurately
individuals perform the movement. The learning process of faster and more accurate movement
is proposed that the discrete individual sequences merge into a higher-level representation of the
entire sequence (Lashley 1951; Rosenbaum et al. 2007; Yokoi and Diedrichsen 2019). As the
individual sequences are learned, an abstract representation of the sequences may emerge in a
form of ‘chunks’ (Lashley 1951). The chunk clusters a set of subcomponents of the sequences in
a specific order. Eventually, the representation of the entire sequences appears that encompasses
multiple chunks to allow easy access to plan and recall the sequence of actions (Rosenbaum et al.
1983). Given the representation of the entire sequences, individuals perform sequence learning
tasks faster and more accurately by planning a number of chunks, or the entire sequence,
together (Rosenbaum et al. 1983; Sakai et al. 2003).

1.1.2. Motor adaptation
Motor adaptation is a process that modifies an already well-practiced motor command in
response to changes in the environment or the body (Krakauer et al. 2019). For instance, when an
individual steps over an unconventionally high obstacle, she is likely to modify the existing
motor command to step over the obstacle, rather than creating a new motor command entirely

11
different from the existing ones. This adjustment is fundamental in rehabilitation after a brain
injury, where the dynamics of the interaction between the body and the brain has changed.
Numerous approaches have showed motor adaptation in laboratory settings. Especially,
motor adaptation during gait, or locomotor adaptation, is often explored using a split-belt
treadmill, which has separate belts for the right and left legs (Bastian 2008; Choi and Bastian
2007; Dietz et al. 1994). When the two belts move in a different speed, this speed mismatch
results in asymmetric walking patterns, such as step length, and the asymmetry converts to more
symmetric walking patterns in an adaptive manner (Dietz et al. 1994; Reisman et al. 2005). Once
the motor system is recalibrated to this perturbed environment, performance is relatively stable
(Malone et al. 2012) or subtly changing in the context of perturbation (Sánchez et al. 2020). A
removal of this perturbation causes a deviation of performance to the opposite direction to the
perturbation direction (Malone et al. 2012; Yokoyama et al. 2018). The deviation rapidly returns
to the level of performance prior to the perturbation (Malone et al. 2012; Yokoyama et al. 2018).
This implicit and rapid aftereffect is commonly referred to as ‘washout’. The existence of
washout supports the idea that an already well-learned motor command is subject to be modified
by the external perturbations, which is the fundamental aspect of motor adaptation. Other
approaches for locomotor adaptation, such as elastic force fields (Fortin et al. 2009), robotic
exoskeletons (Emken et al. 2007; Gordon and Ferris 2007), or added loads on the limb (Smith
and Martin 2007), as well as motor adaptation during upper extremity tasks, such as visuomotor
rotation (Krakauer et al. 2000), robotic devises (Duarte and Reinkensmeyer 2015;
Reinkensmeyer and Patton 2009), prism goggles (Redding et al. 2005), or force fields (Burdet et
al. 2001; Shadmehr and Mussa-Ivaldi 1994; Smith et al. 2006), demonstrate ubiquitous processes
of motor adaptation as described above.

12

1.1.3. Phases of Motor Skill Learning
The classic view on motor skill learning by Fitts and Posner suggests that learning occurs
in three transient phases: Cognitive, associative, and autonomous phases (Fitts and Posner 1967;
Schmidt and Lee 2011) (Figure 2-1). First, during the cognitive phase, performance is inaccurate,
time-consuming, and effortful (Fitts 1964; Lee et al. 1994). This phase is predominantly
accompanied with verbal instructions (Bobrownicki et al. 2015) and cognitive efforts (Lee et al.
1994) to learn the skill. Correspondingly, the skill is gained quickly (Fitts 1964) with the
assistance of error detection, typically shown during early motor skill acquisition. This phase is
often equated to the explicit process of motor learning (Krakauer et al. 2019).
Second, during the associative phase, movements become faster and more accurate (Fitts
1964), but the learning process is slower and subtler than the cognitive phase (Taylor and Ivry
2012). Performance is subtly adjusted to be more automatic, and a person often barely perceives
this change in performance. This phase occurs within hours or days of skill practice and is shown
during a later or final period of motor skill acquisition or short-term retention (Dayan and Cohen
2011). However, movement is context-specific such that the accuracy and automaticity of
performance suffer when performance context changes (Marinelli et al. 2017; Ruitenberg et al.
2012).
Lastly, the autonomous phase refers to the time when the skill has become automatic
after it has been practiced for a much longer time—i.e., months or years—or during long-term
retention (Fitts 1964). In this phase, a person uses only little attentional resources and is hardly
interfered with by secondary activities (Abernethy 1988). Concurrently, the movement becomes
effector-specific rather than context-specific, meaning that changes in context do not interfere

13
with the specific performance of the skill (Ruitenberg et al. 2012). The transition to effector-
specific movement allows individuals to transfer the learned motor skill to novel contexts
(Ahmed and Wolpert 2009; Fitts 1964). Overall, the movement in this phase is characterized as
being effortless, precise, and efficient without awareness of movement details (Marinelli et al.
2017). This phase of motor learning is equated to a more implicit process (Krakauer et al. 2019).

Figure 2-1. A schematic of the classic view of motor learning. Motor learning occurs in
three transient phases: Cognitive, associative, and autonomic phases. During the cognitive
phase, performance error rapidly reduces. During the associative phase, performance error
slowly reduces, and the skill is further refined. During the autonomous phase, performance
error no longer reduces and performance plateaus. Conventionally in motor learning
research, skill acquisition is referred to as the cognitive and associative phases, and skill
retention is referred to as the autonomic phase. During skill acquisition, studies investigate
how fast the skill is acquired (learning rate) and how much the skill is acquired (amount of
improvement). During skill retention, studies find how much the skill is retained (amount of
retention).

14
In motor adaptation, these early and late phases of learning are described by two learning
processes. Smith and colleagues first demonstrated that motor adaptation composites two distinct
timescale learning curve, modeled by a state-space model (Smith et al. 2006). These two learning
curves are fast initial learning, simultaneously followed by gradual and slower learning. These
components of motor adaptation are named a “fast” process and a “slow” process, respectively
(Smith et al. 2006). The fast process includes a high learning rate, especially in the beginning,
followed by a rapid decay. The slow process includes a slower, gradual learning rate, but with a
slower decay. The sum of these two components is the global performance. This two-component
motor adaptation model accurately captures the trajectory of motor adaptation (Smith et al.
2006), indicating the two-state-space model can mimic the statistical representation of motor
adaptation. However, further investigation suggests that there may be more than two
components of learning processes during motor adaptation, particularly multiple slow processes
(Lee and Schweighofer 2009).

1.1.4. Neural representations of Motor Skill Learning
Various brain structures forming corticostriatal and corticocerebellar circuits contribute
to mediate different phases of motor skill learning (Bostan and Strick 2018; Doyon and Benali
2005; Ferrazzoli et al. 2018). The patterns of neural activations over the course of motor learning
follow anatomically and functionally known corticostriatal (Alexander et al. 1986; Haber 2016)
and corticocerebellar circuitry (Koziol et al. 2014) involving cognition and motor functions.
Considerable works also suggest topologically organized communication between the cortex, the
basal ganglia, and the cerebellum (Bostan and Strick 2018; Hoshi et al. 2005). In this section, we
will review the neural representations of motor adaptation and motor sequence learning, then

15
combine these to a theoretical model of motor skill learning proposed by Doyon and Benali
(Doyon and Benali 2005).
The representation of motor sequence learning spans many brain regions such as
prefrontal and motor cortex, the basal ganglia, and the cerebellum. Especially, during the early
stage of motor sequence learning, the caudate nucleus, the dorsolateral prefrontal cortex
(DLPFC) (Doyon et al. 2002, 2009; Hikosaka et al. 2002; Lehéricy et al. 2005; Wu et al. 2015a),
and posterior cerebellar cortices (Floyer-Lea and Matthews 2004) are activated. However, during
the late stage of motor sequence learning, sensorimotor regions of the brain are more activated.
These regions include the posterior putamen (Wei et al. 2014), the primary motor cortex (M1),
the supplementary motor area (Wu et al. 2015; Lehéricy et al. 2005), and anterior cerebellar
cortex including deep cerebellar nuclei (Albouy et al. 2013; Floyer-Lea and Matthews 2004;
Sami et al. 2014). When this sequence skill is over-practiced and well-automatized, areas
involved in the sensorimotor basal ganglia circuit are more activated than cerebellum (Haslinger
et al. 2004; Tzvi et al. 2014; Wu et al. 2004).
Although motor adaptation is commonly known for a cerebellum-based process (Cullen
and Brooks 2015; Shadmehr et al. 2010), recent findings suggest that other brain regions such as
the basal ganglia (Seidler et al. 2006) or the cerebral cortex (Overduin et al. 2009; Seidler et al.
2006; Wise et al. 1998) contribute to different phases of motor adaptation. Ample studies in
people with cerebellar damage have found deficits in adapting to visuomotor perturbations (Rabe
et al. 2009; Schlerf et al. 2013; Tseng et al. 2007) such as a split-belt treadmill (Morton and
Bastian 2006; Statton et al. 2018), a prism goggle (Hanajima et al. 2015; Weiner et al. 1983), or
a force-field (Smith and Shadmehr 2005). Other brain areas were also activated during motor
adaptation, but in a specific phase of motor adaptation. During early phase of motor adaptation,

16
prefrontal cortex such as the DLPFC (Cassady et al. 2018; Ruitenberg et al. 2018) and pre-
supplementary area (Krakauer et al. 2004) appear to be activated. The basal ganglia also appear
to play a role during the phase of rapid motor skill acquisition, rather than during long-term
retention (Cassady et al. 2018; Krakauer et al. 2004; Ruitenberg et al. 2018; Seidler et al. 2006).
During the later phase of motor adaptation, sensorimotor cortices show activations such as the
posterior parietal cortex (Della-Maggiore et al. 2004; Krakauer et al. 2004), the M1 (Della-
Maggiore and McIntosh 2005; Landi et al. 2011; Li et al. 2001) as well as the basal ganglia
(Seidler et al. 2006). The disruption of M1 using transcranial magnetic stimulation (TMS) also
reveal a phase-specific influence during the later phase of motor adaptation (Orban de Xivry et
al. 2011; Richardson et al. 2006). Overall, this suggests that during the early phase of motor
learning, there may be a global contribution from the cerebellum, the basal ganglia, and the
cortex to acquire the motor skills, whereas during the late phase of motor learning, the
cerebellum may play more important role in motor adaptation.
Combined together (Figure 2-2), distinct cortico-striato-cerebellar circuits are involved
during different phases of motor skill learning. During the early phase of motor learning, the
associative striatum, such as the caudate nucleus (Doyon et al. 2002, 2009; Floyer-Lea and
Matthews 2004; Hikosaka et al. 2002; Lehéricy et al. 2005; Wu et al. 2015a), posterior cerebellar
cortices (Floyer-Lea and Matthews 2004), and frontal associative cortical regions, such as
dorsolateral prefrontal cortex (DLPFC), are activated while individuals rapidly acquire motor
skills. During this phase, individuals utilize cognitive processes to learn what to learn (Marinelli
et al. 2017). As practice continues, sensorimotor striatum such as the putamen, deep cerebellar
nuclei, and sensorimotor and parietal cortical regions are activated more than the associative
regions (Doyon and Benali 2005). These neuroanatomical structures are associated with

17
sensorimotor integration and processes. As learning is consolidated, retained, and automatized,
the activated neural substrates differ in the two types of motor skill learning. The sensorimotor
striatum and sensorimotor cortices are significantly activated during motor sequence learning,
while the cerebellum and sensorimotor cortices are predominantly activated during motor
adaptation.

18

Figure 2-2. Theoretical framework of corcito-striato-cerebellar circuitry involved in different
phases of motor skill learning based on changes in brain activities. Both the basal ganglia and
cerebellum contribute to skill acquisition regardless of motor skill learning processes. During
early learning, neural substrates known for ‘associative’ area such as the caudate nucleus,
anterior putamen, posterior cerebellar cortices, and frontal associative cortices are activated.
During late learning, neural substrates known for ‘sensorimotor’ area such as the putamen,
deep cerebellar nuclei, and sensorimotor and parietal cortices are activated. When the skill is
consolidated and retained, areas of neural activations differ between the type of motor skill
learning. During motor sequence learning, the putamen and sensorimotor cortices are
activated. Whereas, during motor adaptation, the cerebellum and sensorimotor cortices are
activated.

1.2. Obstacle Negotiation as a Locomotor Skill
Negotiating an obstacle can be compartmentalized into adjusting step lengths to be longer
or shorter a few steps before stepping over an obstacle (Chien et al. 2018; Chou and Draganich
1998a; Crosbie and Ko 2000), lifting an appropriate height of the foot (Amir A Mohagheghi

19
2004; Chou and Draganich 1998b), then placing the leading foot after the obstacle based on the
location of the trailing leg (Chen et al. 1991). An individual would execute these components in
a predetermined order to successfully step over an obstacle. This precise control of foot
placement and foot height before, during, and after the crossing (Marigold et al. 2011; Moraes
and Patla 2006) are executed while individuals adapt to environmental constraints such as the
different size of obstacles or walking speed (Maidan et al. 2018; da Silva et al. 2011). As visual
information provides one of the strongest influences on motor performance, obstacle negotiation
also depends on the integration of visual information about the body and environment with
ongoing motor commands.
1.2.1. Visual information during obstacle negotiation
When visual information about either the body and environment is missing such as
dimming light or carrying a large object, this can lead to an insufficient integration of necessary
information to successfully negotiating obstacles. The effects of insufficient visual information
on obstacle negotiation performance may result from a lack of spatial information about the
environment (exteroceptive information) and/or a lack of information about the body’s state
relative to the environment (exproprioceptive information). The role of visual information during
obstacle negotiation has commonly been studied by occluding portions of the visual field during
the approach to an impending obstacle. By occluding available visual information during the
penultimate step prior to an obstacle, several studies have found that individuals decrease
precision by making more collisions (Patla and Greig 2006; Rietdyk and Rhea 2011) and
increasing foot placement variability (Patla and Greig 2006) compared to trials with full vision
(Matthis and Fajen 2014a; Patla 1998). People also attempt to increase safety margins by lifting
their legs higher (Amir A Mohagheghi 2004; Graci et al. 2010; Rhea and Rietdyk 2007; Rietdyk

20
and Rhea 2006a; Timmis and Buckley 2012) and placing their leading and trailing feet further
away from the obstacles (Amir A Mohagheghi 2004; Rhea and Rietdyk 2007; Rietdyk and Rhea
2006a; Timmis and Buckley 2012). Although previous studies provide clues about how visual
information is used to guide obstacle negotiation, these studies often remove visual information
about the environment and the body simultaneously. As a result, it remains to be seen how visual
information about the body is integrated with information about the environment to facilitate
skillful obstacle negotiation.
1.3. Motor Skill Learning in Virtual Reality
In recent years, virtual reality (VR) has been increasingly used to provide engaging,
interactive, and task-specific locomotor training (Fung et al. 2006; Jaffe et al. 2004; Mirelman et
al. 2011, 2016; Parijat et al. 2015; Rizzo and Kim 2005; Shema et al. 2014; Yang et al. 2008).
VR can also be a platform to easily manipulate motor tasks or environments for training that may
not be possible or safe in the real world (Canning et al. 2020). VR-based locomotor training
frequently includes obstacle negotiation because it is an essential locomotor skill in the
community (Mirelman et al. 2011; Shema et al. 2014; Yang et al. 2008), and tripping over
obstacles is a common cause of falls in many patient populations (Stolze et al. 2004). The
efficacy of VR in a clinical setting is based on whether the learned motor skill is retained and
transferred to the real world.
The clinical application of VR-based training interventions is predicated on the idea that
practice in VR will lead to lasting changes in trained skills and that these changes will influence
real-world behavior. The presence of lasting changes in motor skills resulting from practice is a
hallmark of motor learning. Retention of motor skills has been examined in response to VR
training, particularly in fields such as flight and medical procedural training. For example,

21
complex surgical and medical skills are performed faster and more accurately during a retention
session following a single day of VR-based training (Ghanbarzadeh et al. 2014; Maagaard et al.
2011; Siu et al. 2016; Vaughan et al. 2016). Although healthy adults show improvement and
retention of skills related to operating a complex system using upper extremity, it remains to be
determined if individuals can retain locomotor skills during obstacle negotiation.
Skill transfer, which is defined as “the gain or loss in the capability for performance in
one task as a result of practice or experience on some other task” (Schmidt and Lee 2011), is
another key feature of motor learning. Skill transfer is particularly critical when skill acquisition
occurs in a context that differs from the environment in which the skill is expressed, such as VR-
based training. Following VR-based training of obstacle negotiation on a treadmill, participants
showed increased walking speeds in the lab (Mirelman et al. 2011; Shema et al. 2014) and
community (Yang et al. 2008). However, the evaluation of transfer in these VR-based training
studies was based on outcome measures such as walking speed that did not reflect the objective
of the training task, which was the control of foot clearance obstacle negotiation.

1.4. Motor Skill Learning in Parkinson Disease
1.4.1. Challenges with Obstacle Negotiation
Parkinson Disease (PD) is a progressive neurodegenerative disorder resulting from a loss
of dopaminergic neurons in the substantia nigra pars compacta (Ehringer and Hornykiewicz
1998), affecting over 1 million individuals over age 65 in the United States and over 7 million
worldwide (Kalia and Lang 2015; Ross and Abbott 2014; Weintraub et al. 2008). The loss of
dopamine leads to clinical manifestation in individuals with PD, primarily motor functions such
as postural instability, impairments in gait, rigidity, resting tremor, and bradykinesia (Erro et al.

22
2013; Grabli et al. 2012; Jankovic 2008; Jankovic et al. 1990; Ma et al. 2015; Weintraub et al.
2008), but also a wide range of cognitive dysfunctions such as deficits in executive functions and
visuospatial processing (Chaudhuri et al. 2006; Davidsdottir et al. 2005; Weintraub et al. 2008).
These symptoms exacerbate as the disease progresses, and lead to declines in locomotion and
high risk of falls (Gray and Hildebrand 2000). Consequently, 50-68% of individuals with PD
experience at least one fall every year while walking, and they report obstacle negotiation as one
of the major barriers of community participation (Ashburn et al. 2001; Stack and Ashburn 1999).
A gold-standard treatment for individuals with PD is dopamine replacement therapy
(DRT) to alleviate many PD-related motor symptoms such as bradykinesia, rigidity, and resting
tremor. However, other motor symptoms such as gait impairments and postural instability, even
in the early stage of PD, do not respond well to DRT (Peterson and Horak 2016). Moreover,
DRT eventually becomes inadequate in the later course of PD, resulting in a progressive
deterioration in mobility and activities of daily living. Increasing evidence supports that task-
specific, goal-based motor skill training promotes neuroplasticity and reduces motor impairments
in individuals with PD (Fisher et al. 2013; Petzinger et al. 2010, 2013). Therefore,
neurorehabilitation focusing on gait and postural training is commonly accompanied by
medication management therapy.
Complex gait training such as negotiating an obstacle course is an effective intervention
for people with PD. Complex locomotion such as obstacle negotiation requires an appropriate
modification of gait when approaching and stepping over an obstacle based on visuospatial
information about the current state of the body and the surrounding environments (Matthis and
Fajen 2014; Patla and Greig 2006; Rietdyk and Rhea 2011). However, PD impairs the ability to
appropriately modify their gait during the course of obstacle negotiation (Wu et al. 2015a),

23
which may lead to a high risk of falling. Therefore, obstacle negotiation is effective training to
challenge both motor and cognitive abilities for individuals with PD.
1.4.2. Motor and Locomotor Skill Learning
PD results in a progressive loss of dopaminergic pathways from the posterior to the
anterior part of basal ganglia, leading to deficits in motor learning in different phases across a
disease progression. During skill acquisition, people with mild to moderate PD consistently show
preserved the capability to acquire new motor skills. Individuals can acquire motor skills using
upper extremity such as bimanual coordination tasks (Swinnen et al. 2000; Verschueren et al.
1997), reaching tasks (Behrman et al. 2000; Ghilardi et al. 2003), a throwing task (Pendt et al.
2011), and finger sequence learning tasks (Lee et al. 2015; Onla-or and Winstein 2008). They
also demonstrate preserved motor skill acquisition using lower extremity such as an obstacle
negotiation skill (Michel et al. 2009) and the whole body such as a postural sequence learning
task (Smiley-Oyen et al. 2006) and maintaining balance tasks (Peterson et al. 2016; Van
Ooteghem et al. 2017). However, some studies found that the learning rate is slower than healthy
controls (Ghilardi et al. 2003; Michel et al. 2009; Smiley-Oyen et al. 2006). These findings
indicate that people mild to moderate PD do not abolish initial skill acquisition, but PD may slow
down the skill acquisition process.
Retention of motor skills in people with mild to moderate PD depends on attentional and
cognitive engagement. People with PD showed retention with a sequential reaching task
(Behrman et al. 2000), tracing predetermined movement trajectory (Onla-or and Winstein 2008),
throwing (Pendt et al. 2011), a postural sequence learning task (Smiley-Oyen et al. 2006), and
maintaining balance (Peterson et al. 2016; Van Ooteghem et al. 2017). On the other hand, in a
reaching task where external perturbation was large, people with PD showed slower and reduced

24
motor adaptation compared to healthy controls (Contreras-Vidal and Stelmach 1995). Moreover,
training with commercialized exercise games showed limited retention in people with PD,
possibly due to the high level of cognitive engagement with those games (Mendes et al. 2012).
1.4.3. Context-Dependent Learning
People with PD also have difficulty executing acquired motor skills in different contexts.
More strikingly, the difficulty executing the skill also occurs when seemingly irrelevant contexts
(i.e., incidental context) such as a background color of the scene are different (Lee et al. 2015).
This challenge is commonly referred to as context-dependent learning (CDL). CDL is defined as
superior retention of learned skills when retention is performed in the same context as the
training performance was compared to different contexts (Wright and Shea 1991). Over-reliance
on irrelevant contexts may limit the flexible generalization of learned skills. Particularly in
rehabilitation, CDL can be a major obstacle where learned motor skills in a clinical setting are
not well-generalized in community environments. Previous studies have demonstrated that
people with PD exhibit CDL during activities of daily living (ADLs) and finger sequence tapping
tasks (Lee et al. 2015; Nieuwboer et al. 2001; Onla-or and Winstein 2008). Nieuwboer and
colleagues revealed that home physical therapy treatment for patients with PD improves ADLs,
such as chair transfer or bed mobility. However, they performed better when the skills were
tested at home, which was the same context as the training context, compared to a hospital,
which was a different context (Nieuwboer et al. 2001). Moreover, Lee and colleagues also
demonstrated that people with PD could learn a laboratory-based finger sequence tapping task,
but their retention performance deteriorated when incidental context (i.e., background color and
location of sequence appearance) during retention differed from practice (Lee et al. 2015).

25
The potential neural mechanism of CDL is the deficit in context filtering in the striatum
by the lack of dopamine release, leading to encoding inappropriate contexts in the DLPFC
through the frontostriatal circuit (Dominey and Boussaoud 1997; Lee et al. 2016; Wise et al.
1996). Arguably, the DLPFC encodes all contextual information regarding the task, then the
striatum selects task-relevant contextual information and behaves as a filter to keep out
incidental information to select optimal motor command (D’Ardenne et al. 2012; Dominey and
Boussaoud 1997; Frank et al. 2001; Marcos et al. 2018; Wise et al. 1996). Lee and Fisher
demonstrated that down-regulation on the DLPFC using low frequency repetitive transcranial
magnetic stimulation (rTMS) reduces CDL in people with PD and healthy older adults, arguing
that the potential neural mechanism of CDL is the over-reliance on the frontostriatal circuit
during motor learning (Lee and Fisher 2017). The greater reduction of CDL by rTMS on the
DLPFC in people with PD compared to the healthy adults further suggests that the over-reliance
of the DLPFC may be due to the compensatory response for the impaired striatum. Therefore, it
may be possible that the excessive involvement of cognition during locomotor learning also
leads to context-dependency. However, it remains to be determined whether people with PD
exhibit CDL during locomotor skill tasks in a similar manner as upper extremity motor skills.

1.5. Explaining Individual Differences of Motor Learning in People with Parkinson
Disease
Effective motor learning evolves from an interaction between the task and learner’s
characteristics in addition to a specific practice condition. However, people with PD present a
wide range of variability in their clinical manifestation and disease progression, possibly caused
by varying pathophysiological and/or genetic expressions (Greenland et al. 2019). This

26
heterogeneity of PD can create difficulty in developing an intervention program and establishing
a disease prognosis.
One way to explain the individual difference in motor learning is through resting-state
functional connectivity (rsFC). Particularly, resting-state functional magnetic resonance imaging
(rsfMRI) can offer a different than the task-based fMRI, yet complementary perspective on how
the brain functional organization is associated with motor learning. Recent studies suggest that
intrinsic functional organization of the brain, which can be measured using rsFC, may reflect a
readiness to perform or learn future motor skills. rsfMRI measures large-scale covariance of low-
frequency spontaneous fluctuations in the blood oxygen level-dependent (BOLD) signal during
rest (van den Heuvel and Hulshoff Pol 2010). The strength of the correlation between two or
more remote brain regions reflects the degree of functional connectivity (FC), which indicates
temporal synchrony of those brain regions (van den Heuvel and Hulshoff Pol 2010). The
temporal synchrony between two or more brain regions at rest is referred to as resting-state
functional connectivity (rsFC) (van den Heuvel and Hulshoff Pol 2010; Smith et al. 2009).
1.5.1. Altered Resting-state Functional Connectivity in People with Parkinson
Disease
Ample studies have demonstrated that PD alters intrinsic functional connectivity related
to corticostriatal rsFC (Anderkova et al. 2017; Hacker et al. 2012; Helmich et al. 2010; Kim et al.
2017c; Luo et al. 2014; Nieuwhof and Helmich 2017; Owens-Walton et al. 2018; Yang et al.
2013). People with PD who are drug-naïve or in the early phase during OFF medication show
that FC between the posterior striatum and the sensorimotor cortex decreases (Anderkova et al.
2017; Ham et al. 2015; Helmich et al. 2010; Luo et al. 2014; Tahmasian et al. 2015; Wu et al.
2011; Yang et al. 2013). A whole-brain analysis also revealed disturbance of the balance in

27
functional brain networks in people with PD compared to healthy controls (Berman et al. 2016;
Koshimori et al. 2016; Luo et al. 2015; Sang et al. 2015; de Schipper et al. 2018; Tinaz et al.
2017; Wei et al. 2014; Yu et al. 2013; Zhang et al. 2015). Specifically, global functional
connectivity is reduced in the SMA, prefrontal cortex, and putamen—i.e., sensorimotor
corticostriatal circuit—in people with PD without DRT. The topologically different part of the
putamen as a seed demonstrated an opposite result than the posterior putamen. Helmich and
colleagues found a stronger FC between the anterior putamen and sensorimotor cortex (Helmich
et al. 2010), possibly as a compensatory remapping of striatum functions without DRT. This
suggests the importance of selecting the topologically segregated striatum to precisely
understand how the functional connectivity alters as a result of PD.
Pharmacological management for dopamine depletion partly normalizes the altered rsFC
in people with PD to a similar level of age-matched adults. The decrease in FC between the
posterior striatum and the sensorimotor cortex in people with PD without DRT is normalized or
further increased after DRT (Hacker et al. 2012; Tahmasian et al. 2015; Yang et al. 2013).
However, a whole-brain analysis demonstrated that people with PD with DRT showed that the
whole-brain rsFC stays in an ‘integrated’ brain state, rather than a ‘segregated’ brain state, longer
than healthy older adults (Kim et al. 2017c; Nieuwhof and Helmich 2017). This indicates that
people with PD may utilize less distinctive functional networks in the longer integrated brain
state for specific functional tasks. This disturbance in the balance between integrated and
segregated networks in people with PD may imply a lack of readiness to perform a task or
movement.
1.5.2. Associations between Resting-state Functional Connectivity and motor skill
acquisition and learning

28
As the corticostriatal circuits serve an important link to motor skill learning, the corticostriatal
circuits are good candidates to explain the individual differences in motor skill learning in people
with PD. In healthy young adults, baseline rsFC is associated with future motor skill acquisition
and learning (Bonzano et al. 2015; Faiman et al. 2018; Hamann et al. 2014; Mary et al. 2017;
Stillman et al. 2013; Wu et al. 2014, 2017). For instance, rsFC between the primary motor cortex
(M1) and the posterior putamen was negatively correlated with learning rate, reflecting initial
fast learning, during motor sequence learning. This indicates that faster learning rate can be
explained by the weaker putamen-motor cortex rsFC (Mary et al. 2017). Based on the motor
learning framework in the corticostriatal circuits, the sensorimotor circuit (putamen-sensorimotor
cortex) contributes to the later slow learning and retention. Whereas, the associative circuit
(caudate-associative cortex) contributes to the initial fast learning. The negative correlation
between fast learning and sensorimotor corticostriatal rsFC may suggest that specific intrinsic
functional connectivity can explain a different phase of motor learning. However, this study used
the M1 as their seed and did not specifically investigate corticostriatal rsFC in association with
motor skill learning. Therefore, it remains unclear whether distinct corticostriatal circuits can
explain different phases of motor learning. Given the clinical and neural heterogeneity of PD,
understandings of the rsFC may offer an additional explanation for the inter-individual
differences in motor skill learning in people with PD.

29
CHAPTER 3
The quality of visual information about the lower extremities influences
visuomotor coordination during virtual obstacle negotiation
This work was published in the Journal of Neurophysiology in 2018.

Kim A, Kretch KS, Zhou Z, Finley JM. The quality of visual information about the lower extremities
influences visuomotor coordination during virtual obstacle negotiation. J Neurophysiol. 2018;120(2):839-
847. doi:10.1152/jn.00931.2017

Abstract
Successful negotiation of obstacles during walking relies on the integration of visual
information about the environment with ongoing locomotor commands. When information about
the body and environment are removed through occlusion of the lower visual field, individuals
increase downward head pitch angle, reduce foot placement precision, and increase safety
margins during crossing. However, whether these effects are mediated by loss of visual
information about the lower extremities, the obstacle, or both remains to be seen. Here, we used
a fully immersive, virtual obstacle negotiation task to investigate how visual information about
the lower extremities is integrated with information about the environment to facilitate skillful
obstacle negotiation. Participants stepped over virtual obstacles while walking on a treadmill
with one of three types of visual feedback about the lower extremities: no feedback, end-point
feedback, or a link-segment model. We found that absence of visual information about the lower
extremities led to an increase in the variability of leading foot placement after crossing. The
presence of a visual representation of the lower extremities promoted greater downward head
pitch angle during the approach to and subsequent crossing of an obstacle. In addition, having
greater downward head pitch was associated with closer placement of the trailing foot to the
obstacle, further placement of the leading foot after the obstacle, and higher trailing foot
clearance. These results demonstrate that the fidelity of visual information about the lower

30
extremities influences both feed-forward and feedback aspects of visuomotor coordination
during obstacle negotiation.

Introduction
Locomotor skills such as obstacle negotiation depend on the integration of visual
information about the body and environment with ongoing motor commands. Although people
typically focus on the travel path during approach to impending obstacles (Patla and Vickers,
1997), the amount of time spent fixating an obstacle varies with obstacle height (Patla and
Vickers, 1997). This suggests that there is an important relationship between visual information
about impending obstacles and preparation for obstacle crossing. The role of visual information
during obstacle negotiation has commonly been studied by occluding portions of the visual field
during the approach to an impending obstacle. By occluding available visual information during
the penultimate step prior to an obstacle, several studies have found that individuals decrease
precision by making more collisions (Patla and Greig 2006; Rietdyk and Rhea 2011) and
increasing foot placement variability (Patla and Greig 2006) compared to trials with full vision
(Matthis et al. 2015; Matthis and Fajen 2014b; Patla 1998). People also attempt to increase safety
margins by lifting their legs higher (Amir A Mohagheghi 2004; Graci et al. 2010; Rhea and
Rietdyk 2007; Rietdyk and Rhea 2006b; Timmis and Buckley 2012) and placing their leading
and trailing feet further away from the obstacles (Amir A Mohagheghi 2004; Rhea and Rietdyk
2007; Rietdyk and Rhea 2006b; Timmis and Buckley 2012).
The effects of restricting visual information on obstacle negotiation performance may
result from a lack of spatial information about the environment (exteroceptive information)
and/or a lack of information about the body’s state relative to the environment (exproprioceptive

31
information). Foot placement is planned in a feed-forward manner approximately 2-2.5 steps
prior to an obstacle or target based on exteroceptive information about the obstacle’s spatial
location (Matthis et al. 2015; Matthis and Fajen 2014b; Timmis and Buckley 2012). This
suggests that restricting the view of the environment more than 2.5 steps prior to the obstacle
may impair planning and result in a higher collision rate. In addition, leading foot clearance is
fine-tuned in a feedback manner during obstacle crossing using exproprioceptive information
about the leading leg’s location relative to the obstacle (Patla 1998; Rhea and Rietdyk 2007;
Rietdyk and Rhea 2006b). In this case, loss of visual information about the position of the lower
extremities relative to an impending obstacle would likely impair the online control of the
leading limb. Although previous studies provide clues about how visual information is used to
guide obstacle negotiation, these studies often remove visual information about the environment
and the body simultaneously. As a result, it remains to be seen how visual information about the
body is integrated with information about the environment to facilitate skillful obstacle
negotiation.
Obtaining useful visual information about the environment requires coordination of gaze
behavior with ongoing locomotor commands. Visual input is dependent on how we orient our
head and where we fixate our eyes while we walk. Although directly fixating a target or obstacle
provides high-resolution spatial information, peripheral vision is often sufficient for guiding
obstacle negotiation (Franchak and Adolph 2010; Marigold et al. 2007; Timmis et al. 2016). The
human visual field is approximately 135° in the vertical direction (Harrington 1981), and, as a
result, we maintain visual access to the ground approximately 75 cm in front of our feet when
standing (Franchak and Adolph 2010) if the head and eyes are in a neutral position. During
locomotion over uneven terrain, individuals rotate their heads down slightly, approximately 16°

32
from neutral (Marigold and Patla 2008) which brings the lower edge of the field of view
approximately 22 cm from the body. This suggests that the leading limb may be within the field
of view during walking and this information could, therefore, be used for online control during
obstacle negotiation.
To better understand how visual information about the body is utilized during
locomotion, a number of studies have used immersive virtual reality (VR) with head-mounted
displays (HMDs) to investigate how behaviors such as distance estimation and affordance
perception are modified in the presence of a visual avatar (Bodenheimer et al. 2007;
Bodenheimer and Fu 2015; Leyrer et al. 2011, 2015; Lin 2014; Lin et al. 2015; McManus et al.
2011; Mohler et al. 2008, 2010; Phillips et al. 2010; Renner et al. 2013; Ries et al. 2009;
Thompson et al. 2004). Distance estimation in virtual environments becomes more accurate with
a full-body avatar compared to having no avatar when assessed verbally (Leyrer et al. 2011) or
by having participants walk a specified distance (Mohler et al. 2008, 2010; Phillips et al. 2010;
Ries et al. 2009). Moreover, people’s estimate of their ability to step over an object, step down a
ledge or duck below an object most closely matches real-world ability when a first-person avatar
is provided (Bodenheimer and Fu 2015; Lin et al. 2015). Given the importance of visual
information about the body for improving the perception of distance and estimating performance
capability, it is plausible that this information is also used to facilitate obstacle negotiation.
Here, we investigated how visual information about the lower extremities influences the
coordination between head pitch angle and obstacle negotiation. We used a virtual obstacle
negotiation task because this allowed us to have fine experimental control over the fidelity of
visual feedback provided to the user. We tested virtual obstacle negotiation performance with
three types of visual feedback about the lower extremities: no visual feedback, information about

33
the feet only, and rendering of the feet and legs using a link-segment model. We hypothesized
that (1) provision of visual information about the body will lead to reduced variability of foot
placement during obstacle negotiation, (2) individuals will increase downward head pitch angle,
a proxy of gaze, when approaching obstacles and that this effect would be heightened by the
presence of visual information about the lower extremities (3) greater downward head pitch
angle will be associated with lower foot clearance due to higher certainty about the relative
positions of the obstacle and foot and (4) the relationship between head pitch angle and foot
clearance will be strongest in the presence of visual information about the lower extremities. The
results of this study highlight how visual information about the lower extremities is integrated
with ongoing locomotor behavior to achieve successful obstacle negotiation in VR. Moreover,
understanding how visual information about the lower extremities affects obstacle negotiation
may help designers of VR-based clinical interventions determine the necessary fidelity of lower
body feedback required for users to achieve natural obstacle crossing performance in immersive
VR.
Methods
Participants
Eighteen healthy young individuals participated in this study (10M, 8F, average age and
standard deviation of 26±4 years). All participants had normal vision or corrected-to-normal
vision. Study procedures were approved by the Institutional Review Board at the University of
Southern California, and all participants provided written, informed consent before testing began.
Experimental Setup
Participants’ lower extremity kinematics were tracked used infrared-emitting markers
(Qualisys) placed on the following landmarks: toe (approximately the 4
th
metatarsal head), heel

34
(back of a shoe), knee (lateral epicondyle) and hip (greater trochanter). An additional marker was
placed on the HMD (Oculus Rift Development Kit 2, Oculus VR LLC) to measure approximate
eye height. Marker trajectories were used to control the virtual leg models, and these trajectories
were recorded for the duration of the experiment.
Virtual Obstacle Negotiation Task
The virtual environment was developed using Sketchup (Trimble Navigation Limited,
USA), and the participants’ interaction with the environment was controlled using Vizard
(WorldViz, USA). The virtual environment consisted of a corridor with obstacles spanning its
width (Figure 3-Figure 2-A). A total of 40 virtual obstacles were placed along the corridor at
random intervals of between 5m and 10m. The obstacles had a height and depth of 0.14m and
0.10m, respectively. The virtual environment was displayed within the HMD which had a 100
degree horizontal and vertical field of view, a resolution of 960X1080 pixels for each eye, a mass
of approximately 450g and 100% binocular overlap.
Participants viewed the scene through the HMD while walking on a treadmill (Bertec
Fully Instrumented Treadmill, USA). The velocity of the scene was synchronized with the
treadmill at 1.0 m/s, and the orientation of the viewpoint was controlled by an IMU within the
HMD. Participants began by walking on the treadmill while wearing the HMD for a period of
one minute to become familiar with the setup (Figure 3-2A). During this familiarization trial, no
visual information about the lower extremities was provided, and no virtual obstacles were
placed in the corridor. Then, participants performed three obstacle negotiation trials and three
obstacle-free trials. For each type of trial, three feedback conditions were included: (1) no body
model, (2) an endpoint foot model, and (3) a link segment leg model (Figure 3-B). The endpoint
model was used to determine if visual information about the endpoint of the limb improved

35
obstacle negotiation performance. The link segment model was included as previous studies have
observed that the perceived size of the body influences the perceived size of objects in the
environment (van der Hoort et al. 2011; Hoort and Ehrsson 2016). Hence, the addition of
segmental length information could encourage more consistent obstacle negotiation performance.
The locations of the models’ joints were determined by marker positions, and thus the models
were naturally calibrated to each participant’s anthropometry. Participants viewed the
environment and the lower extremity model from a first-person perspective (Figure 3-A).
During the obstacle negotiation trials,
participants were instructed to avoid collisions with
obstacles and maximize their score. The virtual
obstacles had collision sensors capable of detecting
when any of the markers entered the obstacle volume.
Participants began with a score of 40, and one point was
deducted for each collision. In addition, the obstacle
changed color from brown to blue at the onset of a
collision. During the three obstacle-free trials,
participants were instructed to walk down the hallway.
By examining behavior during trials with and without
obstacles, we could determine how the presence of
obstacles influenced head pitch angle independent from
the effects of lower extremity visual feedback. The
order of the feedback conditions was randomized, and

Figure 3-1. A) Virtual corridor with
obstacles. Participants were able to
see their score and one point was
deducted for each collision. B)
Visual feedback conditions from a
third person viewpoint. Spheres
represent the position of markers
placed on the lower extremities.
Segments connecting the spheres
were used to provide a visual
representation of limb segment
length.

36
the order of the obstacle and obstacle-free trials was counterbalanced across participants (Figure
3-2A).
Data Recording and Analysis
Toe, heel, and head marker positions were recorded at 100 Hz, and the data were
processed in MATLAB R2016b (Natick, MA). We first assessed the effects of lower extremity
feedback condition on a set of metrics that characterized obstacle crossing performance. The
following dependent variables were calculated to measure obstacle negotiation performance: (1)
collision rate, (2) average and standard deviation of foot clearance (the minimum vertical
distance between the toe and the obstacle), (3) average and standard deviation of foot placement
before the obstacle for the trailing limb (horizontal distance between the toe and rear edge of the
obstacle) and (4) average and standard deviation of foot placement after crossing for the leading
limb (horizontal distance between leading foot position and front edge of the obstacle). Each
metric of distance was normalized by each individual’s leg length. The participants’ average leg
length (greater trochanter to the floor) was 0.87 ± 0.06m.
Only successful trials were used for foot clearance and foot placement analyses. Eight out
of 2160 obstacles were omitted from our analyses because the software generated false
collisions. We also had false negatives in 282/2160 cases for the leading limb and 215/2160
cases for the trailing limb, meaning that there was an actual collision that was recorded as a
success. In these cases, we retained the data for our analyses, but we changed the outcome status
of each case to a collision. False positives and false negatives were identified in our data
processing when either the value of foot placement or clearance did not match with the outcome
status. For instance, if an obstacle with a negative value of foot clearance was identified as a
success, we checked the raw data by plotting the marker trajectory. If the marker trajectory truly

37
showed a mismatched outcome status, we changed the outcome status as a collision. These
mismatch likely resulted from brief lapses in network communication between Qualisys Track
Manager and Vizard.
We used head pitch angle, measured from the IMU embedded in the HMD, as an indirect
measure of participants’ strategy for acquiring visual information while walking. We quantified
the average head pitch angle for the four steps leading up to each obstacle limb (n-4, n-3, n-2,
and n-1, respectively), the crossing step with the leading limb (n), and the crossing step with the
trailing limb (n+1, Figure 3-2B). Averages were computed over the duration of each step (from
heel strike on one limb to heel strike on the opposite limb). A head pitch angle of 0°
corresponded to having the head in a neutral position, while negative values represented rotation
toward the ground (forward pitch). We evaluated how the presence of obstacles influenced
acquisition of visual information by comparing head pitch angle during the obstacle crossing
trials with the behavior averaged during the respective obstacle-free trials.
Statistical analysis
Statistical analyses of all dependent variables were performed using linear mixed-effects
(LME) models in R (R Project for Statistical Computing). We tested whether the type of visual
information provided about the lower extremities affected the following dependent variables: (1)
collision rate, (2) leading and trailing foot placement, (3) leading and trailing foot clearance, (4)
variability of leading and trailing foot placement, (5) variability of leading and trailing foot
clearance. We also tested for effects of visual feedback and step number (n-4, n-3, n-2, n-1, n,
n+1) on head pitch angle. For all models, we included a random intercept of each participant
because are it is typically recommended to include a random intercept for each participant in
longitudinal studies to account for differences between participants (Seltman 2012). We tested

38
each model to determine whether random slopes for each explanatory variable were necessary
using a log-likelihood test.
Foot clearance and placement of the leading and trailing limbs were tested for
associations with head pitch angle one-step before crossing to determine how the acquisition of
visual information about the environment influenced subsequent control of foot trajectory. Our
interests were in the main effect of the head pitch angle on foot clearance/placement variables
and the interaction between head pitch angle and visual feedback conditions. These models
required a random slope for feedback condition in addition to the random intercept for each
participant. For all of our statistical analyses, significance was set at p < 0.05 level. Post-hoc
comparisons were performed for significant main effects or interactions using Bonferroni
corrections for multiple comparisons. We used the R package lme4 to fit the models, package
lmerTest to calculate model p values using Satterthwaite approximations for the degrees of
freedom for the LME analyses, and the package multcomp for multiple comparisons.
Satterthwaite approximations adjusted the degrees of freedom based on differences in variance
between conditions.

39

Figure 3-2. A) Experimental protocol. Participants were counterbalanced to begin the
experiment with either obstacle-free trials or obstacle negotiation trials. The box above each
trial represent three feedback conditions: 1) no model as represented by a horizontal line at the
floor level, 2) end-point feedback as represented by two dots and 3) link-segment feedback as
represented by dots at the foot, knee and hip and sticks connecting the dots. B) Schematic
figure showing dependent variables. The solid and dashed lines represent the leading and
trailing limbs during obstacle crossing, respectively. Each step number was defined as the time
period between two consecutive heel strikes (HS). C) Representative time series of head pitch
angle during an obstacle negotiation trial. The solid black line refers to the head angle. The
vertical grey lines refer to point along the corridor where the center of the obstacle was
located. D) Representative time series data for heel trajectory. The dashed grey line refers to
the left heel trajectory and the solid grey line refers to the right heel trajectory. Black boxes are
scaled to the height and width of the obstacles along the path.

40
Results
Foot placement and clearance
Visual information about the lower extremities had a systematic effect on obstacle
crossing behavior. A representative example of the foot trajectory during obstacle negotiation is
shown in Figure 3-2D. The results of the likelihood ratio tests revealed that all dependent
variables including collision rate, and foot placement and clearance metrics were better fit
without random slopes. We found a significant main effect of the type of visual information
provided about the lower extremities (no model, endpoint model, or link-segment model) on the
variability of leading foot placement (F(2,33)=4.30, p=0.02) (Figure 3-3). Specifically, the
variability of leading foot placement was significantly reduced when end-point (Bonferroni
corrected p = 0.03) and link-segment (Bonferroni corrected p = 0.04) feedback were presented
compared to no visual feedback. There were no significant differences in any other variables
related to obstacle crossing across conditions (Table 1).

Figure 3-3. Box and whisker plot illustrating
leading foot placement variability. Horizontal
lines within each box indicate median values,
and the bottom and top boundaries of the box
indicate the 25th and 75th percentiles. Points
outside each box indicate outliers. The asterisk
denotes a significant difference at the
Bonferroni corrected p < 0.05 level.

We also found that some foot placement metrics differed between success and collision
trials. Both the trailing foot before crossing and the leading foot after crossing were placed
further from the obstacle during collision trials compared to success trials (F(1,82)= 9.32,
p=0.003 for the trailing foot and F(1,82)= 7.02, p=0.01 for the leading foot). In addition, trailing
foot placement variability was higher in collision trials than in success trials (F(2,79)= 3.16,

41
p=0.048) and this difference was larger in the no visual feedback condition than in the end-point
only condition (Bonferroni corrected p = 0.036).
Table 3-1. Average and standard deviation of foot placement and clearance. Values are means
(M) ± standard deviation (SD).
Variables No model End-point Link-segment p-value
Trailing foot
placement (m)
M 0.28±0.14 0.28±0.14 0.29±0.13 0.96
SD 0.11±0.04 0.10±0.04 0.11±0.04 0.85
Leading foot
clearance (m)
M 0.15±0.12 0.15±0.12 0.15±0.11 0.86
SD 0.08±0.04 0.07±0.04 0.07±0.03 0.50
Leading foot
placement (m)
M 0.26±0.08 0.27±0.07 0.25±0.06 0.67
SD 0.06±0.03 0.04±0.02* 0.05±0.02* 0.02*
Trailing foot
clearance (m)
M 0.19±0.08 0.19±0.10 0.20±0.08 0.74
SD 0.06±0.03 0.06±0.03 0.06±0.02 0.89
* p<0.05 significant difference compared to the no feedback condition
Collision rate
Although the type of visual information provided about the lower extremities influenced
crossing behavior, it did not affect the overall collision rate or the frequency of each type of
collision. There was no significant effect of the type of visual information provided about the
lower extremities on collision rates (F(2,34)=1.01, p=0.38). The average number of collisions per
trial in the no model, end-point, and link-segment feedback conditions was 17±13, 14±13, and
13±12, respectively.
Changes in head pitch angle during obstacle negotiation
As participants approached and crossed impending obstacles, we observed a consistent
time course of changes in head pitch angle. During the approach steps, participants gradually
increased downward head pitch angle until they crossed the obstacle with the leading limb
(Figure 3-2C). After crossing, participants reduced the magnitude of downward head pitch angle
toward the level observed in the trials without obstacles. The results of the likelihood ratio test

42
revealed that the model of head pitch angle as a function of the type of visual information
provided about the lower extremities and step number required random slopes for each type of
visual information in addition to a random intercept for each participant. The model returned
significant main effects of the type of visual information provided about the lower extremities
(F(2,299)=14.84, p<0.001) and step number (F(5,299)=70.03, p<0.001) on the head pitch angle
(Figure 3-4). Post-hoc analyses revealed that the magnitude of the head pitch angle at step n-4
was significantly less than the magnitude at all other steps (all Bonferroni corrected p<0.05),
while the magnitude of the head pitch angle during the crossing step (step n) was significantly
greater than all other steps (all Bonferroni corrected p<0.05). On average, the magnitude of the
head pitch angle was 6±1º greater (Bonferroni corrected p<0.001) with the end-point model and
3±1º greater (Bonferroni corrected p=0.004) with the link-segment model compared to when no
visual feedback was provided. These changes in head pitch during approach and crossing were
the result of negotiating obstacles rather than the novelty of the lower extremity feedback. This
was supported by our findings that the average head pitch angle was between 0 and -5º while
walking with no obstacles and pitch angle did not differ across visual feedback conditions in the
no obstacle trials (F(2,26)=3.00, p=0.07). Lastly, there was no difference in head pitch angle
during success versus collision trials.

43

Figure 3-4. Average head pitch angle as a function of
step number and feedback condition. Error bars
indicate the standard error. The black line refers to the
condition when no visual information about lower
extremities was provided, the dark grey line refers to
the condition when an end-point foot representation
was provided, and the light grey line refers to the
condition when the link-segment leg representation
was provided. ***: Bonferroni corrected p<0.005,
****: Bonferroni corrected p<0.001
Associations between head pitch angle before crossing and obstacle crossing performance
We next investigated whether head pitch angle during the approach to an obstacle was
associated with subsequent obstacle crossing performance. Multiple metrics of crossing behavior
at the individual obstacle level were predicted by head pitch angle during the approach step.
There was a main effect of head pitch angle during the approach step (n-2) on the trailing foot
placement before the obstacle (F(1,1482)=32.62, p<0.001, β = 2e-3 m/deg, standard error = 4e-4
m/deg) and head pitch angle during the final approach step (n-1) on leading foot placement after
the obstacle (F(1,2015)=82.01, p<0.001, Figure 3-5A). Specifically, greater downward head
pitch angle before crossing was associated with closer placement of the trailing foot to the
obstacle and farther placement of the leading foot beyond the obstacle after crossing. There was
also an interaction between the type of visual information provided about the lower extremities
and head pitch angle during the final approach step (n-1) on leading foot placement after the
obstacle (F(2,609)=4.92, p=0.008, Figure 3-5A). The correlation between head pitch angle and
leading foot placement was less strong with link-segment feedback compared to when no visual
feedback about the lower extremities was provided (Bonferroni corrected p=0.006).
Associations between head pitch angle during obstacle crossing and crossing strategy

44
Crossing performance at the individual obstacle level was also associated with head pitch
angle during obstacle crossing. There was a main effect of head pitch angle during leading limb
crossing (n) on leading foot placement (F(1,1900)=14.51, p<0.001, Figure 3-5B). Specifically,
greater downward head pitch angle during crossing was associated with farther placement of the
leading foot after crossing. Head pitch angle was also associated with measures of foot clearance.
There was a significant main effect of the head pitch angle during the crossing step of the leading
limb (n) on trailing foot clearance (F(1,1847)=41.06, p<0.001, β = -1e-3 m/deg, standard error =
2e-4 m/deg) such that greater downward head pitch during the previous step was associated with
higher trailing foot clearance. Moreover, there was a main effect of the head pitch angle during
trailing foot crossing (n+1) on trailing foot clearance (F(1,2017)=16.63, p<0.001). Greater
downward head pitch angle was associated with higher trailing foot clearance. Lastly, there was
an interaction between head pitch angle during trailing foot crossing (n+1) and the type of visual
information provided about the lower extremities on trailing foot clearance (F(2,696)=6.04,
p=0.003, Figure 3-5C). Specifically, the correlation between head pitch angle and concurrent
trailing foot clearance was significantly stronger with link-segment feedback compared to no
feedback (Bonferroni corrected p=0.002). Together, these results suggest that visual information
about the body and/or the obstacles was used for online control of obstacle negotiation.

45

Figure 3-5. Estimated regression coefficients for the dependent variables from linear models
describing the relationship between (A) leading foot placement after the obstacle and head
angle during the approach step, (B) leading foot placement after the obstacle and head angle
during lead foot crossing, and (C) trailing foot clearance and head angle during trailing foot
crossing in each feedback condition. Negative coefficients indicate that greater downward
head angle was associated with larger values for each crossing variable. The illustrations in the
left column represent the period analyzed for head angle and the corresponding crossing
variable. The solid black lines indicate the analyzed step for head angle (thin black arrows)
and the respective obstacle crossing performance variable (solid grey arrows). Error bars
represent standard errors. HS: Heel Strike. **: Bonferroni corrected p<0.01, ***: Bonferroni
corrected p<0.005.

Discussion
The objective of this study was to determine how the presence and acquisition of visual
information about one’s body influences obstacle negotiation performance in a virtual

46
environment. To this end, we used a virtual reality environment to determine how the provision
of different levels of visual information about the lower extremities influenced negotiation of
virtual obstacles while walking on a treadmill. We found that the variability of leading foot
placement after crossing was reduced when visual information about the foot was provided, and
downward head pitch angle was increased when either form of visual feedback was provided.
Further, at the individual obstacle level, trailing foot placement before the obstacle, leading limb
placement after the obstacle, and trailing foot clearance were associated with head pitch angle
prior to and during crossing. Moreover, the association between trailing foot clearance and head
pitch angle during crossing was strongest when the link-segment leg representation was
provided. Our results demonstrate that visual information about the lower extremities is
integrated with ongoing locomotor commands to modulate acquisition of visual information and
foot placement during obstacle negotiation.
Provision of visual information about the lower extremities facilitates precise foot placement
during obstacle negotiation
In line with our hypothesis, the precision of leading foot placement increased when visual
information about the lower extremities was provided. This agrees with previous research where
the variability of foot placement during real-world obstacle negotiation increased when the lower
visual field was occluded (Amir A Mohagheghi 2004; Rhea and Rietdyk 2007; Timmis and
Buckley 2012). This was also consistent with previous research which found when online visual
feedback of endpoint trajectory was provided along with proprioceptive feedback, participants
reduced movement variability (Franklin et al. 2007). Our finding that foot placement was also
more variable during collision trials in no visual feedback condition further supports our
conclusion that visual information was used to promote more consistent foot placement.
Together, these results show that online visual information about both the upper and lower

47
extremities is used to improve the precision of visually guided motor behaviors such as obstacle
negotiation.
Coordination between head pitch angle and obstacle crossing performance
During the approach to impending obstacles, individuals increased their downward head
pitch angle until they reached the obstacle, then gradually decreased the head pitch angle after
crossing an obstacle. The magnitude and timing of changes in head pitch angle differed from
what has previously been reported in the real world. In our task, participants reached a peak head
pitch angle of approximately 40~50°, while previous studies reported that head pitch angle is
approximately 16° while walking over uneven terrain (Marigold and Patla 2008). This
discrepancy likely results from the lack of peripheral vision in VR with an HMD. Individuals
may have compensated for the narrower vertical field of view in the HMD compared to the real
world by increasing their downward head pitch angle to obtain the visual information that would
naturally be acquired in the real world. The presence of this compensation suggests that
information in the peripheral visual field, including information about the environment and
information about the body, is utilized to perform obstacle negotiation.
When any visual information about the lower extremities was provided, participants
increased head pitch angle relative to the no model condition. Given the increase in precision of
leading foot placement when endpoint information was provided, this suggests that the visual
information about the lower extremities accessed by increasing head pitch angle was used to
guide foot placement. We also found that trailing limb clearance was most strongly coordinated
with head pitch angle during crossing when link segment information was provided. Thus, for a
given head pitch angle, participants had a greater safety margin for the trailing limb in the
presence, not absence, of visual information about the lower extremities. This finding is

48
inconsistent with previous studies which demonstrate that increased downward gaze and
increased safety margins are typically observed when the lower visual field is occluded (Amir A
Mohagheghi 2004; Marigold and Patla 2008; Rhea and Rietdyk 2007; Rietdyk and Rhea 2006b;
Timmis and Buckley 2012). However, a potential explanation for this discrepancy is that
participants in our study, particularly specific to fully immersive VR, used visual information
about the trailing limb to increase the probability of successful clearance due to the absence of a
visual representation of the trunk. Therefore, future work will need to determine how occlusion
of the trailing limb by the trunk influences visuomotor coordination during obstacle negotiation.
Greater safety margins may be also be a consequence of closer placement of the trailing foot to
the obstacle before crossing as this would require modification of the foot’s trajectory to
guarantee successful clearance.
An increase in head pitch angle was also associated with closer placement of the trailing
limb to the obstacle, further placement of the leading limb after the obstacle, and higher trailing
foot clearance. These associations were observed when head pitch was measured during the step
prior to crossing or during the crossing step, even in the absence of visual information about the
lower extremities. Acquiring more visual information about the obstacle may have led our
participants to reduce the safety margin associated with trailing foot placement due to increased
certainty about the body’s position relative to the obstacle. This would necessarily result in
further leading foot placement after the obstacle if participants maintained a consistent step
length during obstacle crossing. The increase in trailing foot clearance may have been necessary
to reduce the probability of collisions by the trailing foot given the closer placement of the
trailing foot to the obstacles. The observation of correlations between head pitch angle during the
step prior to crossing and metrics of obstacle negotiation may reflect the use of visual

49
information about the lower extremities in a feed-forward manner to control the trajectory of the
lower extremities.
Changes in correlations between obstacle negotiation performance metrics and head pitch
angle when a link-segment representation was provided further suggest that visual information
about the lower extremities is used during obstacle negotiation. Having a link-segment
representation reduced the correlation between leading foot placement after crossing and head
pitch angle prior to crossing compared to when no visual feedback about the lower extremities
was provided. This dissociation may indicate that in the absence of visual information about the
lower extremities, leading foot placement relies on feed-forward planning, but when visual
information about the lower extremities is available, leading foot placement is more reliant on
online control.
Performance metrics did not differ between success and collision trials. More
specifically, foot placement and head pitch angle variables revealed similar patterns in success
and collision trials. Although we observed changes in performance based on the amount of
information provided about the lower extremities, the use of this information does not appear to
differ during successful trails versus trials with collisions.
Clinical implications
Due to recent advances in VR technology, cost-effective and accessible VR systems have gained
an attention as a potential tool for rehabilitation (Arias et al. 2012; Gobron et al. 2015; Kim et al.
2017a; Malik et al. 2017). However, there has been little investigation of the features of the
virtual environment influence locomotor performance. Our results show that the types of visual
information provided about the lower extremities may contribute to different obstacle negotiation
strategies. These effects might be more prominent in populations with neurological impairments

50
where processing of multi-modal sensory information is damaged (van Hedel et al. 2005; Malik
et al. 2017; Michel et al. 2009). Therefore, careful consideration of how visual information about
the lower extremities is provided and utilized by people with neurological impairments during
interactions with virtual environment may have a significant influence on the potential utility of
VR in rehabilitation.
Limitations
There are a few primary limitations of this study. First, collision with obstacles was
indicated to participants only by binary visual feedback, which may have caused difficulty
recognizing the timing and side of the collision. This could be remedied in future studies through
the provision of specific haptic feedback that is coordinated with the onset of virtual obstacle
collision. Second, we did not measure gaze directly, but we used head pitch angle as a proxy of
gaze. Although acquiring visual input partially depends on the orientation of individuals’ head
while walking, a direct measure of gaze during obstacle negotiation would provide a more
accurate understanding of how we use visual information during obstacle negotiation. Third, we
did not use a realistic representation of the body (i.e. an avatar) and this may have affected the
observed obstacle negotiation strategy. Although provision of an end-point representation and a
sense of vertical scale via segmental information contributed to changes in obstacle negotiation
performance, further investigation is required to determine how the inclusion of a realistic avatar
would influence crossing behavior. Lastly, obstacle negotiation performance in VR may differ
from that in the real world and, as a result, future studies should characterize differences in
spatiotemporal coordination and gaze control during obstacle negotiation in real and virtual
environments.

51
Conclusion
This study aimed to investigate how visual information about one’s own body is utilized
during obstacle negotiation in a virtual environment. We presented three levels of visual
feedback about the lower extremities in VR and explored how this information influenced
participants’ obstacle negotiation strategy. Our results revealed that visual information about the
lower extremities promoted more consistent obstacle crossing behavior. Moreover, we observed
that individuals actively acquired visual information about the environment by increasing
downward head pitch angle during obstacle negotiation. Finally, visual information acquisition
strategies were associated with obstacle negotiation performance at the individual obstacle level,
and provision of visual information about the lower extremities impacted these associations. Our
findings suggest that visual information about the lower extremities is used in planning,
execution and online control during obstacle negotiation. These results support the need for
considering the type of visual information provided about the lower extremities when designing
locomotor tasks in VR.

52
CHAPTER 4
Locomotor skill acquisition in virtual reality shows sustained transfer to the real
world
This work was published in the Journal of NeuroEngineering and Rehabilitation in 2019.

Kim, A., Schweighofer, N. & Finley, J.M. Locomotor skill acquisition in virtual reality shows sustained
transfer to the real world. J NeuroEngineering Rehabil 16, 113 (2019). https://doi.org/10.1186/s12984-
019-0584-y

Abstract
Background: Virtual reality (VR) is a potentially promising tool for enhancing real-world
locomotion in individuals with mobility impairment through its ability to provide personalized
performance feedback and simulate real-world challenges. However, it is unknown whether
novel locomotor skills learned in VR show sustained transfer to the real world. Here, as an initial
step towards developing a VR-based clinical intervention, we study how young adults learn and
transfer a treadmill-based virtual obstacle negotiation skill to the real world.
Methods: On Day 1, participants crossed virtual obstacles while walking on a treadmill, with the
instruction to minimize foot clearance during obstacle crossing. Gradual changes in performance
during training were fit via non-linear mixed effect models. Immediate transfer was measured by
foot clearance during physical obstacle crossing while walking over-ground. Retention of the
obstacle negotiation skill in VR and retention of over-ground transfer were assessed after 24
hours.
Results: On Day 1, participants systematically reduced foot clearance throughout practice by an
average of 5 cm (SD 4 cm) and transferred 3 cm (SD 1 cm) of this reduction to over-ground
walking. The acquired reduction in foot clearance was also retained after 24 hours in VR and
over-ground. There was only a small, but significant 0.8 cm increase in foot clearance in VR and

53
no significant increase in clearance over-ground on Day 2. Moreover, individual differences in
final performance at the end of practice on Day 1 predicted retention both in VR and in the real
environment.
Conclusions: Overall, our results support the use of VR for locomotor training as skills learned
in a virtual environment readily transfer to real-world locomotion. Future work is needed to
determine if VR-based locomotor training leads to sustained transfer in clinical populations with
mobility impairments, such as individuals with Parkinson’s disease and stroke.

Background
In recent years, virtual reality (VR) has been increasingly used to provide engaging,
interactive, and task-specific locomotor training (Fung et al. 2006; Jaffe et al. 2004; Mirelman et
al. 2011, 2016; Parijat et al. 2015; Rizzo and Kim 2005; Shema et al. 2014; Yang et al. 2008).
These studies have simulated walking in different environments such as park or streets (Fung et
al. 2006; Yang et al. 2008), walking on a slope (Fung et al. 2006), or walking while avoiding
obstacles (Fung et al. 2006; Mirelman et al. 2011; Shema et al. 2014; Yang et al. 2008). VR-
based locomotor training frequently includes obstacle negotiation because it is an essential
locomotor skill in the community (Mirelman et al. 2011; Shema et al. 2014; Yang et al. 2008)
and tripping over obstacles is a common cause of falls in many patient populations (Stolze et al.
2004). The clinical application of VR-based training interventions is predicated on the idea that
practice in VR will lead to lasting changes in trained skills and that these changes will influence
real-world behavior. Therefore, understanding how locomotor skills acquired in VR are retained
and how these skills generalize to the real world is critical for determining the long-term utility
of VR for locomotor rehabilitation.

54
The presence of lasting changes in a motor skill as a result of practice is a hallmark of
motor learning and this retention process has been examined across a wide variety of real and
virtual learning contexts. Retention of motor skills has been examined in response to VR
training, particularly in fields such as flight and medical procedural training. For example,
complex surgical and medical skills are performed faster and more accurately during a retention
session following a single day of VR-based training (Ghanbarzadeh et al. 2014; Maagaard et al.
2011; Siu et al. 2016; Vaughan et al. 2016). Retention of locomotor skills is often explored in
studies that analyze how people adapt to external perturbations such as a split-belt treadmill
which has separate belts for the right and left legs (Day et al. 2018; Leech et al. 2018; Malone et
al. 2011), elastic force fields (Fortin et al. 2009), robotic exoskeletons (Gordon and Ferris 2007),
or added loads (Smith and Martin 2007). For instance, studies of the split-belt treadmill learning
process have revealed that the increases in step length asymmetry observed during initial
exposure to the belts moving at different speeds significantly decreased with subsequent
exposures to the device (Day et al. 2018; Leech et al. 2018; Malone et al. 2011). A recent study
by Krishnan and colleagues also investigated locomotor skill learning during a tracking task in
which participants were instructed to match the pre-defined target of hip and knee trajectories as
accurately as possible during the swing phase of the gait (Krishnan et al. 2018). They found that
the reduction in tracking error achieved through practice is retained the following day.
Although motor skill learning in VR and locomotor learning have been examined in isolation, it
remains to be seen how locomotor skills are acquired and retained following training in a virtual
environment.
Skill transfer, which is defined as “the gain or loss in the capability for performance in
one task as a result of practice or experience on some other task” (Schmidt and Lee 2011), is

55
another key feature of motor learning. Skill transfer is particularly critical when skill acquisition
occurs in a context that differs from the environment in which the skill is to be expressed. One
way in which skill transfer has been evaluated during motor learning is by measuring how the
adaptation of reaching in a robot-generated force field generalizes to unconstrained reaching.
This work has shown that adaptation to reaching in a curl-field leads to increased curvature
during reaching in free space (Cothros et al. 2006; Kluzik et al. 2008). Moreover, studies of
treadmill-based locomotor skill learning often evaluate transfer of learned skills from treadmill
walking to over-ground. For example, during split-belt treadmill adaptation, the learned changes
in interlimb symmetry partially transfer to over-ground walking (Torres-Oviedo and Bastian
2012). Further, VR-based training of obstacle negotiation on a treadmill led to increased walking
speeds in the lab (Mirelman et al. 2011; Shema et al. 2014) and community (Yang et al. 2008).
However, the evaluation of transfer in these VR-based training studies was based on outcome
measures such as walking speed that did not reflect the objective of the training task, which was
the control of foot clearance obstacle negotiation. Therefore, it remains to be seen if the elements
of skill from VR-training transfer to over-ground walking.
Underlying individual differences in learning can influence motor skill retention and
transfer to new environments. For example, a recent study demonstrated that healthy older adults
and people post-stroke who acquire a motor sequence skill at a faster rate also show greater
retention of that skill (Wadden et al. 2017). Similarly, the rate of skill acquisition for a reaching
task during early training predicts faster trial completion time at 1-month follow-up (Schaefer
and Duff 2015). Lastly, the magnitude of improvements in reaching speed during skill
acquisition predicts long-term changes in reaching speed in healthy individuals (Park and
Schweighofer 2017). Studies of individual differences in transfer have most often sought to

56
understand how the practice of a skill with one limb influences performance of the same skill
with the untrained limb. For example, interlimb transfer of motor skills acquired through
visuomotor adaptation varies with handedness (Chase and Seidler 2008) and individual
differences in baseline movement variability (Lefumat et al. 2015). However, far less work has
sought to understand how individual differences in skill acquisition affect the transfer of learned
skills to new environments. Overall, the influence of individual differences in skill acquisition on
locomotor skill retention and sustained transfer has yet to be determined.
Here, we determined how individual differences in locomotor skill learning during virtual
reality treadmill-based training influence retention and transfer of learned skills to over-ground
walking in the real world. We used a VR-based version of a previously established precision
obstacle negotiation task (Erni and Dietz 2001; van Hedel and Dietz 2004) and asked 1) whether
healthy young adults could learn to minimize clearance during virtual obstacle negotiation, 2) if
the learned skill transferred to over-ground walking, 3) if the learned skill was retained in both
VR and the real world after 24 hours, and 4) if individual differences in the amount or rate of
skill acquisition could predict retention and transfer. We hypothesized that 1) participants would
reduce foot clearance in VR during practice on Day 1 and that 2) the reduced foot clearance in
VR would transfer to over-ground obstacle negotiation. We also hypothesized that 3) the
reduction in foot clearance in VR and over-ground would be retained in each environment after a
24-hour retention period. Lastly, given that the rate and magnitude of the performance
improvement during skill have been established as predictors of skill retention in previous
studies, we also hypothesized that 4) these measures would predict retention of the learned skill
in VR and over-ground. Given the growing use of VR for motor skill learning, our results may

57
provide a unique opportunity to understand the factors that influence how training in VR might
lead to long-term improvements in skilled locomotion.
Methods
Participants
Nineteen healthy young adults participated in the study (10 female, average age of 26 ± 4
years). All participants had normal vision or corrected-to-normal vision. Study procedures were
approved by the Institutional Review Board at the University of Southern California and all
participants provided written, informed consent before testing began. All aspects of the study
conformed to the principles described in the Declaration of Helsinki.
Experimental Protocol

58
Participants completed a VR-based version of a previously established obstacle
negotiation task (Erni and Dietz 2001; van Hedel and Dietz 2004) where they were instructed to
minimize foot clearance when crossing the obstacle. This task was specifically chosen because it
allowed us to examine a form of skill acquisition, which requires participants to learn a precise

Figure 4-1. Experimental setup and protocol. (A) Virtual corridor with obstacles and an eye-
level display of participants' current score. (B) Visual feedback of the lower extremities
viewed from a third-person perspective. Spheres represent the position of markers placed on
the lower extremities. Line segments connecting the spheres were used to provide a visual
representation of limb segment length. During the study, participants viewed the
representation of the lower extremities from a first-person viewpoint. (C) Schematic diagram
of the mapping between the participant’s performance and the auditory feedback they
received. (D) Over-ground obstacle negotiation setup. (E) Experimental protocol illustrating
the day of the study, the trial type, number of obstacles per trial, and whether auditory
performance feedback was provided.

59
mapping between the perception of the spatial location of virtual obstacles and the control of foot
trajectory. Participants walked on a treadmill (Bertec Fully Instrumented Treadmill, USA) while
wearing a head-mounted display (HMD) and interacting with the virtual environment. The
velocity of the virtual environment was synchronized with the treadmill at 1.0 m/s, and an IMU
within the HMD controlled the orientation of the viewpoint. The virtual simulation was run at 60
Hz and the motion capture system had a real-time delay of 3.5 ms. All participants were
instructed to lightly hold on to the handrails while walking on the treadmill. Participants viewed
the environment (Figure 4-A) and the virtual representation of their legs from a first-person
perspective (Figure 4-B). Their body was represented by a set of spheres located at the hip, knee,
heel, and toe, bilaterally and lines connecting the spheres to represent the limb segments (Kim et
al. 2018). Marker placement details are described below in the Data Collection and Processing
section.
The virtual environment consisted of a corridor with obstacles (Figure 4-A) and was
developed using Sketchup (Trimble Navigation Limited, USA). The interaction between
participants and the virtual environment was controlled using Vizard (WorldViz, USA). A total
of 40 virtual obstacles, 20 each on the right and left side, were randomly placed along the
corridor at intervals of between 5 m and 10 m to provide sufficient space for participants to fully
recover their typical gait pattern after crossing the previous obstacle. Previous work has also
established that a distance of 2 m between obstacles is sufficient for people to cross each obstacle
as if it was independent of the others (Krell and Patla 2002). The height of the obstacles was
adapted from the previous study, which used the same objective of minimizing foot clearance,
but with a physical obstacle (Michel et al. 2009). The obstacles were 0.14 m in height and 0.10 m
in depth. The placement of the obstacles was lateralized so that participants either crossed the

60
obstacle with the right or left leg. In VR, the mediolateral distance between the feet was
constrained such that participants only saw movement of their legs in the sagittal plane. We
imposed this constraint to ensure bilateral obstacle negotiation. An Oculus Rift Development Kit
2 HMD was used to display the virtual environment. The HMD had a 100-degree horizontal and
vertical field of view, a resolution of 960 X 1080 pixels for each eye, a mass of approximately
450 g and 100 % binocular overlap.
The study consisted of two visits on consecutive days (Figure 4-E). On Day 1,
participants first walked over-ground while stepping over a single physical obstacle ten times
without the HMD (BASE, Figure 4-D). If there was a collision during physical obstacle
negotiation, participants repeated the trial. Then, they moved to the treadmill and donned the
HMD. For the first treadmill trial, they walked while stepping over 20 obstacles while receiving
auditory feedback in the form of an unpleasant sound following collisions with the obstacles
(BASE_VR). Following BASE_VR, participants had three practice blocks of obstacle
negotiation in VR and each block consisted of 40 obstacles. Between blocks 2 and 3 we
evaluated immediate transfer by having participants perform a physical obstacle negotiation task
ten times over-ground without the HMD (TF). In all trials, participants were instructed to cross
each obstacle using only the limb ipsilateral to the obstacle and to minimize the vertical distance
between the foot and the obstacles during the crossing. During the over-ground obstacle
negotiation task, we instructed participants to use either the right or left leg to step over the
obstacle prior to walking. The right-left order of the ten trials was randomized for each
participant. Moreover, for over-ground walking, participants were instructed to maintain their
walking speed throughout obstacle negotiation and avoid slowing down during the approaching
and crossing step. During the training blocks, participants received three types of auditory

61
performance feedback: 1) a pleasant sound when foot clearance was within a target range of 0-2
cm, which was used to be consistent with previous studies (Erni and Dietz 2001; van Hedel and
Dietz 2004), 2) an error sound whose frequency scaled with foot clearance when clearance was
greater than 2 cm, and 3) a failure sound following collisions with the obstacle (Figure 4-C).
Participants began each trial with 40 points and lost one point for each collision with the
obstacle. After 24 hours, participants returned to the laboratory and completed one retention
block of 40 obstacles in VR with no auditory feedback other than collision feedback (RET_VR).
We then assessed over-ground retention (RET_OG) in the same manner as BASE.
Data Collection
Participant’s lower extremity kinematics were tracked using infrared-emitting LEDs
(Qualisys, Sweden) placed on the following landmarks bilaterally: toe (approximately the second
toe), heel, lateral femoral epicondyle, and greater trochanter. During virtual obstacle crossing,
the vertical distance between the obstacle and both the toe and heel markers was calculated
throughout the crossing step based on the raw marker position data, and the lowest value
between the two markers was used as our measure of foot clearance. For over-ground obstacle
trials, marker positions were recorded at 100 Hz in Qualisys Track Manager and were post-
processed with a 4
th
order Butterworth low pass filter with a cutoff frequency of 6 Hz. Here, the
measure of foot clearance was the same as trials during virtual obstacle crossing.
Statistical Analysis
The change in foot clearance during baseline and training trials was modeled using a
nonlinear, exponential mixed-effects (NLME) model to capture the exponential time course of
learning (Newell et al. 2001) and individual differences in the initial and final performance. The
advantage of using NLME models over individual exponential fits is that NLME models can

62
provide more precise parameter estimates and explain more variance than individual exponential
fits (Winter and Wieling 2016). Moreover, NLME models also capture fixed effects that are
common across individuals. Data from the baseline trial were included because participants were
instructed to minimize foot clearance during this session and exhibited improvements in
performance over the trial. The NLME model consisted of an exponential decay term to capture
the reduction in foot clearance during acquisition and constants that captured initial performance
and the performance plateau (Equation 1):
𝐹𝐹𝐹𝐹
�
𝑖𝑖 , 𝑗𝑗 = 𝐴𝐴 𝑖𝑖 × 𝑒𝑒 − 𝑗𝑗 𝑡𝑡 𝑡𝑡𝑡𝑡 𝑖𝑖 + 𝐷𝐷 𝑖𝑖 (1)
Here, 𝐹𝐹𝐹𝐹
�
𝑖𝑖 , 𝑗𝑗 was the estimated foot clearance for each participant (i = 1:19) and each
obstacle (j = 1:N) where N is the number of obstacles that were crossed successfully (maximum:
140). Ai represents the approximate reduction in foot clearance over the course of practice for our
sample. taui represents the individual acquisition rate. Only obstacles crossed without collisions
were included in the model because foot clearance during collisions would result in negative
values. 𝐹𝐹𝐹𝐹
�
𝑖𝑖 , 𝑗𝑗 , Ai, and Di were expressed in meters and taui was expressed in units of obstacles.
Model parameters were estimated using NLME fit with stochastic Expectation-Maximization
algorithm function, nlmefitsa, from MATLAB R2017a (Natick, MA). The sum of the fixed and
random effects from these models represented participant-specific effects. R
2
values were used to
measure the goodness of fit of the final models. There was no significant difference in foot
clearance during skill acquisition between the left and right legs (t(184)=-0.17, p=0.87).
Therefore, we analyzed all obstacles together in a single ensemble.
The definitions of all dependent variables that were derived from the NLME model and
measured foot clearance are found in Table 1. We estimated initial and final clearance, the

63
relative and absolute amount of skill acquisition, and lastly, acquisition rate. Together, these
variables were used to test how skill acquisition and performance during practice related to
retention. Initial clearance represented the estimated foot clearance at the first obstacle. Final
foot clearance was estimated by summing the amplitude of the exponential term at the end of
practice and the asymptote. The absolute amount of skill acquisition was estimated from the
model as the change in clearance from the first obstacle to the last obstacle. The relative amount
of skill acquisition represented the change in performance from each individual’s initial foot
clearance to final foot clearance normalized by the initial foot clearance and expressed as a
percentage.
Table 4-1. Dependent variables derived from the NLME model.
Variable Definition
𝑨𝑨 𝒊𝒊 + 𝑫𝑫 𝒊𝒊 Initial foot clearance
𝑨𝑨 𝒊𝒊 × 𝒆𝒆 − 𝑵𝑵 𝒕𝒕 𝒕𝒕𝒕𝒕
𝒊𝒊 + 𝑫𝑫 𝒊𝒊
Final foot clearance
𝑨𝑨 𝒊𝒊 − 𝑨𝑨 𝒊𝒊 × 𝒆𝒆 − 𝑵𝑵 𝒕𝒕 𝒕𝒕𝒕𝒕
𝒊𝒊
Absolute amount of skill acquisition
𝒕𝒕 𝒕𝒕𝒕𝒕
𝒊𝒊 Rate of skill acquisition
𝑨𝑨 𝒊𝒊 − 𝑨𝑨 𝒊𝒊 × 𝒆𝒆 − 𝑵𝑵 𝒕𝒕 𝒕𝒕𝒕𝒕
𝒊𝒊 𝑨𝑨 𝒊𝒊 + 𝑫𝑫 𝒊𝒊 × 𝟏𝟏𝟏𝟏𝟏𝟏
Relative amount of skill acquisition
i: Participant ID, N: The number of successful clearances on the treadmill.
We quantified the absolute magnitude of transfer to over-ground walking on Day 1
(absolute transfer) as the difference in foot clearance between the TF and BASE blocks. Each of
these metrics was computed as the average foot clearance over the ten obstacles during each trial.
We conducted an additional analysis using a linear mixed-effects model with a fixed effect of
trial and a random intercept for each participant to test whether there were any within-block
changes in clearance over the ten obstacles. We also quantified the fraction of the improvement

64
in skill in VR that was transferred to over-ground walking on Day 1 (relative transfer) as the
ratio of absolute transfer to the amount of skill acquisition estimated from the NLME model
expressed as a percentage. Foot clearance during RET_VR was computed as the average foot
clearance over all obstacles after removing obstacle crossings where collisions occurred.
Retention of the locomotor skill in VR on Day 2 was calculated as the difference in foot
clearance between RET_VR and END_VR where each of these measures was the average of all
successful obstacle crossings in each block. Similarly, over-ground retention was computed as
the difference in clearance between RET_OG on Day 2 and TF on Day 1. Because the goal of
the task was to reduce foot clearance, more negative values indicated a larger improvement. Foot
clearance during over-ground trials was averaged across ten trials.
Dependent variables calculated for absolute transfer, relative transfer, and over-ground
retention were tested for normality using the Lilliefors test in MATLAB. If the variables satisfied
the normality test, single-sample t-tests were performed to test whether participants transferred
the obstacle negotiation skill to over-ground walking, whether they showed relative transfer
during over-ground walking, and whether they retained the reduction in foot clearance in VR and
over-ground on Day 2. If the variables did not satisfy the normality test, the one-sample
Wilcoxon signed-rank tests were performed instead.
We also used multiple linear regression to determine whether the amount or rate of
locomotor skill acquisition during practice predicted retention of the obstacle negotiation skill in
VR and over-ground on Day 2, respectively. Specifically, we hypothesized that the relative
amount and rate of skill acquisition would predict retention in VR, similar to what has been
observed in previous studies (Park and Schweighofer 2017; Schaefer and Duff 2015; Wadden et
al. 2017). The set of predictors for retention in VR included final foot clearance and the absolute

65
and relative amount of skill acquisition from Equation 1. The predictors for over-ground
retention included each of these predictors, foot clearance during BASE and TF, and the change
in over-ground foot clearance on Day 1. The predictors included in each regression model were
tested for multicollinearity using the Variance Inflation Factor (VIF). If the predictor had VIF
higher than 5, which indicates that the predictor was highly correlated with other predictors, that
predictor was removed from the model. After removing collinear variables, we used the best
subset selection method for variable selection and selected the model with the lowest Bayesian
Information Criterion (BIC) (James et al. 2013). The alpha level was set at p < 0.05. All
statistical analyses were done in MATLAB.
Results
Acquisition and transfer of skilled obstacle negotiation
All participants successfully reduced their foot clearance throughout practice on Day 1
(Figure 4-2). The average R
2
for the NLME model was 0.33 ± 0.22 (SD). The average of the
exponential amplitude parameter Ai was 0.09 ± 0.04 m. The average time constant taui was 21 ±
18 obstacles, which indicated that, on average, individuals needed to successfully negotiate 21
obstacles to achieve 66% of their total reduction in clearance. Lastly, the average of the constant
Di was 0.04 ± 0.01 m. Based on our parameter estimates, foot clearance during virtual obstacle
negotiation in VR was reduced by 69 ± 15% during the acquisition period on Day 1. The average
number of collisions was 3±3 out of 20 obstacles and 20±10 out of 120 obstacles in BASE_VR
and practice blocks, respectively. Practice in VR led to a transfer of reductions in foot clearance
to over-ground walking at the end of Day 1 (Figure 4-3A). The average walking speed over-
ground was 0.9±0.1 m/s, which was similar to the speed on the treadmill. Single-sample t-tests
revealed that there was a reduction in foot clearance of 0.03 ± 0.01 m (t=-10.28, p<0.001) from

66
BASE to TF on Day 1 and this corresponded to a relative transfer of 32% (interquartile ratio
(IQR) 16%, Wilcoxon signed rank test, z=3.82, p<0.001). Additionally, we tested for potential
within-block improvement during walking over-ground. There was no significant difference in
foot clearance when comparing the first trial and last trial during BASE (t(180)=0.21, p=0.83) or
during TF (t(180)=0.78, p=0.43), which indicates that there was no within-block improvement
during over-ground trials (Figure 4-3C).

67

Figure 4-2. Individual foot clearance data and fit from the NLME model. Gray points
represent foot clearance during each obstacle crossing in VR on Day 1, the black curve
represents the participant-specific fit from the NLME, and the dashed curve represents the
group level fit of the NLME. The black points and error bars after the gray dashed vertical line
represent average and standard deviation foot clearance in VR on Day 2, respectively.

Retention of skilled obstacle negotiation
When participants returned to the lab 24 hours later, they generally retained the level of
performance they achieved at the end of Day 1 (Figure 4-2). The average number of collisions
was 2±2 out of 40 obstacles. Single-sample t-tests demonstrated that there was a small, but
significant increase in foot clearance of 0.008 ± 0.01 m (t=2.57, p=0.02) from END_VR on Day
1 to RET_VR on Day 2, which indicates that there was some forgetting of the skill. However,

68
there was no significant change in clearance from TF on Day 1 to RET_OG on Day 2, which
indicates that this skill was retained after a 24-hour retention interval (Figure 4-3B). There was
no significant difference between the first and the last trials during RET_OG (t(180)=1.13,
p=0.26), which indicates that participants sustained the level of performance throughout the trials
over-ground on Day 2 (Figure 4-3C).

Figure 4-3. Over-ground transfer on Day 1 and Day 2. (A) Transfer to over-ground walking
on Day 1. Here, reductions in foot clearance indicate improvements in skill. (B) Over-ground
retention on Day 2. All data are reported as boxplots with the horizontal lines within each box
indicating median values and the bottom and top boundaries of each box indicating the 25th
and 75th percentiles. Dark gray points represent individual data points and the gray lines
connecting the represent the change in foot clearance across trials. (C) Trials during over-
ground obstacle negotiation. Black points represent average foot clearance across all
participants and gray vertical lines represent standard deviations. BASE: baseline block for
over-ground on Day 1, TF: transfer block for over-ground on Day 1, and RET_OG: retention
block for over-ground on Day 2. The asterisks (***) indicate statistically significant
differences from zero at p<0.001.

Prediction of retention in VR and over-ground
After obtaining individual model parameters Ai, taui, and Di, we then calculated
acquisition-related variables (Table 1) to identify the predictors of performance during retention.
The final model with the lowest BIC included only the estimated final foot clearance during
acquisition as a significant predictor of retention in VR (Adjusted R
2
=0.40, t=3.60, p=0.002,

69
=1.14, standard error (SE)=0.33, Figure 4-4A). Participants who achieved lower foot clearance
at the end of Day 1 were more likely to maintain a low foot clearance on Day 2. This association
indicates that the performance at the end of practice is an important predictor of 24-hour
retention.
Variable selection for the model of over-ground retention demonstrated that foot
clearance during transfer to over-ground on Day 1 was the only significant predictor of retention
in over-ground (Adjusted R
2
=0.42, t=3.72, p=0.002, β=0.50, SE=0.13, Figure 4-4B).
Specifically, lower clearance during RET_OG was associated with lower foot clearance during
TF on Day 1. Together, the results suggest that the degree of transfer to a new context on Day 1
is an important predictor of how performance in that context is retained 24 hours later.

Figure 4-4. Associations between performance on Day 1 and retention on Day 2. (A)
Relationship between foot clearance during retention in VR on Day 2 (RET_VR) and the final
foot clearance in VR on Day 1. (B) Relationship between over-ground foot clearance during
retention on Day 2 (RET_OG) and over-ground foot clearance during transfer (TF) on Day 1.
Each participant is represented by a single data point, the solid black line represents the
regression fit, and the dashed gray lines are 95% confidence intervals.

70
Discussion
The objectives of this study were to determine how individual differences in obstacle
negotiation skill acquisition in VR influence retention and transfer to over-ground walking. We
found that our participants successfully reduced foot clearance as instructed during practice
trials, transferred the reduced foot clearance to over-ground obstacle negotiation, and retained the
reduced foot clearance after 24 hours. Furthermore, retention in each environment was associated
with measures of performance during practice in the same environment. Together, our results
demonstrate that locomotor skills can be learned in VR and that measures of performance during
skill acquisition predict retention in a context-dependent manner.
A previous study that used the same instruction to minimize clearance during physical
obstacle negotiation found that participants reduced their foot clearance by about 40% during a
single day of acquisition (van Hedel and Dietz 2004). Based on our NLME results, foot
clearance during virtual obstacle negotiation in VR was reduced by a mean of 69 ± 15% during
locomotor skill acquisition on Day 1. Further, we found evidence of forgetting after a 24-hour
retention period, but the performance decrement was small.
The similarity in performance improvement in physical training from the previous study
(van Hedel and Dietz 2004) and virtual training from our study suggests that obstacle negotiation
skills are learned similarly in VR and the real world. However, the absolute initial and final foot
clearance during practice was higher in our study than the previous study. Initial and final foot
clearance in the van Hedel and Dietz study was approximately 5 cm and 2 cm, respectively while
in our study was approximately 13 cm and 4 cm, respectively. The exaggeration of foot
clearance in VR may be due to the overestimation of obstacle heights as individuals tend to
overestimate height in VR (Asjad et al. 2018). This over-estimation may limit the degree to

71
which participants are capable of reducing foot clearance to avoid a collision. Nevertheless, these
results suggest that VR has the potential to be used as a locomotor skill learning tool that can
facilitate skill transfer and retention in the real world.
Transfer and retention of locomotor skills learned in virtual reality
In line with our hypothesis, the reduction in foot clearance in VR during practice on Day
1 was transferred to over-ground walking. After practicing in VR, individuals reduced their foot
clearance during over-ground walking by a mean of 21 ± 9% relative to baseline. When this
change was expressed relative to the change in performance from the beginning to the end of
practice in VR in a similar manner as Torres-Oviedo and Bastian (Torres-Oviedo and Bastian
2012), this corresponded to a median relative transfer of 32% (IQR 16%). This level of relative
transfer was comparable to what has been observed during adaptation to a gradual perturbation
during split-belt walking where aftereffects in over-ground walking were ~35% of the difference
in asymmetry between baseline and catch trials (Torres-Oviedo and Bastian 2012). Together,
these results demonstrate that locomotor skills acquired through both implicit adaptive learning,
which primarily involves the cerebellum (Morton and Bastian 2006; Taylor and Ivry 2014), and
more explicit skill-based learning, which is more associated with the prefrontal cortex
(Destrebecqz et al. 2005; Taylor and Ivry 2014), transfer to environments that differ from that in
which training occurred. Given that the neural processes underlying these two types of learning
differ markedly from one another, the similarity in transfer between these studies may reflect a
fundamental feature of how skills learned on a treadmill transfer to over-ground walking.
It would be of interest for future studies to determine how transfer differs following over-ground
VR-based training relative to when VR training is implemented on a treadmill. This could help

72
further establish which features of the training environment are most important for promoting
transfer of learning to the real world.
Our results are not consistent with those from a recent study of skilled upper extremity
motor learning which did not transfer from VR to the real world (Anglin et al. 2017). When
participants acquired a sequential isometric pinch task in VR, subsequent performance of the
same task in the real world was significantly slower and approximately 29% less accurate than
performance in the last block of practice in VR, suggesting that there was a decrement of
performance from VR to the real world. However, this discrepancy may stem from differences in
our measure of transfer. Here we calculated transfer as the difference in performance from a
baseline over-ground trial, which occurred before practice and a post-practice trial which also
occurred over-ground. In contrast, Anglin and colleagues calculated transfer as the difference in
performance between the end of practice in VR and a subsequent practice trial in the real world.
Therefore, their results may reflect aspects of transfer of skill acquisition, context-dependent
differences in performance, or both. Future studies should record a baseline trial in each context
to assess transfer of motor skills.
The reduction in foot clearance on Day 2 was generally retained in VR and over-ground
relative to the last block of each environment on Day 1. Although we observed significant
forgetting of the locomotor skill in VR after 24 hours, the increase in foot clearance on Day 2
compared to the end of practice was only 8 mm. Given that the reduction in foot clearance during
skill acquisition was 9 cm, an 8 mm performance decrement can be considered negligible.
Studies of split-belt treadmill adaptation consistently reported a ~80% reduction in step length
asymmetry after multiple days of practice (Day et al. 2018; Leech et al. 2018; Malone et al.
2011). Moreover, individuals who learned to track a target during the swing phase of gait

73
retained approximately an 38% reduction in tracking error compared to baseline on Day 1
(Krishnan et al. 2018). We observed a 59% reduction of foot clearance from baseline to retention
in VR. These results demonstrate that locomotor skills can be acquired and retained following
multiple forms of practice including treadmill-based VR training.
We also found that there were no changes in foot clearance during over-ground retention
relative to the over-ground transfer trial on Day 1. There are two potential explanations for this
observation. First, this may reflect that retention during over-ground walking was comparable to
retention in VR. Some of the observed retention during over-ground walking on Day 2 may have
resulted from the additional practice performed during the retention block in VR as this was
always tested before retention in over-ground. These potential explanations could be
disambiguated in future studies where the evaluation of retention in the non-trained environment
and retention in VR are counterbalanced. Together, these results indicate that many types of
motor skills can be learned and retained in VR, which supports the use of VR as a useful training
modality.
When considering VR-based rehabilitation, however, it is important to note that aging
and pathology can affect the level of retention and sustained transfer. Learning impairments
could result from neurodegenerative changes such as atrophy of prefrontal gray matter or
impairment of the frontostriatal network, which are important in motor learning (Cabeza 2001;
Nieuwboer et al. 2009; Olson et al. 2019; Voelcker-Rehage 2008). For example, Parkinson’s
disease may reduce an individual’s ability to transfer learned motor skills due to declines in
executive functions such as working memory and cognitive flexibility or an inability to form
accurate motor memories of the task (Marinelli et al. 2017).

74
Performance in each environment was a strong predictor of retention in the same environment
The level of foot clearance measured during retention in VR on Day 2 was strongly
associated with the final foot clearance during acquisition in VR such that participants who
achieved lower clearance at the end of acquisition also achieved lower clearance during
retention. In contrast to previous literature where the amount and rate of performance change
during skill acquisition predicted motor skill learning (Park and Schweighofer 2017; Schaefer
and Duff 2015; Wadden et al. 2017), neither the amount nor the rate of skill acquisition were
important predictors of retention in VR on Day 2. Although previous studies did not explicitly
investigate performance at the end of the practice block as a potential predictor, final
performance may have been associated with other significant predictors found in previous
studies such as the amount and rate of skill acquisition. Other work has shown that the level of
performance during acquisition is not always related to performance during retention
(Guadagnoli and Lee 2004; Schmidt and Lee 2011; Schmidt and Bjork 1992; Wulf and
Lewthwaite 2016). For instance, Schmidt and Djork (1992) found that features of practice that
impaired performance during acquisition such as a random practice schedule, reduced feedback
frequency, or contextual interference enhanced long-term retention (Schmidt and Bjork 1992).
Understanding how the practice structure used during VR-based training influences long-term
retention and transfer is an important area of future study.
Retention of foot clearance during over-ground walking on Day 2 was also associated
with foot clearance during the over-ground transfer block, which is equivalent to final foot
clearance over-ground on Day 1. Individuals who achieved lower foot clearance during transfer
on Day 1 exhibited lower foot clearance during retention on Day 2. This is consistent with the
primary predictor of retention in VR and together provides evidence that the final level of

75
performance during practice is an important predictor of retention. Given the moderate amount
of variance accounted for by our models of retention performance, there are likely important,
potentially subject-specific, factors such as variance in the amount of forgetting between days
that also influence retention performance. Overall, the context-specificity of predictors of
retention suggests that different memory processes may underlie the expression of the learned
locomotor skills in differing contexts. Therefore, measures of transfer in addition to retention are
important to fully predict lasting effects of VR training on locomotor performance.
Conclusion
Retention and transfer of motor skills are integral features of motor skill learning and
form the basis by which training in virtual environments can impact real-world behavior. Here,
we demonstrated that individuals could acquire a strategy for skilled obstacle negotiation in VR,
transfer the skill to over-ground walking, and retain the skill after 24 hours. Moreover, the extent
of retention and transfer was predicted by individual differences in the final performance in a
context-dependent manner. Specifically, performance measures in each environment were
strongly correlated with retention performance in the same environment. Overall, our findings
support the use of VR as an effective tool to train skilled obstacle negotiation and facilitate the
transfer of improvements in obstacle negotiation to the real-world.

76
CHAPTER 5
People with Parkinson disease acquire a locomotor skill faster but retain less than
age-matched adults

Abstract
Locomotor learning and generalization are critical foundations for gait rehabilitation in
people with Parkinson disease (PD). However, there is considerable heterogeneity in how well
people with PD retain learned motor skills. This heterogeneity may result from early cognitive
impairments in people with PD. Therefore, we investigated whether PD influences locomotor
skill learning and which domains of cognition explain individual differences in locomotor
learning in people with PD. On Day 1, participants with PD and age-matched control participants
practiced stepping over virtual obstacles while walking on a treadmill. Participants were
instructed to achieve a desired foot clearance during crossing. On Day 2, we assessed retention in
one of two environmental conditions: the same as Day 1 or in a SWITCH context where the
color of the walls and ceiling differed from Day 1. We found that people with PD had a faster
learning rate, but retained less than controls. However, there was no effect of context on
retention performance. In addition, a clinical assessment of memory was positively associated
with locomotor skill retention in people with PD. These results demonstrate that PD negatively
influences retention of acquired locomotor skills, and that clinical assessments of memory may
be important for identifying individuals who might require additional locomotor training.

Keywords
Locomotor learning, Obstacle negotiation, Virtual reality, Cognition, Parkinson disease

77
Introduction
Parkinson disease (PD) is a progressive neurodegenerative disorder resulting from a loss
of dopamine, and manifesting clinically as gait impairments such as short step length and
postural instability during gait, resting tremor, bradykinesia, and a varying level of cognitive
deficits (Weintraub et al. 2008; Weintraub and Burn 2011). A gold-standard treatment for people
with PD is dopamine replacement therapy (DRT) to alleviate motor symptoms such as resting
tremor and bradykinesia (Peterson and Horak 2016). However, DRT does not relieve other motor
symptoms such as instability during gait even in the early stage of PD (Peterson and Horak
2016). One complementary approach to alleviate motor impairments that are not responsive to
DRT is task-specific, goal-based motor skill training (Fisher et al. 2013; Petzinger et al. 2010,
2013). Increasing evidence shows that motor skill training promotes neuroplasticity and reduces
motor impairments in people with PD (Fisher et al. 2013; Petzinger et al. 2010, 2013). Therefore,
skill-based motor learning that focuses on gait and posture accompanied by DRT can aid
effective management for gait and postural impairments in people with PD.
Training to negotiate an obstacle course is commonly prescribed in the clinic to challenge
both motor and cognitive abilities for people with PD (Morris 2006). Obstacle negotiation
requires motor planning and precise control of foot placement and foot height during crossing
(Kim et al. 2018; Marigold et al. 2011; Moraes and Patla 2006), and flexible adaptation to
environmental constraints such as the different size of obstacles or walking speed (Maidan et al.
2018). Although PD impairs the ability to precisely control movement and utilize executive
function (Marinelli et al. 2017; Wu et al. 2015a), people with PD can learn to modify foot height
during obstacle negotiation in response to augmented feedback, but do so more slowly than age-

78
matched controls (Michel et al. 2009). However, it is unclear if differences in skill acquisition
influenced retention of the practiced skill.
Studies have reported mixed results on whether PD influences retention of motor skills.
Impaired retention has been observed during motor tasks such as Wii-based negotiation of virtual
obstacle courses (Mendes et al. 2012), visuomotor adaptation of reaching (Marinelli et al. 2009),
and reaching movement scaling task (Smiley‐Oyen et al. 2002). Other studies found intact
retention compared to healthy controls during standing balance (Peterson et al. 2016; Van
Ooteghem et al. 2017), throwing (Pendt et al. 2011), and sequential arm reaching (Behrman et al.
2000). These inconsistent findings suggest that understanding of how people with PD respond to
locomotor skill learning and retain the specific locomotor skills is necessary.
In addition to retention of a skill when examined in the same environment in which it was
learned, generalizability is a vital aspect of motor learning. The goal of motor learning and
rehabilitation is to be able to perform motor skills outside the context of practice (Winstein et al.
2014). However, people with PD have difficulty executing acquired motor skills in different
contexts (Lee et al. 2015; Nieuwboer et al. 2001; Onla-or and Winstein 2008; Verschueren et al.
1997) even if only the incidental context (e.g., the color) differs (Lee et al. 2015, 2019; Lee and
Fisher 2017; Nieuwboer et al. 2001). This is commonly referred to as context-dependent learning
(CDL): superior retention of learned skills when retention is assessed in the same context as the
training (Wright and Shea 1991). Over-reliance on incidental context may limit the flexible
generalization of learned skills. Lee and colleagues also showed that a laboratory-based finger
sequence tapping task could be learned by people with PD, but their performance deteriorated
when the incidental context during retention differed from practice (Lee et al. 2015, 2019).

79
Although CDL is demonstrated during upper extremity tasks and ADLs, it is unclear if people
with PD also exhibit CDL during a locomotor skill task.
As people with PD commonly have both motor and cognitive impairments, the degree of
cognitive impairment may be an important factor in explaining individual differences in motor
learning. Cognitive impairment is a common non-motor symptom for people with PD, even in
the early stage of the disease (Goldman et al. 2018; Watson and Leverenz 2010). Due to the
close interplay between motor and cognitive abilities, cognition is often used to explain the
heterogeneity of motor learning in people with PD (Marinelli et al. 2017; Nieuwboer et al. 2009).
However, previous studies have focused on investigating the relationship between a single aspect
of cognition or overall cognition on motor learning within a single day without considering
specific cognitive domains of interest (Deroost et al. 2006; Muslimović et al. 2007; Price and
Shin 2009; Vandenbossche et al. 2009). For instance, the reaction time difference between final
practice and immediate retention blocks within the same day for finger sequence learning is
associated with executive function in people with PD (Price and Shin 2009). Moreover, people
with PD with low overall cognitive scores showed less learning during finger sequence learning
(Vandenbossche et al. 2009). One study investigated differences in motor sequence learning in
people with PD who had higher scores of cognition, measured using a battery of
neuropsychological assessments, than those who had lower scores (Deroost et al. 2006). They
found that people with PD who had better motor sequence performance at the end of practice
also had better scores on learning and memory cognitive assessments. However, these studies
investigated single-day motor learning, and did not investigate associations between cognition
and motor retention after period of memory consolidation. Therefore, it remains to be seen

80
whether cognition can explain individual differences in long-term locomotor learning in people
with PD.
Here, we investigated 1) the influence of PD on locomotor skill acquisition, measured by
learning rate, relative to age-matched controls, 2) the influence of PD on locomotor retention
after 24 hours during a precision-based obstacle negotiation skill learning relative to age-
matched controls, 3) how changes in an environmental context affected locomotor skill retention
in people with PD compared to age-matched controls, and 4) which domains of cognition are
associated with inter-individual variability in motor learning in people with PD. We
hypothesized that 1) people with PD would acquire the locomotor skill more slowly compared to
healthy controls, 2) people with PD would have less retention compared to healthy controls, 3)
people with PD would have a greater performance error in a different environmental context
compared to healthy controls, and 4) assessments of executive function and memory would be
positively associated with locomotor skill retention. The results of our study would help explain
the effects of PD on locomotor skill learning and highlight which cognitive domains contribute
to inter-subject variability of locomotor skill learning in people with PD. This information can
provide useful ingredients to explain inter-subject variability of response to gait interventions
and further refine targeted gait interventions for people with PD.

Methods
Participants
We recruited a total of 35 individuals, including 22 people with PD from the neurology
clinic at the University of Southern California and 13 age-matched adults from the community
(Table 5-1). Inclusion criteria for individuals with PD were: 1) a Montreal Cognitive Assessment

81
(MoCA) score of 19 or above, which indicates absent or mild cognitive impairment, 2) ability to
provide informed consent, 3) confirmed diagnosis of PD based on the UK Brain Bank criteria,
and 4) Hoehn and Yahr (H&Y) stage 1 to 3. Exclusion criteria for individuals with PD were 1)
other neurological, cardiovascular, orthopedic, and psychiatric diagnoses that may affect
walking, 2) L-dopa induced hallucinations, and 3) freezing of gait. Inclusion criteria for healthy
older adults were the same as 1) and 2). Exclusion criteria for healthy older adults were
neurological, cardiovascular, orthopedic, and psychiatric diagnoses. The Institutional Review
Board at the University of Southern California approved study procedures, and all participants
provided written, informed consent before testing began. All aspects of the study conformed to
the principles described in the Declaration of Helsinki.
Table 5-1. Demographics and clinical gait and balance assessments. Results are presented as the
mean and standard deviation except for sex and Hohen and Yarh (H&Y). Sex and H&Y are
presented as the number of participants. Overground speed was calculated using 10 Meter Walk
Test (10MWT).
Age-matched control (n=13) PD (n=22) p-value
Age (yrs) 66±8 64±10 0.86
Sex 3 M, 10 F 12 M, 10 F 0.07
Treadmill speed (m/s) 0.94±0.24 0.99±0.22 0.72
Overground speed (m/s) 1.40±0.29 1.38±0.20 0.83
miniBEST 24±4 23±4 0.75
MDS-UPDRS part III - 21±10 -
H&Y - 1=2, 2=20 -
Disease duration (yrs) - 5.3±7.4 -

Experimental Setup
Participants viewed a virtual environment through the HMD in a first-person viewpoint
while walking on a treadmill (Figure 5-1A). The virtual environment was developed using

82
Sketchup (Trimble Navigation Limited, USA), and the participants’ interaction with the
environment was controlled and recorded using Vizard (WorldViz, USA). The virtual
environment consisted of a corridor with obstacles on the left and the right side (Figure 5-1B). A
total of 32 virtual obstacles were placed along the corridor at random intervals between 4 m and
8 m. There were two different heights of obstacles: 0.05 m (LOW) and 0.18 m (HIGH), (Figure
5-1E). Each obstacle height was associated with a different success range above obstacles (0.05-
0.09 m and 0.01-0.05 m, respectively). Given the different associations between the obstacle
height and success range, the task was sufficiently complex for participants to acquire the skill
based on our pilot study. The depth of the obstacles was set to 0.10 m. Movement in the
mediolateral direction was prevented in the virtual environment. Therefore, participants stepped
over obstacles with the leg on the same side as the obstacle. The environment was displayed
within the HTC Vive head-mounted display (HMD), which has a 110-degree field of view, a
resolution of 1080 X 1200 pixels per eye, and a mass of approximately 550 g.
The velocity of the scene and the treadmill speed were synchronized using Python-based
code in Vizard, and an IMU within the HMD controlled the orientation of the viewpoint.
Participants’ lower extremity representations were presented in the first-person viewpoint
consisting of spheres at the toe, heel, knee, and hip on each side and line segments that connected
the spheres (Figure 5-1C).

Experimental Protocol
The schematic experimental protocol is in Figure 5-1D. On Day 1, we performed a
comprehensive neuropsychological battery for people with PD. On Day 2, participants
completed a novel obstacle negotiation task adapted from Michel et al. in a virtual environment

83
on a treadmill (Bertec Fully Instrumented Treadmill, USA) at their self-selected speed.
Participants donned a head-mounted display and a harness, and lightly held onto handrails while
walking (Figure 5-1A). During the baseline block (BASE), we instructed participants to step
over the obstacles as naturally as possible and to avoid collisions. If a collision occurred,
participants heard an unpleasant sound (Figure 5-1F). The BASE condition was divided into two
separate blocks for low and high obstacles, respectively.
After completing BASE, we provided instructions to participants as follows: “You will
step over obstacles of two different heights. Each obstacle has a different success range. This
success range is invisible to you, so you will explore to find this range and maintain your foot
within the success range while crossing. You will receive auditory feedback according to your
performance” (Figure 5-1E). Then, participants listened to examples of auditory feedback
(Figure 5-1DError! Reference source not found.). Three example sounds were played for
participants: 1) a pleasant sound if the foot clearance was within the success range, 2) an
increment sound if the foot clearance was lower than the success range, indicating the need to
increase foot clearance on the next obstacle with the same height, and 3) a decrement sound if
the foot clearance was higher than the success range, instructing them to decrease foot clearance
on their next obstacle with the same height. The duration of the sound was scaled to the distance
from the foot and the success range during the crossing. After participants listened to all possible
sounds, they were asked to identify a set of 30 sound examples (sound test) including the
pleasant sound, the unpleasant sound, two increments, and two decrements sounds that differed
in duration (100% and 40% duration). If participants achieved less than 80% accuracy, they took
another test with the same set of possible sounds until they achieved 80% or more accuracy.
After they passed the sound test, participants started the skill practice blocks.

84
Participants practiced stepping over a total of 192 obstacles in six 32-obstacle bouts.
Auditory feedback was provided according to the performance during skill practice blocks
(Figure 5-1F), and a score representing the number of successful obstacles was presented on the
top of the virtual environment (Figure 5-1B). During the second to last bout, the auditory
feedback was removed except for the collision feedback (NFB) to test immediate retention. The
goal of the task was the same as other practice blocks. For the last bout, participants received
performance feedback again (Figure 5-1D).
After approximately 24 hours, participants revisited the laboratory to complete retention
tests (Day 3). First, we assessed symptom severity of PD using the Movement Disorder Society –
Unified Parkinson’s Disease Rating Scales (MDS-UPDRS). Then, participants completed one of
two retention conditions in a counterbalanced manner. The conditions differed in the
environment that participants observed: either the same environmental condition as Day 2
(SAME) or a different environmental condition (SWITCH). In the SWITCH condition, the color
of the virtual floor was switched (white to grey), and the color of the walls and the ceiling was
switched (light blue to orange) (Figure 5-1D). We tested the influence of the environmental
context on locomotor learning by changing the background color as this type of context
negatively affected finger sequence learning in people with PD (Lee et al. 2015). During the
retention test, participants received no auditory feedback regarding their performance other than
collisions. After the retention test, we asked participants to select the training and testing
environments from a set of 15 pictures with different wall and floor colors to account for the
attentional recognition of the incidental context on their performance. After this interview, we
assessed participants’ gait and balance using clinical assessments. These assessments included a
10-meter walk test (10MWT) and Mini-Balance Evaluation Systems Test (mini-BESTest).

85

Data Acquisition
A 10-camera Qualisys Oqus motion capture system (Qualisys AB, Sweden) captured
position data at 100 Hz from reflective markers placed on the following landmarks on
participants’ lower extremities: second toes, heels, lateral epicondyle of the femur, and greater

Figure 5-1. Experimental setup and protocol. (A) Schematic of the experimental setup. (B)
The virtual environment. (C) Model of the lower extremities. Participants viewed the model
from a first-person perspective. (D) Experimental protocol. BASE: baseline block, RET:
retention block. (E) An example of foot trajectories for two obstacles with prescribed success
ranges in yellow. (F) Collision (empty sound icon) and task performance (filled sound icon)
sound feedback.

86
trochanter. Foot clearance was defined as the minimum foot distance from the height of the
obstacle (Figure 5-2D), and was calculated using the foot trajectory in the sagittal plane when the
foot was passing over the obstacle in real-time during experiments.

Cognitive assessments
We performed a comprehensive neuropsychological assessment to examine five domains
of cognition: executive function, attention, memory, language, and visuospatial processing. Each
domain was assessed by computing composite scores from several assessments. Executive
function was assessed from the Switching fluency, Color-word inhibition, Color-word
inhibition/switching components of the Delis Kaplan Executive Function System (D-KEFS)
assessment (Delis et al. 2001) and the number of errors from the Wisconsin Card Sorting Test
(WCST) (Heaton 1981). Attention composite was assessed from Adaptive Digit Ordering Test
(DOT-A) (Werheid et al. 2002) sequencing and forward span, D-KEFS Color naming, and D-
KEFS Word reading time. Episodic memory composite was assessed from California Verbal
Learning Test (CVLT-II) Trials 1-5 scores (Delis et al. 2000), CVLT-II short delay free recall,
CVLT-II long delay free recall, Brief Visuospatial Memory Test (BVMT-R) (Benedict 1997)
immediate recall total, and BVMT-R delayed recall. Language performance composite was
assessed from D-KEFS letter fluency, category fluency, and Boston Naming Test (BNT) (Kaplan
et al. 1983). Lastly, the visuospatial ability composite was assessed from the Bentons Judgement
of Line Orientation (JLO) (Franzen 2000) and Hooper Visual Organization Test (VOT) (Hooper
1983). All measures for the tests were converted into a z-score based on the sample mean and
SD. Then, we averaged the z-scores across all tests that make up the composite to obtain a score
for that domain of cognition.

87
Data analysis
The primary outcome measures were 1) foot clearance during BASE, 2) learning rate
during skill learning and 3) changes in performance error (PE) between Day 1 and Day 2 as a
measure of retention. PE was calculated as the difference between foot clearance and closest
success range threshold. For instance, if foot clearance was lower than the success range, PE was
the difference between foot clearance and the lower success range threshold. We measured
retention as the difference of performance error between the first four obstacles during practice
(initial practice) trials on Day 1 and during retention trials on Day 2 (PEinit – PEret). Therefore, a
larger value indicates more learning.
State-space model for locomotor learning
We used a state-space model to estimate learning rate and interference. The state-space
model allows to explain trial-by-trial, error-based adaptation behavior (Smith et al. 2006). The
model reflects a process of motor adaptation by updating internal models, or states, based on
sensory prediction error arisen by the discrepancy between desired and actual performance
outcome (Smith et al. 2006). A simple form of the state-space model in motor adaptation
estimates a forgetting index and learning rate. Previous research expanded this form to estimate
interference between two tasks (Kim et al. 2015). Here, we fit a state-space model with an
interference term to estimate learning rate and interference between two obstacles during our
learning task.
We modeled foot clearance during practice from the experiment using the following
state-space model. Foot clearance (FC) at each obstacle n is estimated by:
𝒙𝒙 ( 𝑛𝑛 + 1)
𝑇𝑇 = 𝐴𝐴 ∙ 𝒙𝒙 ( 𝑛𝑛 )
𝑇𝑇 + 𝐵𝐵 ∙ 𝑒𝑒 ( 𝑛𝑛 ) ∙ 𝒄𝒄 ( 𝑛𝑛 ) (1)

88
where 𝒙𝒙 was a N X 2 state vector representing individuals’ estimation of the location of the
success range relative to their baseline foot clearance ( 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵𝐵𝐵 𝐵𝐵 ) for both obstacles, 𝒙𝒙 =
[𝑥𝑥 𝐿𝐿𝐿𝐿 𝐿𝐿 𝑥𝑥 𝐻𝐻 𝐻𝐻 𝐻𝐻𝐻𝐻
]
𝑇𝑇 (1). Due to the natural fluctuation of performance from baseline to practice, we
replaced x of the first obstacle for each height with initial deviation terms for LOW 𝑥𝑥 𝐿𝐿𝐿𝐿 𝐿𝐿 (1) =
𝑥𝑥 𝐿𝐿𝐿𝐿 𝐿𝐿 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 and for HIGH 𝑥𝑥 𝐻𝐻 𝐻𝐻 𝐻𝐻𝐻𝐻
(1) = 𝑥𝑥 𝐻𝐻 𝐻𝐻 𝐻𝐻𝐻𝐻 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 . These terms accounted for the initial deviation
of performance from BASE to practice. The forgetting index, A, corresponded to the relative
change in the estimated target height from one obstacle to next obstacle of the same type.
Learning rate, B, corresponded to the rate of the reduction in subsequent foot clearance for an
error 𝑒𝑒 , where B ranged from 0 (no motor error is accounted in x) to 1 (full motor error is
accounted in x). Learning rate also accounted for the interference between two obstacles as 𝒄𝒄 =
(1 q)
𝑇𝑇 for LOW or 𝒄𝒄 = (q 1)
𝑇𝑇 for HIGH obstacles, where the parameter q modulated the level
of interference between two heights. q ranged from 0 (no interference) to 1 (full interference).
No interference indicated that the different heights of obstacles did not affect each other such
that they were learned independently, whereas full interference implies that errors on a given
obstacle influenced subsequent foot clearance of both obstacles.
𝑒𝑒 ( 𝑛𝑛 ) = �
𝑇𝑇 𝑈𝑈𝑈𝑈𝑈𝑈 𝐵𝐵 𝑈𝑈 − 𝐹𝐹𝐹𝐹 ( 𝑛𝑛 ), 𝑖𝑖𝑖𝑖 𝐹𝐹𝐹𝐹 ( 𝑛𝑛 ) > 𝑇𝑇 𝑈𝑈𝑈𝑈𝑈𝑈 𝐵𝐵 𝑈𝑈 𝑇𝑇 𝐿𝐿𝐿𝐿 𝐿𝐿 𝐵𝐵 𝑈𝑈 − 𝐹𝐹𝐹𝐹 ( 𝑛𝑛 ), 𝑖𝑖𝑖𝑖 𝐹𝐹𝐹𝐹 ( 𝑛𝑛 ) < 𝑇𝑇 𝐿𝐿𝐿𝐿 𝐿𝐿 𝐵𝐵 𝑈𝑈 0, 𝑖𝑖𝑖𝑖 𝐹𝐹𝐹𝐹 ( 𝑛𝑛 ) ≤ 𝑇𝑇 𝑈𝑈𝑈𝑈𝑈𝑈 𝐵𝐵 𝑈𝑈 𝑎𝑎 𝑛𝑛 𝑎𝑎 𝐹𝐹𝐹𝐹 ( 𝑛𝑛 ) ≥ 𝑇𝑇 𝐿𝐿𝐿𝐿 𝐿𝐿 𝐵𝐵𝑈𝑈
(2)

𝑒𝑒 corresponded to the performance error, which is the difference between the current foot
clearance and the given upper (TUPPER) or lower (TLOWER) threshold (success range). If foot
clearance was within the threshold, 𝑒𝑒 was zero.
𝐹𝐹𝐹𝐹 ( 𝑛𝑛 ) = 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵𝐵𝐵 𝐵𝐵 ( 𝑛𝑛 ) + 𝒙𝒙 ( 𝑛𝑛 ) 𝒄𝒄 ( 𝑛𝑛 ) + ε (3)

89
FC was estimated as a sum of estimated target locations for LOW and HIGH, accounted by
interference between two obstacles (Figure 5-3A). Based on the height of the n
th
obstacle,
𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵𝐵𝐵 𝐵𝐵 changed to the mean of either LOW or HIGH.
Statistical analysis
We used hierarchical Bayesian estimation with Markov Chain Monte Carlo (MCMC) to
estimate learning rate and interference from the state-space model. Hierarchical Bayesian
estimation with MCMC estimates both individual parameters for each participant and the
population parameters (Browne and Draper 2006). This allows a greater regularization of the
parameter estimation for each participant by drawing subject-specific estimates towards group-
level estimates, so that overall variance of the model reduces. Additionally, MCMC provides a
probability distribution of the parameter estimates, which shifts the attention of the analysis to
parameter uncertainty rather than finding a point estimate (Kruschke 2014). MCMC requires a
likelihood function and a prior probability for each parameter. The algorithm then yields the
posterior probability which represents the credibility of the prior probability with new
observations taken into account based on Bayes’ rule (Kruschke 2014). The posterior probability
can be summarized with the median and 89% highest density interval (HDI) of the probability
representing the 89% range of posterior probability distribution (Makowski et al. 2019).
The inputs of the MCMC were set as follows. The likelihood function was our state-
space model with free parameters of forgetting (A), learning rate (B), and interference (q). For
the priors, A, B, and q for both groups were sampled in the logistic space as (4)-(6)

𝐴𝐴 ( 𝑝𝑝 ) ~
1
1 + exp � − 𝑁𝑁 �𝜇𝜇 𝐵𝐵 ( 𝑝𝑝 ) + 𝜇𝜇 𝐵𝐵 𝐺𝐺𝐺𝐺𝐺𝐺 𝑡𝑡 𝐺𝐺 × 𝐺𝐺𝐺𝐺 𝐺𝐺 𝐺𝐺 𝑝𝑝 ( 𝑝𝑝 ), 𝜎𝜎 2
𝐵𝐵 ( 𝑝𝑝 ) � �

(4)

90

𝐵𝐵 ( 𝑝𝑝 ) ~
1
1 + exp � − 𝑁𝑁 � 𝜇𝜇 𝐵𝐵 ( 𝑝𝑝 ) + 𝜇𝜇 𝐵𝐵 𝐺𝐺𝐺𝐺𝐺𝐺 𝑡𝑡 𝐺𝐺 × 𝐺𝐺𝐺𝐺 𝐺𝐺 𝐺𝐺 𝑝𝑝 ( 𝑝𝑝 ), 𝜎𝜎 2
𝐵𝐵 ( 𝑝𝑝 ) � �

(5)

𝑞𝑞 ( 𝑝𝑝 ) ~
1
1 + exp � − 𝑁𝑁 �𝜇𝜇 𝑞𝑞 ( 𝑝𝑝 ) + 𝜇𝜇 𝑞𝑞 𝐺𝐺𝐺𝐺𝐺𝐺 𝑡𝑡 𝐺𝐺 × 𝐺𝐺𝐺𝐺 𝐺𝐺 𝐺𝐺 𝑝𝑝 ( 𝑝𝑝 ), 𝜎𝜎 2
𝑞𝑞 ( 𝑝𝑝 ) � �

(6)
where p corresponded to the participant number, and Group was a vector of P (total number of p)
where the PD group was indexed as 0 and the control group was indexed as 1. Therefore, 𝜇𝜇 of
each parameter for the PD and control groups were 𝜇𝜇 and 𝜇𝜇 + 𝜇𝜇 𝐻𝐻 𝐺𝐺 𝐺𝐺𝐺𝐺𝐺𝐺 , respectively. The
participant-specific priors were selected by sampling hyper-parameters, reflecting group
parameters (7)(10). We chose uninformative priors for the hyper-parameters so that the posterior
distributions were primarily influenced by the data.

𝜇𝜇 𝐵𝐵 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝐵𝐵 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝑞𝑞 ~ 𝑁𝑁 (0, 10
3
) (7)

𝜇𝜇 𝐵𝐵 𝐺𝐺𝐺𝐺𝐺𝐺 𝑡𝑡 𝐺𝐺 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝐵𝐵 𝐺𝐺𝐺𝐺𝐺𝐺 𝑡𝑡 𝐺𝐺 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝑞𝑞 𝐺𝐺𝐺𝐺𝐺𝐺 𝑡𝑡 𝐺𝐺 ~ 𝑁𝑁 (0, 10
3
)
(8)

𝜎𝜎 2
𝐵𝐵 ~ 𝜎𝜎 2
𝐵𝐵 ~ 𝜎𝜎 2
𝑞𝑞 ~ 1/Γ(10
− 3
, 10
− 3
) (9)

𝜎𝜎 𝜀𝜀 2
( 𝑝𝑝 ) ~ 1/Γ(10
−3
, 10
−3
)
(10)
MCMC sampling for the hierarchical Bayesian state-space model was implemented in
Rjags (v4-10) in R (version 3.5.1). We estimated parameters using three chains with 20,000
iterations in each chain, 10,000 iteration burn-in, and thinning of 10 for three chains, totaling
3000 samples. We represented the posterior distribution of the parameters of interest for our
study (B and q) (Figure 5-3B-C). We tested the MCMC convergence of the Bayesian model
using 𝑅𝑅 �
as an initial criterion (Gelman and Rubin 1992) and rank normalized 𝑅𝑅 �
as a final
criterion (Vehtari et al. 2021) to ensure that MCMC searched a similar sampling space across
three chains and the final results were similar regardless of the initial sample. The JAGS script
used to build the hierarchical Bayesian state-space model is in Appendix A. We conducted a
model selection analysis using simulated data to determine whether the model with the current

91
parameters was the most favorable among other possible combinations of the parameters. We
also performed a validation analysis to validate the use of the hierarchical Bayesian model by
generating data from a known state-space model and retrieving the model parameters. Details of
these studies are in Appendix A.
We first assessed demographic differences using a two-sample independent t-test.
Moreover, we assessed differences on foot clearance during BASE between group using a linear
mixed-effects model. To determine the influence of PD on locomotor skill acquisition, we used
Bayesian inference such that we compared the median and 89% HDI of the posterior probability
of the group variables for learning rate and interference using a distribution of their effect size.
However, we found a significant Pearson correlation between learning rate and performance
error during initial practice, indicating that performance error during initial practice may be a
confounding variable for learning rate. Therefore, we additionally performed a subset analysis.
We excluded control participants in a step-wise manner, starting from the highest performance
error during initial practice, until the mean performance error of the two groups was not
significantly different. This resulted in the exclusion of three control participants. We estimated
the group-level learning rate from the Bayesian estimation with the remaining ten control
participants and compared the posterior probability between the two groups. The influence of PD
on locomotor skill retention was tested with the measure of retention in SAME using the linear
mixed-effects model. To determine the effect of context in people with PD, we tested the main
effect of context and interaction between group and context in addition to the main effect of
group.
We also calculated the effect size of the group differences. For learning rate, interference,
retention in SAME, and context effect on retention, we used Cohen’s d. We used a conventional

92
threshold for small (> 0.2), medium (> 0.5), and large (> 0.8) effect size. Further, for learning
rate and interference, we calculated % posterior probability that the effect size was greater than
the conventional threshold that included the median effect size, indicating certainty of the effect
size.
To determine if measures of cognition explained between-subject differences in the
amount of learning in people with PD we performed LASSO regression in MATLAB (R2019b,
The MathWorks, USA). The dependent variable was the amount of learning, and the independent
variables were executive function, attention, memory, language, and visuospatial processing. We
added age and MDS-UPDRS III as covariates. Significance was set at p < 0.05. We quantified
the severity of multicollinearity using the Variance Inflation Factor (VIF) in the best model
selected by LASSO. The VIF higher than 5 indicates a concern for multicollinearity (James et al.
2013). We also used Cohen’s f
2
to quantify effect size as this is a commonly used metric for
multiple regression models (Cohen 1988; Steiger 2004). By convention, we used a threshold for
small (> 0.02), medium (> 0.15), and large (> 0.35) effect size (Cohen 1988).
Results
Baseline
We tested differences in foot clearance during BASE between two groups to ensure that
they had similar performance at the beginning of the experiment. There was no significant in foot
clearance during BASE between groups (F(1,66) = 0.01, p = 0.90) or obstacle heights (F(1,66) =
1.52, p=0.22). Moreover, there was no significant interaction between group and obstacle height
(F(1,66) = 1.03, p = 0.31). The average foot clearance for low obstacles was 0.14 ± 0.06 m for
the PD group and 0.16 ± 0.09 m for the control group. Foot clearance for high obstacles was 0.12
± 0.06 m for the PD group and 0.17 ± 0.06 m for the control group.

93

Figure 5-2. Foot clearance during BASE for (A) low obstacles and (B) high obstacles. Each
dot represents one participant.

Skill Acquisition
People with PD acquired the locomotor skill slower (Figure 5-3E-F) and had higher
interference between two obstacles (Figure 5-3H-I) compared to age-matched controls.
Goodness-of-fit measured by the average and SD of the RMSE for the Bayesian state-space
model was 0.03 ± 0.01 m (Supplementary Figure 5-1). The group parameter from the
hierarchical Bayesian estimation showed a medium effect that people with PD had a faster
learning rate than controls (median and HDI: -0.56 [-1.15 – 0.05]) (Figure 5-3F). The probability
that the effect size was greater than the threshold for a medium effect was 56%. The median of
the group parameter for learning rate for people with PD was 0.22 [0.16 – 0.28] and for the
control was 0.17 [0.11 – 0.23] (Figure 5-3E).

94

Figure 5-3. Results of Bayesian modeling. (A) Observed and estimated foot clearance in the
course of trials for an example participant. Each point represents the observed foot clearance
on a given obstacle and the blue line represents estimated foot clearance from the state-space
model. (B) The posterior probability of the estimated learning rate for an example participant.
(C) The posterior probability of the estimated inference for an example participant. (D)
Median of the subject-specific posterior probabilities for learning rate for each group. The
black horizontal line indicates the median learning rate of each group. (E) The posterior

95

Faster learning rate was significantly associated with greater performance error during
initial practice, measured at the first four obstacles during skill acquisition (Figure 5-4B). During
initial practice, performance error was lower for people with PD (0.05 ± 0.03 m) compared to
controls (0.08 ± 0.05 m) (F(1,33) = 9.17, p = 0.005, Figure 5-4A). Moreover, there was a
significant Pearson’s correlation between performance error during initial practice and learning
rate regardless of the group (r
2
= 0.38, p < 0.001). Due to the potential confounding effect of the
initial performance error, we performed an additional subset analysis. The selection criteria for
participants included in this analysis is described in the Methods. The median for people with PD
was 0.22 [0.16 – 0.28] and for the controls was 0.18 [0.11 – 0.27] (Supplementary Figure 5-2A).
The group parameter for learning rate now showed a small effect that people with PD had faster
learning rate than controls (-0.27 [-0.91 – 0.36]) with 57% probability (Supplementary Figure 5-
2B).
probability for the group-level learning rate. (F) The posterior probability of the effect size of
group differences on learning rate. The black dotted line indicates the median of the posterior
probability, and the gray dotted line indicates zero. (G) Subject-specific median interference
for each group. (H) The posterior probability of the group-level inference. (I) The posterior
probability of the effect size for the group-level interference. Red and blue represent people
with PD and age-matched controls, respectively.

96

Figure 5-4. (A) Initial performance error in PD and Control group. Red dots represent
participants with PD and blue dots represent control participants. The black solid line
represents a median for each group. (B) A correlation between performance error during initial
practice and learning rate. The black solid line and gray shaded area indicate a fit and 95%
confidence intervals from the correlation regardless of the group.

There was also a small effect size that people with PD had greater interference than the
controls (median and HDI: -0.39 [-1.07 – 0.31]) (Figure 5-3I). However, the probability that the
effect size was greater than the threshold for a small effect was 66%. The median of the group
parameter for interference for people with PD was 0.31 [0.13 – 0.49] and for the control was
0.25 [0.12 – 0.37] (Figure 5-3H).
Retention and CDL
When comparing retention performance to initial performance during skill acquisition,
we found that people with PD learned less (-0.02 ± 0.02 m) than controls (-0.05 ± 0.06 m) in the
SAME context (F(1,31) = 12.03, p = 0.002) (Figure 5-5A). However, this result may be due to
the initial difference between groups as the control group had a higher initial performance error
than the PD group and there was no difference in performance during retention between groups
(0.03 ± 0.02 and 0.03 ± 0.01 m, respectively, F(1,33) = 0.41, p = 0.53).

97
Environmental context only affected the control group during locomotor skill learning
(Figure 5-5B). There was no significant main effect of Context on retention in people with PD (β
= -0.002, F(1,31) = 0.01, p = 0.91) and no significant main effect of Group on retention in the
SWITCH context (β = 0.007, F(1,31) = 0.17, p = 0.68). However, there was a significant main
effect of Context on retention in controls (β = 0.05, F(1,31) = 7.45, p = 0.01), indicating
retention in the SAME context was higher than that in the SWITCH context in controls.
Particularly, there was significant interaction between context and group (β = -0.05, F(1,31) =
7.45, p = 0.02) on retention. Specifically, the difference in retention between the context in the
control group was higher than that in the PD group (β = -0.05, F(1,31) = 4.39, p = 0.04).

Figure 5-5. Performance error during retention compared to initial practice as a measure of
retention in the SAME and SWITCH context in people with PD and controls. SAME context
illustrates retention and SWITCH context illustrates context-dependent learning. Red dots
represent the PD group and the blue dots represent the control group. Solid lines represent
medians for each condition for each group. Dotted lines represent zero.

Cognition
Clinical assessments of cognition explained individual differences in the amount of
learning in people with PD. Because there was no significant effect of context in people with PD,

98
we combined participants with PD regardless of the environmental context that they were in for
retention and determined which domains of cognition were associated with locomotor learning.
We found that a model with visuospatial processing, memory, and age was the best model to
explain differences in learning between participants (Table 5-2). The amount of learning was
positively associated with the composite score for memory (Figure 5-6Table 5-2), indicating that
people with PD who had higher memory function had greater learning of the locomotor skill.
The VIF for the three variables in this model was less than 2, suggesting a negligible concern of
the multicollinearity.
Table 5-2. Final regression result to explain the amount of locomotor learning with cognition.
Variables β Standard error t p Cohen’s f
2

Intercept -0.28 0.17 -1.63 0.12 -
Visuospatial -0.26 0.22 -1.19 0.25 0.18
Memory 0.46 0.19 2.45 0.025* 0.22
Age 0.51 0.16 3.24 0.005* 0.32

Figure 5-6. Adjusted regression plots for
the amount of learning as a function of
memory. The gray dots represent
participants with PD. The solid red line
represents the regression fit, and the
dashed lines are 95% confidence intervals.

Discussion

99
The objective of this study was to determine how PD influences obstacle negotiation skill
learning and which domains of cognition are associated with inter-individual variability in
locomotor learning in people with PD. We examined locomotor learning with a virtual obstacle
negotiation task and we performed a comprehensive neuropsychological battery to determine
cognitive domains associated with locomotor learning. We found that people with PD learned the
locomotor skill faster, but learned less than healthy controls. However, we did not find any effect
of environmental context on locomotor learning in people with PD. Furthermore, we found that
assessments of memory were positively associated with locomotor skill retention in people with
PD. Our results demonstrate that people with PD have an intact ability to acquire locomotor
skills but may have impaired retention compared to healthy controls. In addition, our results
suggest that clinical assessments of memory may help identify people with PD who may need
additional training to promote long-term retention of locomotor skills.

Locomotor skill acquisition and retention
There was a medium effect of faster learning rate for people with PD compared to
healthy controls. This result was inconsistent with previous literature (Michel et al. 2009;
Mongeon et al. 2013; Roemmich et al. 2014). Interestingly, the faster learning rate was strongly
correlated with performance error during initial practice. We observed that people with PD had
lower initial performance error than healthy controls at the beginning of the acquisition phase.
As the location of the success range was always lower than participant’s baseline foot clearance,
the lower initial performance error may result from reduced willingness to exert effort to move in
people with PD (Le Heron et al. 2018; Salamone et al. 2018). Willingness to exert effort is
mediated by both the dopaminergic projections from the ventral striatum to ventromedial

100
prefrontal cortex (Levy and Glimcher 2012) and non-dopaminergic neurotransmitters such as
serotonergic (Meyniel et al. 2016) and cholinergic systems (Nunes et al. 2013). PD results in
damages in not only the dopaminergic system, but also non-dopaminergic systems, including
serotonergic and cholinergic systems (Fox 2013), resulting in the decrease in an individual’s
willingness to exert effort. In our locomotor learning task, the reduced willingness to exert effort
may have a favored effect during skill acquisition, demonstrating faster learning rate in people
with PD. Participants with PD may quickly decrease their foot clearance, which consequently
achieved successful foot clearance for the learning task due to the reduced willingness to exert
effort. Therefore, future studies can further examine the associations between locomotor and
willingness to exert effort in people with PD.
Faster learning rate is particularly striking because this result was in the opposite
direction of the group difference from Michel and colleague’s study, even though the obstacle
crossing learning task was relatively similar. A possible explanation for this inconsistent result
may be due to the inclusion of people with PD who had mild balance impairments. Michel and
colleagues included participants with PD who had modified H&Y (mH&Y) primarily 2.5 while
our study included people with PD who had H&Y primarily 2. mH&Y 2.5 reflects mild balance
impairment while H&Y 2 indicates no balance impairment. This may suggest that the speed of
locomotor learning reduces as the disease progresses, especially as the balance becomes
impaired in people with PD. Therefore, future studies can further examine the potential effect of
balance impairments in learning rate during locomotor learning while matching the initial
performance to understand the influence of PD on locomotor skill acquisition.
We also found that people with PD had less locomotor learning compared to healthy
controls, consistent with our hypothesis and previous literature (Marinelli et al. 2009; Mendes et

101
al. 2012; Smiley‐Oyen et al. 2002). PD causes a loss of dopaminergic projections in the posterior
striatum, including the putamen, in the early stage of the disease, and this damage spreads
anteriorly as the disease progresses (Poewe et al. 2017). The putamen is a structure that is part of
the sensorimotor circuit of the basal ganglia (Alexander et al. 1986), which is associated with
performing a well-known, habitual movement (Lehéricy et al. 2005; Redgrave et al. 2010). It is
possible that impairment in the sensorimotor circuit is the underlying cause of impaired retention
during locomotor learning. Alternatively, less retention in people with PD compared to healthy
controls may be due to the difference in initial performance error. We calculated the difference in
performance error between the initial and retention trials as an indicator of retention. The lower
initial performance error in the PD group may lead to a reduction in the overall change between
the two time points, which results in the “worse” retention. Future studies can match initial
performance error of the locomotor task and increase the sample size to confirm this result.

Context-dependent locomotor learning
Contrary to our hypothesis, there was no effect of incidental context on locomotor
learning in people with PD. A recent study found that context-dependency during motor
sequence learning was present only in people with PD who have freezing of gait (FOG), but not
in people with PD who do not have FOG and controls (Lee et al. 2019). Lee and colleagues also
argued that the presence of CDL in their previous studies (Lee et al. 2015; Lee and Fisher 2017)
might be due to the inclusion of people with advanced PD, although they did not assess disease
severity. Another study investigating home-based physical therapy for balance also included
participants with more advanced PD (Nieuwboer et al. 2001). The H&Y scores for Nieuwboer
and colleagues’ study were 2.5-3, while participants with PD in our study scored H&Y 1-2.

102
Moreover, although Nieuwboer and colleagues did not specify the inclusion or exclusion of
FOG, one of their outcome measures was the occurrence of freezing (Nieuwboer et al. 2001),
suggesting that some participants may have experienced FOG. As we included people with only
mild to moderate PD and excluded participants who have FOG, our participants with PD may not
have developed over-reliance on the incidental context during motor learning. However, more
strikingly, we found context-dependency in our control participants. This may have resulted from
one participant who remarkably increased his performance error from initial practice to retention.
Skill acquisition from this participant showed no consistent reduction of performance error until
the end of skill acquisition, which may lead to the unusual increase in performance error during
retention. Overall, people with mild to moderate PD may be able to retain locomotor skills
regardless of changes in incidental context.

Associations between cognition and locomotor learning
Finally, a clinical assessment of memory was a significant predictor of locomotor
learning in people with PD. Previous studies generally report that low cognitive function is
associated with worse motor learning in a single-day setting (Deroost et al. 2006; Price and Shin
2009; Vandenbossche et al. 2009). Particularly, people with PD who had better motor sequence
learning at the end of practice also had a better learning and memory score, assessed by a battery
of cognitive assessments (Deroost et al. 2006). Our results extend these findings by
demonstrating that the reliance on memory during motor sequence learning may extend to time-
delayed retention. Other domains of cognition were not significant predictors in our study, even
though previous studies found positive associations between executive function and motor
learning (Price and Shin 2009; Vandenbossche et al. 2009). One reason for the finding in

103
memory, but not executive function, in our study might be that the previous studies measured
retention within the same day of the motor skill practice. This may indicate that executive
function may be an important contributor to reflect short-term retention, while memory plays an
important role in predicting longer-term, time-delayed retention. Therefore, memory may be a
valuable domain of cognition during locomotor learning and may be useful to identify subgroups
of people with PD who require further training to promote better retention.
Interestingly, the LASSO approach selected visuospatial processing in the model, but this
was not a significant predictor of locomotor learning. A recent study found that visuospatial
processing was positively associated with retention of a fine motor skill in older adults (Lingo
VanGilder et al. 2018). However, this association may differ in people with PD that the
variability of visuospatial processing among participants may not be sensitive enough to explain
inter-individual variability in locomotor learning. Further, we found a significant correlation
between clinical scores of memory and visuospatial processing, indicating that visuospatial
processing also may contribute to explain the variance in the model. Future studies can
investigate a wider range of PD that may impair visuospatial processing to determine the
association between visuospatial processing and locomotor learning.

Limitations
There are two limitations of this study. First, we had a small sample size in each group,
and the sample size of the PD group was almost twice larger than the control group. This
unbalanced and small sample size made the retention and CDL analysis difficult as the groups
were divided into two conditions. Second, people with PD and healthy control had a significant
difference in initial performance error for our locomotor learning task. The difference in initial

104
performance error between the groups led to difficulty dissociating the true effect of PD on
learning rate and retention. Therefore, future studies should match the initial performance
between the two groups and determine if the results differ from ours.

Conclusions
This study aimed to investigate if PD influences locomotor skill learning and determine
which domain of cognition contributes to locomotor skill learning in people with PD. Our results
demonstrated that people with PD could acquire the locomotor skill but had less retention of the
skill. Moreover, we observed that environmental context did not affect locomotor skill retention
in people with PD. Finally, a clinical assessment of memory was positively associated with the
level of retention. Our findings suggest that people with mild to moderate PD may have impaired
retention but may not experience context-dependent learning during locomotion. Moreover, our
results support that a clinical assessment of memory may be a useful measure for targeted
physical interventions by identifying people with PD who may require more training.

105
Supplementary Figure 5-1. Observed and estimated foot clearance. Each panel represents a
participant. Gray dots represent observed foot clearance in the course of locomotor learning. Red
and blue solid lines represent foot clearance estimated by the Bayesian modeling for people with
PD and the controls, respectively.

106
Supplementary Figure 5-2. Group-level posterior probabilities for learning rate with the subset
of participants. (A) The posterior probability for the group-level learning rate. (B) The posterior
probability of the effect size of group differences on learning rate. The black dotted line indicates
the median of the posterior probability, and the gray dotted line indicates zero.

107
CHAPTER 6
Corticostriatal resting-state functional connectivity is associated with locomotor
skill learning in people with Parkinson disease

Abstract
Communication between the basal ganglia and cortex is an important mediator of motor
skill learning. Particularly, different stages of motor learning are associated with functionally
defined circuits of the basal ganglia. As motor skill learning during gait, or locomotor skill
learning, underlies gait rehabilitation for people who have walking impairments such as
Parkinson disease (PD), we examined whether corticostriatal resting-state functional connectivity
(rsFC) explained individual differences in different stages of locomotor skill learning in people
with PD. A seed-to-seed analysis was performed using striatal seeds (caudate, anterior putamen,
and posterior putamen) and cortical seeds (primary motor cortex (M1) and dorsolateral prefrontal
cortex (DLPFC)). Then, we investigated the correlation between corticostriatal rsFC and
locomotor skill learning. Moreover, a whole-brain connectivity analysis using a seed-to-voxel
analysis was used to confirm the correlation results and explore other areas associated with
locomotor skill learning. We found that faster learning rate was associated with stronger rsFC
between the anterior putamen and the DLPFC, and the whole-brain connectivity analysis
confirmed this correlation. Our result indicates that the speed of locomotor learning may be
associated with the strength of a cognitive loop of the corticostriatal circuits in the resting brain
in people with PD.

108
Introduction

Motor skill learning is a fundamental component of many everyday activities. An
important brain structure that modulates motor skill learning is the basal ganglia (Ferrazzoli et al.
2018; Krakauer et al. 2019; Makino et al. 2016). Specifically, the functionally segregated
corticostriatal circuits (Alexander et al. 1986) are proposed to process information in different
stages of motor learning (Lehéricy et al. 2005; Redgrave et al. 2010; Wu et al. 2015b). The
cognitive circuit connects the caudate nucleus and anterior putamen and dorsolateral prefrontal
cortex (DLPFC) to process goal-directed, higher-level cognitive information during the initial
phase of learning (Lehéricy et al. 2005; Lipton et al. 2019; Redgrave et al. 2010; Wu et al.
2015b). The sensorimotor circuit connects the posterior putamen to the sensorimotor cortices
such as the motor cortex, somatosensory cortex, and supplementary motor area. This circuit is
involved in well-learned motor skills, or retention, and habits (Lehéricy et al. 2005; Lipton et al.
2019; Redgrave et al. 2010; Wu et al. 2015b). The unique involvement of corticostriatal circuits
in distinct stages of motor learning suggests that injury to different parts of the basal ganglia can
result in compromised motor learning at specific stages. Understanding the differential influence
of the corticostriatal circuits in motor learning may explain individual differences in motor
learning in individuals with damage to the basal ganglia.
An ideal model to investigate the influence of impairments in differing corticostriatal
circuits in human motor learning is Parkinson disease (PD). PD is a progressive
neurodegenerative disorder resulting from a loss of dopaminergic neurons in the basal ganglia
(Ehringer and Hornykiewicz 1998; Weintraub et al. 2008; Weintraub and Burn 2011). During
motor skill acquisition, people with PD consistently show preserved capability to acquire new
motor skills during bimanual coordination tasks (Swinnen et al. 2000; Verschueren et al. 1997),

109
reaching tasks (Behrman et al. 2000; Ghilardi et al. 2003), a throwing task (Pendt et al. 2011),
obstacle crossing (Michel et al. 2009), postural sequence learning (Smiley-Oyen et al. 2006), and
balance maintenance tasks (Peterson et al. 2016; Van Ooteghem et al. 2017). However, some
studies found that people with PD learn more slowly than healthy controls (Ghilardi et al. 2003;
Michel et al. 2009; Smiley-Oyen et al. 2006). In addition, studies of motor retention in people
with PD have reported mixed results. Some studies found that people with PD showed a similar
degree of long-term retention to healthy controls during standing balance tasks (Peterson et al.
2016; Van Ooteghem et al. 2017), throwing (Pendt et al. 2011), and arm reaching (Behrman et
al. 2000). Other studies found that there was impaired retention in people with PD compared to
healthy controls during a Wii-based obstacle negotiation game (Mendes et al. 2012), visuomotor
adaptation (Marinelli et al. 2009), and a scaling task (Smiley-Oyen et al. 2003). These conflicting
results may come from the varying levels of impairment in the corticostriatal circuits involved in
motor skill learning in people with PD. As the cognitive circuit is involved more during early
motor learning, the strength of connectivity within this circuit may influence the speed with
which a skill is learned. As the sensorimotor circuit is more involved during later stages of motor
learning, the degree of long-term retention may be explained by the strength of connectivity
within the sensorimotor circuit. Therefore, assessing the associations between the level of
connectivity within distinct corticostriatal circuits and measures of motor learning aspects in
people with PD can provide valuable insights about the functional roles of corticostriatal circuits
during skill learning.
One way to infer the synchrony of activity within the corticostriatal circuits is through the
quantification of resting-state functional connectivity (rsFC). rsFC is a measure of the temporal
correlation between two spatially distant brain regions and is thought to represent the intrinsic

110
functional connections throughout the brain (Biswal et al. 1995; Cole et al. 2010). Many studies
have demonstrated that PD alters corticostriatal rsFC (Hacker et al. 2012; Helmich et al. 2010;
Kim et al. 2017c; Luo et al. 2014; Nieuwhof and Helmich 2017; Owens-Walton et al. 2018;
Tahmasian et al. 2015; Yang et al. 2013). A seed-based methodology has shown that FC between
the posterior striatum and the sensorimotor cortex decreases in the early phase of PD (Ham et al.
2015; Helmich et al. 2010; Luo et al. 2014; Tahmasian et al. 2015; Yan et al. 2013; Yang et al.
2013). A whole-brain analysis also revealed global functional connectivity is reduced in the
SMA, prefrontal cortex, and putamen—i.e., the sensorimotor corticostriatal circuit—in people
with PD who in the OFF-state with respect to dopamine medications (Berman et al. 2016;
Koshimori et al. 2016; Luo et al. 2015; Sang et al. 2015; de Schipper et al. 2018; Tinaz et al.
2017; Wei et al. 2014; Yu et al. 2013; Zhang et al. 2015). Although this hypoconnectivity is
partly normalized by dopamine medications to a similar level of age-matched adults (Hacker et
al. 2012; Tahmasian et al. 2015; Yang et al. 2013), a whole-brain analysis using a graph analysis
demonstrated that people with PD in the ON-state with respect to dopamine medications still
show a widespread functional connectivity across the brain, rather than functionally specific
connectivity, than healthy older adults (Kim et al. 2017c). The brain regions that show rsFC are
often similarly found during tasks and the strength of rsFC is highly correlated with task-based
coactivation maps (Di et al. 2013; Park et al. 2020; Toro et al. 2008), suggesting that rsFC may
imply readiness to perform a movement. Therefore, corticostriatal rsFC may explain individual
differences in motor skill learning in people with PD.
As the corticostriatal circuits serve an important link to motor skill learning, the two
corticostriatal circuits are good candidates to explain individual differences in motor skill
learning in people with PD. In healthy young adults, baseline rsFC is associated with future

111
motor skill acquisition and learning (Bonzano et al. 2015; Faiman et al. 2018; Hamann et al.
2014; Mary et al. 2017; Stillman et al. 2013; Wu et al. 2014, 2017). For instance, rsFC between
the primary motor cortex (M1) and the posterior putamen was negatively correlated with
learning rate, indicating that disengagement of the sensorimotor circuit may explain faster early
learning (Mary et al. 2017). Moreover, another study found that rsFC between the caudate
nucleus and M1 and the somatosensory cortex (S1) was negatively correlated with the level of
motor skill acquisition during implicit motor sequence learning. This suggests that a mixture of
cognitive (caudate nucleus) and sensorimotor (M1 and S1) circuits may be associated with worse
acquisition of the motor skill (Stillman et al. 2013). Although these studies provided valuable
insights into the relationship between motor learning and corticostriatal connectivity, it remains
unclear how the two distinct corticostriatal circuits involve in different aspects of motor skill
learning. Moreover, previous studies have focused on the associations between rsFC and motor
skill learning during upper extremity tasks. As locomotion is one of the major impairments in
people with PD, understanding the functional relevance of the corticostriatal rsFC for locomotor
skill learning may be critical. Given the clinical and neural heterogeneity of PD, rsFC may offer
an additional explanation for the individual differences in motor skill learning in people with PD.
Therefore, we aimed to investigate whether corticostriatal functional connectivity
explains inter-individual differences in locomotor learning in people with PD. We defined
cognitive corticostriatal circuits with two striatal regions (Helmich et al. 2010): the caudate
nucleus and the anterior putamen. Given higher activation in the cognitive circuits of the basal
ganglia during early motor learning, measured by learning rate, we hypothesized that 1) rsFC
between the caudate nucleus and the DLPFC would be positively associated with learning rate,
2) rsFC between the anterior putamen and the DLPFC would be positively associated with

112
learning rate. We defined the sensorimotor circuit with the posterior putamen (Helmich et al.
2010). Given more activation in the sensorimotor circuits during late motor learning, measured
by retention of motor skills, we hypothesized that 3) rsFC between the posterior putamen and M1
would be positively associated with retention in people with PD. Our results would allow us to
understand the behavioral relevance of the intrinsic organization of the brain with respect to
locomotor learning in people with PD.

Methods
Participants
We recruited 20 people with mild to moderate PD (Table 6-1) from the neurology clinic
at the University of Southern California (USC). Inclusion criteria were 1) a Montreal Cognitive
Assessment (MoCA) score of 19 or above, which indicates normal or mild cognitive impairment,
2) ability to provide informed consent, 3) confirmed diagnosis of idiopathic PD based on UK
Brain Bank criteria (Hughes et al. 1992), and 4) Hoehn and Yahr (H&Y) stage 1 to 3 (Goetz et
al. 2008), which indicates mild to moderate PD. Exclusion criteria were 1) other neurological,
cardiovascular, orthopedic, and psychiatric diagnoses, 2) L-dopa-induced hallucinations, 3)
freezing of gait, and 4) metal implants in the body. All participants maintained their regular
medication schedule for the scans. All participants provided written consent approved by the
Institutional Review Board at USC.
Table 6-1. Participant characteristics. MDS-UPDRS: Movement Disorder Society – Unified
Parkinson’s Disease Rating Scale. H&Y: Hoehn and Yahr scale. LEDD: Levodopa equivalent
daily dosage in mg/day. MoCA: Montreal Cognitive Assessment.
ID Sex Age
More affected
side
Disease
duration
MDS-UPDRS motor
(H&Y)
MoCA
1 M 74 L 36 15 (2) 27
2 M 54 L 1 23 (2) 28

113
3 F 53 L 4 21 (2) 30
4 F 64 L 4 10 (1) 28
5 F 72 R 4 25 (2) 30
6 F 51 L 7 5 (2) 28
7 M 70 L 4 9 (2) 28
8 M 45 L 3 18 (2) 29
9 M 56 L 3 26 (2) 25
10 M 68 Equal 1 17 (1) 27
11 F 81 Equal 2 25 (2) 22
12 F 67 R 10 16 (2) 28
13 M 62 L 5 14 (2) 26
14 M 56 Equal 5 18 (2) 28
15 M 71 R 3 34 (2) 27
16 F 76 L 0.5 32 (2) 30
17 F 72 L 2 29 (2) 28
18 M 46 R 1 32 (2) 27
19 F 64 R 5 11 (2) 28
20 M 73 R 3 51 (2) 24

Experimental Protocol
The full experiment took place over two days (Figure 6-1). On Day 1, structural and
resting-state functional MRI scans were obtained from people with PD. Following the scans,
participants learned a novel obstacle negotiation task in a virtual environment on a treadmill
(Bertec Fully Instrumented Treadmill, USA) at their self-selected walking speed. The details of
the learning task are in Chapter 5.

114

Figure 6-1. Experimental protocol. BASE: Baseline obstacle crossing. MDS-UPDRS:
Movement Disorder Society-Unified Parkinson’s Disease Rating Scale. RET: Retention.

MRI Acquisition
All images followed the Lifespan Human Connectome Project protocol (Harms et al.
2018). MR images were acquired with Siemens Prisma 3T MRI system (Siemens, Erlangen,
Germany) equipped with echo-planar imaging (EPI) capabilities, using a 32-channel head coil
for radio frequency transmission and signal reception at the University of Southern California
Health Science Campus. The outline of the acquisition protocol was 1) T1 acquisition, 2)
fieldmap, and 3) resting-state functional MRI (rsfMRI). High-resolution anatomical imaging
included one sagittal T1-weighted magnetization prepared rapid acquisition gradient echo
(MPRAGE) sequence with a 0.8 X 0.8 X 0.8 mm resolution and repetition time (TR) of 2.4 s and
echo time (TE) 2.22 ms. The flip angle is 8°, and the field of view (FOV) is 256 X 240 mm.
Functional images were acquired using a BOLD contrast sensitive gradient EPI sequence with a
2 X 2 X 2 mm resolution and TR of 0.8 s and TE of 37 ms, which resulted in a total of 420-time
points. The flip angle was 52°, and the FOV was 208 mm. Participants were instructed to focus
on a white cross on a black background, remain as still as possible, and not fall asleep. We
obtained three runs of the functional images, and each functional scan took 5:36 minutes (Figure
6-2).

115
Pre-processing of resting-state functional MRI
Functional imaging data were pre-processed and analyzed with FEAT v6.00 (FMRI
Expert Analysis Tool), part of FSL v5.0 (FMRIB’s Software Library, www.fmrib.ox.ac.uk/fsl)
(Figure 6-2A). The first five time points were deleted to allow the magnetization to reach its
dynamic equilibrium. The remaining scans were spatially realigned using boundary-based
registration. Subsequently, each voxel's time series was realigned temporally to the acquisition of
the first slice to remove any excessive head movement. Anatomical and functional images were
registered using nonlinear registration to a standard EPI template centered in Montreal
Neurological Institute (MNI) space in 2 mm resolution. The distorted functional images due to
magnetic field inhomogeneities were corrected using a fieldmap image. One participant’s
structural image (participant ID 1) showed visible motion artifacts due to dystonia. Therefore, we
excluded this participant from further analyses. We calculated head movement in six degrees of
freedom (DOF, x, y, z, yaw, pitch, and roll) for the remaining 19 participants. Then, we
calculated a mean framewise displacement (Jenkinson et al. 2002) using the six DOF head
movement. We excluded one participant (participant ID 8) from the analysis who had a mean
framewise displacement greater than the 2 * standard deviation of the group mean framewise
displacement (threshold = 0.32 mm) (Yan et al. 2013). With the remaining 18 participants, we
performed an independent component analysis with ICA-AROMA to de-noise motion and other
physiological artifacts that were not the actual BOLD signal (Pruim et al. 2015). The de-noised
images were smoothed with a 5 mm full-width-at-half-maximum (FWHM) Gaussian kernel. The
smoothed images were band-pass filtered at 0.01-0.1 Hz to remove low frequency confounds
such as scanner drifts and remove non-neuronal, non-resting-state-related signals, respectively.
Finally, the functional image was segmented into white matter (WM) and cerebrospinal fluid

116
(CSF) using tissue probability maps. We used the striatal region of interests (ROIs; see below)
and eight nuisance regressors (WM, CSF, six motion parameters) in a partial linear regression
model to obtain a spatial map of β for each run for each participant (Figure 6-2C).

Striatal and Cortical Regions of Interests
Striatal ROIs were anatomically defined from each individual’s structural image,
including bilateral putamen and caudate nucleus, using FIRST (Figure 6-2B). The putamen was
further divided into the anterior and posterior putamen based on the location of the anterior
commissure (Helmich et al. 2010). This subdivision was to account for the functional differences
between these regions (Tian et al. 2020) and asymmetries in basal ganglia impairment between
anterior and posterior putamen in people with PD (Lehéricy et al. 2005). Due to the small
anatomical area of the segregated striatum, three voxels in the anterior and posterior putamen
from the anterior commissure were not included to avoid overlapping signals (Helmich et al.
2010).
We used two methods to define cortical ROIs (Figure 6-2B). For bilateral M1, we defined
it using a probabilistic map from FSL’s Harvard-Oxford Cortical Atlas because M1 is an
anatomically well-defined area. We used a threshold of 50% probability to reduce the influence
of the signal from the outer boundaries of M1. Because the DLPFC is a functionally, not
anatomically, defined region, we used previous literature to obtain the left DLPFC [MNI
coordinates: -30, 30, 45] (Yang et al. 2016) and the right DLPFC [MNI coordinates: 21, 53, 28]
(Elfmarková et al. 2016) as a proxy of regions involved in the cognitive circuit. We created 8
mm spherical kernel ROIs for each DLPFC that were centered on each coordinate provided
above. The investigators visually inspected the ROI registration for each individual.

117

Resting-State Functional MRI Post-processing
We used a seed-to-seed analysis based on our hypotheses: the cognitive circuit was
represented by (1) FC between bilateral caudate nucleus – DLPFC and (2) FC between bilateral
anterior putamen – DLPFC, and the sensorimotor circuit was represented by (3) FC between
bilateral posterior putamen – M1. First, we performed a seed-to-voxel analysis using a general
linear model (GLM) to concatenate three runs for each striatal seed (Figure 6-2D). Then, we
extracted mean parameter estimates as a measure of FC from the corresponding cortical seed
masks for each striatal seed using the parameter estimate image (Figure 6-2E). For instance, if
the left caudate was used as the striatal seed, the left DLPFC was used as a cortical seed mask.

Associations between rsFC and Locomotor Learning
Our main analysis used a correlation between learning parameters and the relevant
parameter estimates of rsFC. We assessed whether learning rate was associated with rsFC within
the cognitive circuits, as measured by connectivity between (1) the caudate nucleus and DLPFC
and (2) the anterior putamen and DLPFC, and whether retention was associated with rsFC within
the sensorimotor circuit, (3) the posterior putamen and M1. We obtained learning rate and
measures of retention as described in Chapter 5. Since learning rate was quantified as a posterior
distribution from hierarchical Bayesian estimation, we used a bootstrap approach to test for
correlations between learning parameters and rsFC. We first randomly selected a learning rate
value from a posterior distribution, which consisted of 3,000 data points, for each participant
using a weighted random sample. Then, the selected learning rate was correlated with cognitive

118
rsFC. We iterated this process 10,000 times and estimated the median and 89% confidence
intervals (CIs) of r
2
and p values. Significance was set at the p < 0.05 level.
As a secondary analysis, we performed a whole-brain analysis using a seed-to-voxel
analysis to find additional functional connectivity with subcortical seeds in relation to either
learning rate or retention. We performed multiple regression analysis using the mixed-effect
GLM implemented in FEAT. The previous GLM results from each participant were modeled
with the fixed effect of demeaned learning rate or retention for striatal seeds. The random effect
included participants. We corrected for multiple comparisons using a cluster-based correction at
z > 3.1 and p < 0.001.

119

Figure 6-2. Flowchart of data processing and analyses. (A) Structural and functional image
pre-processing. (B) Seed extraction. (C) Partial linear regression for each run for each
participant to regress out nuisance regressors. (D) Seed-to-voxel analysis using general linear
modeling to obtain a seed-to-voxel spatial map across all runs for each participant. (E)
Extraction of parameter estimates with the cortical seed masks.

Results
Seed-to-seed analysis
Correlation analyses revealed that the median learning rate estimated from the state-space
model was significantly associated with rsFC between the right anterior putamen and right
DLPFC (R = 0.51, p = 0.03, Figure 6-3A). Specifically, stronger rsFC between these two regions
was associated with a faster learning rate, consistent with our hypothesis. However, we did not

120
find any significant correlations with other corticostriatal rsFC (e.g. sensorimotor circuit, Figure
6-3B).
Given that we estimated a distribution of learning rate from hierarchical Bayesian
estimation, we performed a bootstrap correlation analysis by randomly sampling learning rate
from the distribution. We found a trend toward an association between rsFC between the right
anterior putamen and DLPFC and learning rate (median [89% CI] p-value: 0.05 [0.01-0.21]
(Figure 6-3C) and R: 0.47 [0.31-0.58] (Figure 6-3D)).

121

Figure 6-3. (A) The correlation between the functional connectivity with the DLPFC and learning rate
estimated from the state-space model. Each data point represents a median of the posterior distribution
from the hierarchical Bayesian estimation for each participant. The error bars represent 89% CIs. The
solid line and shaded area indicate the fit and CIs of the correlation, respectively. (B) The correlation
between rsFC between the posterior putamen and M1 and retention. Each data point represents each
participant. The solid line and shaded area indicate the fit and CIs of the correlation, respectively. (C-
D) p-value and R from the Bootstrap results for the correlation between rsFC between the anterior
putamen and the DLPFC and learning rate. The dotted line indicates the median of the bootstrap result.

Seed-to-voxel analysis
To confirm the seed-to-seed analysis and explore additional functional connectivity
associated with motor learning variables, we performed a whole-brain analysis with the striatal
seeds and learning parameters as a secondary analysis. We found significant positive rsFC
between left and right anterior putamen and bilateral DLPFC, and this rsFC was associated with

122
learning rate (Table 6-2, Figure 6-4). We also found that rsFC between the left anterior putamen
and the right anterior cingulate cortex (ACC) was associated with learning rate, and rsFC
between the right anterior putamen and the left precuneus was associated with learning rate.
Overall, the whole-brain analyses were generally consistent with the seed-to-seed analyses.
Table 6-2. Significant functional connectivity with learning parameters.
Striatal seed
Region
(Brodmann area)
Voxel
MNI Coordinate
Z score p
x y z
L Anterior
Putamen
R DLPFC (BA 9) 319 28 50 36 4.87 <0.001
L DLPFC (BA 9) 177 -4 48 22 4.42 <0.001
R Anterior cingulate cortex
(BA 8)
173 4 40 34 4 <0.001
R Anterior
Putamen
L DLPFC (BA 46) 257 -48 38 14 4.43 <0.001
R DLPFC (BA 9) 205 10 30 24 4.55 <0.001
L Precuneus (BA 7) 157 -12 -74 36 4.26 <0.001

123

Figure 6-4. Regions whose functional connectivity was associated with learning rate for (A)
left and (B) right anterior putamen. DLPFC: dorsolateral prefrontal cortex.

Discussion
The objective of this study was to investigate the associations between resting-state
corticostriatal functional connectivity and aspects of locomotor learning in people with PD. We
hypothesized that cognitive circuits of the basal ganglia, measured by 1) rsFC between the
caudate nucleus and the DLPFC and 2) rsFC between the anterior putamen and the DLPFC,
would be positively associated with learning rate, and sensorimotor circuits of the basal ganglia,
measured by 3) rsFC between the posterior putamen and M1, would be positively associated
with retention in people with PD. We found that stronger rsFC between the anterior putamen and

124
the DLPFC was associated with faster learning rate in people with PD. However, other measures
of corticostriatal rsFC were not associated with locomotor learning variables. Our results suggest
that the cognitive circuit of the basal ganglia may be an important region for facilitating
locomotor learning in people with early-stage PD.
Consistent with our hypothesis, the cognitive circuit involving the anterior putamen and
the DLPFC was positively associated with learning rate in people with PD. This is in line with
the functional framework of the basal ganglia circuits, where the cognitive circuit is engaged
when individuals learn new motor skills (Lehéricy et al. 2005; Redgrave et al. 2010) . A recent
study found that the cognitive circuit in the basal ganglia at rest was associated with de novo
motor skill learning in healthy young adults (Choi et al. 2020). Participants learned continuous
tracking of a cursor movement, and the overall success rate at the end of the practice over
multiple days was negatively associated with rsFC between the anterior caudate and the DLPFC.
The authors argued that the greater disengagement of the cognitive circuit was predictive of
better learning performance. However, the authors in this study did not find a positive
association between learning rate and the cognitive circuit of the basal ganglia. The discrepancy
between the negative association from this study and the positive association from our study may
be due to the difference between healthy young adults and people with PD. The strength of the
cognitive circuit of the basal ganglia may be associated with early learning, but in the opposite
direction between two populations. The association between the stronger cognitive circuit and
faster learning rate may indicate an aberrant reorganization to compensate for the damaged basal
ganglia in people with PD. People with PD who are in their early stage show widespread hyper-
rsFC (Gorges et al. 2015). This hyper-connectivity has been suggested as a potential neural
underpinning of a compensatory mechanism to maintain motor behavior in the similar level to

125
the healthy adults (Gorges et al. 2015), which may support our notion that the positive
association between rsFC in cognitive circuit and learning rate may be indicative of aberrant
reorganization. As we did not have a control group in our study, this could be an interesting
future area of investigation to understand the functional changes of the intrinsic connectivity
relevant to motor learning in people with PD.
Particularly, the association between learning rate and the cognitive circuit including the
anterior putamen, but not the sensorimotor circuit, may suggest the importance of the anterior
putamen in people with PD in their early stage. One explanation for the lack of correlation with
behavior for the sensorimotor circuit including the posterior putamen may be that the loss of
dopaminergic neurons in the posterior putamen may lead to a functional reorganization to
maintain motor behavior. A previous study found that rsFC between the posterior putamen and
the sensorimotor area was reduced in people with PD in their early-stage compared to healthy
controls (Helmich et al. 2010). This reduction in functional connectivity was paralleled by an
increase in rsFC between the anterior putamen and the inferior parietal cortex, which is part of
the sensorimotor network. This suggests that functional reorganization of the corticostriatal
circuits may occur with a relatively intact anterior putamen in people with PD in the early stage
(Helmich et al. 2010). Similarly, another possibility may be that there is insufficient variability in
the sensorimotor circuit, including the posterior putamen, to observe a significant correlation
with locomotor learning due to the loss of dopamine in the posterior putamen (Redgrave et al.
2010). Overall, our result further supports the notion that the anterior putamen may reflect
behavioral significance in the early stage of PD.
In the whole-brain analysis, we found that the right anterior cingulate cortex (ACC) was
associated with the left anterior putamen for learning rate. The ACC plays a crucial role in

126
complex cognitive functions, such as adapting to rapidly changing environments, that modulates
downstream cognitive controls to guide willed behavior (Brockett et al. 2020; Paus 2001). In
healthy brains, fMRI studies have shown strong functional connectivity between the ACC and
the caudate nucleus (Graff-Radford et al. 2017; Robinson et al. 2012), reflecting the cognitive
loop involving the caudate nucleus (Alexander et al. 1990). In people with PD in their early
stage, the ACC is further functionally connected to the anterior putamen (Helmich et al. 2010).
This altered functional connectivity may indicate that the cognitive loop becomes stronger as a
result of the damage to the motor loop involving the posterior putamen. Our finding supports the
notion that locomotor learning, particularly the rate of the learning, involves complex cognitive
control that is modulated by the cognitive loop with the anterior putamen.
Moreover, we found that the left precuneus was associated with the right anterior
putamen for learning rate. The precuneus is also shown to link with the basal ganglia, both the
caudate nucleus (Di Martino et al. 2008) and the putamen (Cacciola et al. 2017). The precuneus
is known for its role in oculomotor control (Harsay et al. 2011) as well as conscious information
processing (Doyon et al. 2003; Lehéricy et al. 2005). For instance, the precuneus is more
activated during performing more complex tasks such as longer motor sequences compared to
simpler tasks (Doyon et al. 2003; Lehéricy et al. 2005). The oculomotor system integrates spatial
information with smooth and voluntary saccadic eye movement (Ding and Gold 2013; Hikosaka
et al. 2000; Seger and Miller 2010). Specifically, the caudate nucleus in the oculomotor loop is
activated during visual stimulus processing for categorization (Seger and Miller 2010), which is
often involved in perceptual decisions during learning (Ding and Gold 2013). In people with PD,
there is a general reduction in activities in the precuneus (Ceballos-Baumann 2003; Tessitore et
al. 2019), as well as functional connectivity between the precuneus and both the caudate (Kwak

127
et al. 2010; Tessitore et al. 2019) and the putamen (Tessitore et al. 2019; Yan et al. 2015). The
unique rsFC with the anterior putamen in association with learning rate in our study may suggest
a compensatory reorganization of the oculomotor circuit for spatial information processing for
locomotor learning.
Our study had two major limitations. First, we did not have a control group. Although the
results with people with PD provide valuable insights into inter-individual variability in
locomotor skill learning, having a control group would help disambiguate whether the effect of
associations between rsFC and locomotor learning is PD-specific. Second, the inherent limitation
of the seed-based analysis is that the results may differ based on the selection of seed locations.
Future studies may confirm our results with data-driven analyses such as an independent-
component analysis or a graph analysis.

Conclusions
Heterogeneity in people with PD can cause inter-individual variability in locomotor skill
learning. One way to understand the heterogeneity in people with PD is using resting-state
functional connectivity. Particularly, due to the critical role of the basal ganglia in motor
learning, corticostriatal circuits are valuable candidates to explain the variability in locomotor
skill learning. We found that rsFC between the anterior putamen and the DLPFC was associated
with the speed of locomotor learning. This indicates that the cognitive circuit of the basal ganglia
may be a key marker to explain inter-individual variability in locomotor skill learning in people
with mild to moderate PD. Further research is warranted with larger sample size and inclusion of
a control group to confirm our results.

128
CHAPTER 7
DISCUSSION
Learning to deal with dynamic environmental constraints such as obstacles is a vital
component of rehabilitation. Particularly, locomotor training for people with PD emphasizes
skilled walking intervention such as obstacle negotiation as they have a high risk of falls related
to walking. Due to the damage to the basal ganglia, people with PD often show impaired motor
learning. However, previous literature reports mixed findings about the level of motor learning
impairments likely due to the heterogeneity of PD. We used VR as our tool to provide motor
tasks to address the influence of PD and the individual differences on locomotor learning. VR is
a promising tool to provide more ecological, functional tasks with varying environmental
constraints in a safe and controlled environment (Canning et al. 2020). Although these
advantages of VR are leading to an increase in the clinical application of VR-based training
interventions and research, the premises of using VR rely on the notion that 1) visual information
about the environment and the body can facilitate skilled obstacle negotiation and the 2) changes
in performance in VR will lead to lasting changes in real-world performance.
Therefore, first, we aimed to assess how the quality of visual information about the body
impacts obstacle negotiation performance in VR. Second, we aimed to determine how individual
differences in locomotor skill learning in VR influence retention and transfer of learned skills to
the real world. Third, we aimed to determine how PD influences the acquisition and retention of
skilled locomotor behavior in VR compared to age-matched adults and to assess how changes in
incidental context affect the retention of the recently learned locomotor skill in people with PD
and age-matched adults. Lastly, we aimed to determine the relationship between corticostriatal
rsFC and inter-individual differences in motor learning and CDL in people with PD. Our
findings will help us better understand the effect of PD and the intrinsic functional organization

129
of the brain in people with PD on locomotor skill learning. Ultimately, understanding the
behavioral significance of the altered brain networks aid in designing more effective, targeted
clinical interventions for locomotor learning.

Findings from our work
The presence of visual information about the body influenced the consistency of obstacle
negotiation performance in VR. Greater visual information about the body imposed during
virtual obstacle negotiation allowed individuals to perform more consistently while stepping over
obstacles. We also found that obstacle crossing behavior was associated with gaze behavior
during obstacle crossing. Specifically, greater downward head pitch was associated with closer
placement of the trailing foot relative to the obstacle prior to crossing, and farther lead foot
placement after crossing. These results demonstrate that the quality of visual information about
the body contributes to both feedforward and feedback aspects of visuomotor coordination
during VR-based obstacle negotiation.
Further, we provided support for the use of VR as an effective tool to train skilled
obstacle crossing and facilitate the transfer of improvements in obstacle crossing to the real-
world. We trained healthy young adults with a precision-based obstacle crossing task in VR and
found that they can transfer the acquired locomotor skill from the virtual environment to the real
world, and we also found that the locomotor skill was sustained after 24 hours. Moreover, the
sustained performance after 24 hours was associated with performance at the end of practice
trials, but in a context-specific manner. This may indicate that different memory processes
underlie the expression of learned locomotor skills in differing contexts. Given the growing use

130
of VR for motor skill learning in research and rehabilitation, our results provide evidence of
long-term improvements and transfer of skilled locomotion using VR applications.
However, we found that people with PD learned the locomotor skill faster than healthy
controls but had less retention after 24 hours during obstacle negotiation skill learning in VR.
Moreover, environmental context did not influence retention of this locomotor skill in people
with PD. We also found that a clinical measure of memory was explained individual differences
in locomotor skill retention in people with PD. Our results suggest that PD negatively influences
retention of acquired locomotor skills and that a clinical assessment of memory may be able to
identify individuals who may require additional locomotor training to facilitate proper retention.
Finally, individual differences in locomotor skill acquisition in people with PD were
associated with the cognitive circuit of the basal ganglia, measured using resting-state functional
connectivity. We found that learning rate of skilled locomotion was associated with rsFC
between the anterior putamen and the DLPFC, which are part of the cognitive circuit of the basal
ganglia. This suggests that the cognitive circuit may be an important marker to distinguish the
speed of locomotor learning in people with PD and may be useful to understand disease-related
changes in motor learning behavior in mild to moderate PD.

Impact of Dissertation
VR as a promising tool for research and rehabilitation
The premise of using VR in research and rehabilitation is that users receive sufficient
sensory information from the virtual environment to perform motor tasks consistently, and their
performance is retained and transferred to the real-world performance. Our findings suggest that
visual information about the lower extremity induces more consistent obstacle crossing behavior

131
and overall similar performance to the real-world performance (Kim et al. 2017b). Along with
previous studies that demonstrate the importance of visual information about the surrounding
environment during obstacle negotiation (Jansen et al. 2010; Patla and Greig 2006; Timmis and
Buckley 2012), we further provided evidence of the critical role of visual information about the
body for accurate visuomotor integration during obstacle negotiation. This also advocates the
inclusion of body information in VR applications that incorporate motor skills and movement.
Moreover, the locomotor skill that participants learned in VR was transferred to the real-world
and sustained in both VR and real-world after 24 hours in healthy young adults. We also showed
that VR could provide a platform to examine locomotor skill learning behavior for people with
PD and age-matched adults. Detailed evidence for motor learning in VR supports the efficacy of
VR applications for research and rehabilitation.
Influence of PD and their individual differences on locomotor learning
Mixed results regarding the effects of PD on motor learning in the literature may come
from the heterogeneity of cognition and intrinsic functional connectivity of the brain in people
with PD. Our findings suggest that people with PD indeed have altered locomotor learning
behavior, in which they learned faster, but retained less. Also, inter-subject variability in
locomotor learning in people with PD was partially explained by the clinical assessment of
memory as well as resting-state functional connectivity of the cognitive circuit of the basal
ganglia. Our finding of the cognitive circuit, not the sensorimotor circuit, associated with
locomotor learning indicates that inter-individual differences observed during motor learning
may result from the degree of the intact cognitive circuit of the basal ganglia rather than the
sensorimotor circuit. Our results can be used as potential ingredients of future research
investigating targeted physical interventions to facilitate motor learning for people with PD.

132

Limitations and future work
This dissertation has several methodological limitations. In the work investigating
visuomotor coordination during obstacle negotiation in VR, we did not directly measure gaze.
Although head orientation can be a proxy of gaze, future studies will benefit from a direct
measure of gaze to fully understand how we use visual information during obstacle negotiation.
This ability is now made possible with integrated eye-trackers available in newer head-mounted
displays.
In the work investigating locomotor skill learning in people with PD compared to age-
matched controls, the two groups had different performance errors at the beginning of skill
acquisition despite having the same foot clearance during baseline. The difference in initial
performance error makes it difficult to determine if the significant difference in retention that we
observed between the groups is due to the effect of Parkinson’s disease or if it resulted from the
initial difference performance error. Future studies can incorporate a study design to match two
groups’ performance at the beginning of skill acquisition. One approach for this type of pre-
training is to create a ‘success range’ relative to each participant’s baseline performance
(Appendix B, Experiment 1). Although this provides some window for effective motor learning,
a concern using this approach is that participants may change their behavior or strategy when
skill acquisition begins compared to baseline. In this case, the success range defined relative to
baseline performance is not useful to induce effective motor learning. Moreover, specifically for
obstacle crossing skill learning, there is substantial variability in foot clearance during obstacle
crossing on a treadmill. Due to this variability, it may be difficult to implement a consistent
change of the success range location. For instance, if investigators decide to increase the location

133
of the success range relative to baseline foot clearance, it may be difficult to implement this
approach for participants who already have near maximum foot clearance due to the
biomechanical constraints of the hip flexion during walking. On the other hand, if investigators
choose to decrease the location of the success range relative to baseline foot clearance, it may be
impossible to place the success range for participants who have foot clearance close to the
obstacle.
Another approach is to use a pre-training task to shape participants’ baseline performance
to begin from similar initial performance (Appendix B, Experiment 2). In our pilot study, we
used a pre-training task to minimize foot height while stepping over the virtual obstacles.
Therefore, all participants began the locomotor skill acquisition from their ‘minimum’ foot
clearance during obstacle crossing. Although this approach effectively places participants in
similar initial positions, it may require longer pre-training for individuals who have difficulty in
motor learning. Consequently, the additional pre-training may cause fatigue during the actual
motor skill learning. Therefore, it is important to incorporate a cut-off level of similarity among
participants to prevent fatigue from excessive training blocks.
Lastly, we did not have control participants to investigate how cognition and intrinsic
functional connectivity of the brain explain individual variability of locomotor skill learning. As
we found altered locomotor skill learning in people with PD, these associations may indicate an
aberrant link between movement and cognition. Therefore, the inclusion of the control group will
be critical in future studies to establish whether the associations we found are specific to people
with PD or ubiquitous across populations.
The next step of this work would be to investigate the involvement of cognition and
cognitive circuit of the basal ganglia in varying levels of cognitive impairments in people with

134
PD. Our participants with PD had intact cognitive functions based on a battery of cognitive
assessments. However, mild cognitive impairment (MCI) prevails even in the early stage of PD
(Aarsland et al. 2010; Weintraub et al. 2008; Weintraub and Burn 2011). Therefore, the inclusion
of people with PD with varying levels of cognitive impairments may provide us more sensitive
predictors of locomotor skill learning. Moreover, this will provide us a deeper understanding of
the role of cognition and its associated brain functional connectivity in locomotor skill learning.

135
APPENDIX A
VALIDATION OF THE HIERARCHICAL BAYESIAN STATE-SPACE
MODEL: SIMULATION STUDY
Here, we performed a simulation study to validate the use of the hierarchical Bayesian
estimation for estimating the parameters of our state-space model. We modeled foot clearance
during practice blocks (total number of n (N) = 160) using the following state-space model.
𝒙𝒙 ( 𝑛𝑛 + 1)
𝑇𝑇 = 𝐴𝐴 ∙ 𝒙𝒙 ( 𝑛𝑛 )
𝑇𝑇 + 𝐵𝐵 ∙ 𝑒𝑒 ( 𝑛𝑛 ) ∙ 𝒄𝒄 (1)
𝑦𝑦 ( 𝑛𝑛 ) = 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵𝐵𝐵 𝐵𝐵 + 𝒙𝒙 ( 𝑛𝑛 ) ∙ 𝒄𝒄 + 𝜀𝜀 (2)
𝑒𝑒 ( 𝑛𝑛 ) = �
𝑇𝑇 𝑈𝑈𝑈𝑈𝑈𝑈 𝐵𝐵 𝑈𝑈 − 𝑦𝑦 ( 𝑛𝑛 ), 𝑖𝑖𝑖𝑖 𝑦𝑦 ( 𝑛𝑛 ) > 𝑇𝑇 𝑈𝑈𝑈𝑈𝑈𝑈 𝐵𝐵 𝑈𝑈 𝑇𝑇 𝐿𝐿𝐿𝐿 𝐿𝐿 𝐵𝐵𝑈𝑈
− 𝑦𝑦 ( 𝑛𝑛 ), 𝑖𝑖𝑖𝑖 𝑦𝑦 ( 𝑛𝑛 ) < 𝑇𝑇 𝐿𝐿𝐿𝐿 𝐿𝐿 𝐵𝐵𝑈𝑈
0, 𝐺𝐺 𝑜𝑜 ℎ 𝑒𝑒 𝐺𝐺 𝑒𝑒 𝑖𝑖 𝑒𝑒𝑒𝑒 (3)

𝒙𝒙 is a N X 2 state matrix that represents the deviation from baseline foot clearance
( 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵𝐵𝐵 𝐵𝐵 ) for two heights of obstacles, 𝒙𝒙 = ( 𝑥𝑥 𝐿𝐿𝐿𝐿 𝐿𝐿 𝑥𝑥 𝐻𝐻 𝐻𝐻 𝐻𝐻𝐻𝐻
). 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵𝐵𝐵 𝐵𝐵 was calculated as the mean
foot clearance of the last five obstacles during BASE from the observed data. Forgetting (A)
corresponds to the decay rate of the memory. Learning rate (B) corresponds to the change in
subsequent foot clearance for an error of a given magnitude 𝑒𝑒 , where B ranges from 0 (no motor
error is accounted in x) to 1 (full motor error is accounted in x). 𝑒𝑒 corresponds to the motor error,
which is the difference between the current foot clearance and the given upper (TUPPER) or lower
(TLOWER) threshold (success range). If foot clearance was within the threshold, 𝑒𝑒 was zero.
Subsequent foot clearance is updated based on the error calculated and states updated.
Interference between two heights of obstacles is accounted for in the model, 𝒄𝒄 = (1 𝑞𝑞 )
𝑇𝑇 for
LOW or 𝒄𝒄 = ( 𝑞𝑞 1)
𝑇𝑇 for HIGH obstacles, where the parameter 𝑞𝑞 modulates the level of
interference between two heights. 𝑞𝑞 ranges from 0 (no interference) to 1 (full interference). No
interference indicates that two heights of obstacles did not affect each other such that two heights
of obstacles were learned separately. When there is interference, errors on a given obstacle

136
influenced subsequent foot clearance of both obstacles. Learning rate is scaled by interference so
that higher interference would reduce the overall motor error accounted for x while lower
interference would increase the error accounted. y is observed foot clearance. 𝜀𝜀 is measurement
noise, corresponding to the natural noise of movement execution. Due to the natural fluctuation
of performance from baseline to practice, we also added initial deviation terms for both LOW
( 𝑥𝑥 𝐿𝐿𝐿𝐿 𝐿𝐿 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 ) and HIGH ( 𝑥𝑥 𝐻𝐻 𝐻𝐻 𝐻𝐻𝐻𝐻 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 ). These terms accounted for the initial deviation of
performance during practice from BASE.

Generate simulated behavioral data using the state-space model
We generated 20 artificial datasets using this state-space model. The artificial datasets
were equally separated into two groups, reflecting people with PD and age-matched adults. The
first group was simulated by drawing parameters from normal distributions with the following
means and standard deviations based on experimental data for people with PD: baseline foot
clearance for low and high obstacles (mean and SD: 0.14 ± 0.06 and 0.12 ± 0.06, respectively),
forgetting (A, 0.99 ± 0.003), learning rate (B, 0.27 ± 0.06), and interference index (q, 0.47 ±
0.07). The second group was simulated with experimental data for age-matched adults: baseline
foot clearance for low and high obstacles (0.17 ± 0.04 and 0.16 ± 0.07, respectively), A (0.99 ±
0.003), B (0.05 ± 0.009), and q (0.37 ± 0.09).

Estimating parameters of state-space model using a hierarchical Bayesian estimation with
MCMC
The hierarchical Bayesian state-space model estimation with MCMC has two major
advantages to other optimization algorithms that used to estimate learning-related parameters

137
with a state-space model such as least squares or Expectation Maximization (EM) algorithms
(Albert and Shadmehr 2018). First, unlike the other algorithms in which the model estimates
each individual’s parameters independently, the hierarchical Bayesian estimation with MCMC
estimates individual parameters for each participant and estimates the population estimates of the
parameters (Browne and Draper 2006). This allows a greater regularization of the parameter
estimation for each participant compared to an individual parameter estimation without the
population estimates, leading to a better generalization of the parameters across other studies.
Second, the MCMC methods provide a probability distribution of the parameter estimates while
other algorithms provide a point estimate of parameters. Focusing on the probability distribution
shifts the attention of the analysis to parameter uncertainty rather than finding a point estimate.
The criticisms on point estimates have been widely discussed in the reproducibility crisis in
science (Cumming 2014). Refocusing our statistical inference into parameter uncertainty and
probability rather than a point estimate can allow us to infer certainty in our locomotor learning
for each participant and each group.
Markov Chain Monte Carlo (MCMC) algorithms are commonly used in modern
Bayesian estimation to explore the model’s solution space (Kruschke 2014; van Ravenzwaaij et
al. 2018). MCMC algorithms require inputs of a likelihood function and the prior probability for
each of the parameters. The algorithms then yield the posterior probability that represents a
statistical probability after taking into account the new observations relative to the prior
probability (Kruschke 2014). The detailed MCMC process using Gibbs sampling is in Figure A-
1 (Kruschke 2014). First, we generate prior distributions of parameters of interest (Figure A-1A).
From there, we arbitrarily choose starting values for all parameters of interest (Figure A-1B).
Then, with a 6-dimensional parameter space 𝜃𝜃 = { 𝜃𝜃 1
, … 𝜃𝜃 6
}, MCMC iteratively sample a new

138
proposal for each parameter d from the conditional distribution 𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 | 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 ), where 𝜃𝜃 − 𝑑𝑑 is all
other parameters except for 𝜃𝜃 𝑑𝑑 (Figure A-1D-G). As the posterior probability is a product of all
parameters’ distributions (Chib and Greenberg 1995):
𝑃𝑃 ( 𝜃𝜃 | 𝑋𝑋 ) = 𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 | 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 ) = 𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 | 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 ) 𝑃𝑃 ( 𝜃𝜃 − 𝑑𝑑 | 𝑋𝑋 )
(4)
If we use this to calculate the acceptance ratio 𝛼𝛼 ( 𝜃𝜃 ∗
| 𝜃𝜃 ):

𝛼𝛼 ( 𝜃𝜃 ∗
| 𝜃𝜃 ) = min {1,
𝑃𝑃 ( 𝜃𝜃 ∗
| 𝑋𝑋 ) 𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 | 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 )
𝑃𝑃 ( 𝜃𝜃 | 𝑋𝑋 ) 𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 ∗
| 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 ∗
)
}
(5)

𝛼𝛼 ( 𝜃𝜃 ∗
| 𝜃𝜃 ) = min � 1,
𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 ∗
| 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 ∗
) 𝑃𝑃 ( 𝜃𝜃 − 𝑑𝑑 ∗
| 𝑋𝑋 ) 𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 | 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 )
𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 | 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 ) 𝑃𝑃 ( 𝜃𝜃 − 𝑑𝑑 | 𝑋𝑋 ) 𝑃𝑃 ( 𝜃𝜃 𝑑𝑑 ∗
| 𝑋𝑋 , 𝜃𝜃 − 𝑑𝑑 ∗
)
�
𝛼𝛼 ( 𝜃𝜃 ∗
| 𝜃𝜃 ) = 1
where 𝜃𝜃 ∗
represents a new proposal. The acceptance ratio is always 1, which means the newly
proposed sample is always accepted. If all parameters are evaluated, this completes one iteration.
If there is any parameter that has not been evaluated, repeat the step Figure A-D-G until all
parameters are evaluated. Following the decision, if the number of iterations is less than 10,000,
repeat Figure A-C-H until the number of iterations reaches 10,000. One chain of MCMC process
for each participant consists of these steps.

139

Figure A-1. A flowchart describing artificial dataset generation and Bayesian estimation using
Markov Chain Monte Carlo (MCMC) using Gibbs sampling.

140
The inputs of the MCMC were set as follows. The likelihood function was our state-
space model with free parameters of forgetting (A), learning rate (B), and interference (q). For
the priors, A, B, and q for both groups were sampled in the logistic space as

𝐴𝐴 ( 𝑝𝑝 ) ~
1
1 + exp �− 𝑁𝑁 � 𝜇𝜇 𝐵𝐵 + 𝜇𝜇 𝐵𝐵 _ 𝐻𝐻 𝐺𝐺 𝐺𝐺𝐺𝐺𝐺𝐺 × 𝐺𝐺𝐺𝐺 𝐺𝐺 𝐺𝐺 𝑝𝑝 ( 𝑝𝑝 ), 𝜎𝜎 2
𝐵𝐵 � �

(6)

𝐵𝐵 ( 𝑝𝑝 ) ~
1
1 + exp �− 𝑁𝑁 � 𝜇𝜇 𝐵𝐵 + 𝜇𝜇 𝐵𝐵 _ 𝐻𝐻 𝐺𝐺 𝐺𝐺𝐺𝐺𝐺𝐺 × 𝐺𝐺𝐺𝐺 𝐺𝐺 𝐺𝐺 𝑝𝑝 ( 𝑝𝑝 ), 𝜎𝜎 2
𝐵𝐵 � �

(7)

𝑞𝑞 ( 𝑝𝑝 ) ~
1
1 + exp �− 𝑁𝑁 � 𝜇𝜇 𝑞𝑞 + 𝜇𝜇 𝑞𝑞 _ 𝐻𝐻 𝐺𝐺 𝐺𝐺𝐺𝐺𝐺𝐺 × 𝐺𝐺𝐺𝐺 𝐺𝐺 𝐺𝐺 𝑝𝑝 ( 𝑝𝑝 ), 𝜎𝜎 2
𝑞𝑞 � �

(8)
𝜀𝜀 ( 𝑝𝑝 )~ 𝑁𝑁 (0, 𝜎𝜎 𝜀𝜀 2
)
(9)
where p corresponds to the participant number. Group is a P (total number of p) by 1 vector of 0
and 1 where the first group is indexed as 0 and the second group is indexed as 1. Therefore, 𝜇𝜇 of
each parameter for the first and second groups are 𝜇𝜇 and 𝜇𝜇 + 𝜇𝜇 𝐻𝐻 𝐺𝐺 𝐺𝐺 𝐺𝐺𝐺𝐺
, respectively. The
participant-specific priors were not specified. Instead, the actual priors were selected by
sampling hyper-parameters for population distribution. Hyper-parameters allow us to infer a
weighted composite of the free parameters from each group. We chose uninformative priors for
the hyper-parameters so that the posterior distributions were primarily influenced by the data.
The prior distribution for the precision (an inverse of the variance) used a broad gamma
distribution. The prior distribution of the precision for the measurement noise 𝜀𝜀 was also sampled
in non-informative gamma distribution. The sampling distributions were defined as follows.
𝜇𝜇 𝐵𝐵 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝐵𝐵 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝑞𝑞 ~ 𝑁𝑁 (0, 10
3
)
(10)
𝜇𝜇 _ 𝐻𝐻 𝐺𝐺 𝐺𝐺𝐺𝐺𝐺𝐺 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝐵𝐵 _ 𝐻𝐻 𝐺𝐺 𝐺𝐺𝐺𝐺𝐺𝐺 ~ 𝑁𝑁 (0, 10
3
), 𝜇𝜇 𝑞𝑞 _ 𝐻𝐻 𝐺𝐺 𝐺𝐺𝐺𝐺𝐺𝐺 ~ 𝑁𝑁 (0, 10
3
)
(11)

141
𝜎𝜎 2
𝐵𝐵 ~ 𝜎𝜎 2
𝐵𝐵 ~ 𝜎𝜎 2
𝑞𝑞 ~ 1/Γ(10
− 3
, 10
− 3
)
(12)

𝜎𝜎 𝜀𝜀 2
( 𝑝𝑝 ) ~ 1/Γ(10
−3
, 10
−3
)
(13)
MCMC sampling for the hierarchical Bayesian state-space model was implemented in
Rjags (v4-10). We estimated parameters using 3 chains with 20000 iterations in each chain.
Initially, MCMC algorithms searched for the right sampling space for parameters. During this
period, the sampling space rapidly changes. We set this period as the first 10000 iterations, or
burn-in, so that we did not include this exploratory period in our final parameter sampling.
Moreover, one concern of MCMC sampling is that the subsequent sampling can have a high
correlation. The goal of MCMC sampling is to obtain as many independent samples as possible.
However, if subsequent samples have high correlations, the overall chain of samples do not
represent independent information about the parameter distribution. Therefore, we sampled every
10
th
iteration in our final sample. For each chain, a random initialization function selected initial
parameters (Figure A-1B). The algorithm calculated a posterior distribution of the parameter
estimate per participant p for A, B, and q (Figure A-1B-E). Because we assumed that A was the
same between two groups, the parameters of interest for this simulation were B and q. Thus, we
only presented the results from B and q. Overall, we used every 10
th
MCMC sample from 10,000
post-burn-in sample in each chain for three chains, totaling 3000 samples, to represent the
posterior distribution of the model parameters B and q.
The initial convergence of the model was tested using 𝑅𝑅 �
, which is visualized by trace
plots (Figure A-2A). 𝑅𝑅 �
tests whether the posterior distribution of each chain was similar to the
distribution of the pooled chains and is computed as follows (Gelman et al. 2013; Vehtari et al.
2021). Let Xm1, …, Xmn be the m
th
chain from M as the number of chains and N as the number of

142
samples. Let 𝑋𝑋 �
𝑚𝑚 ∙
be the mean of the m
th
chain and 𝑋𝑋 �
∙ ∙
be the mean over all chains. First, define
within-chain variance denoted as W.
𝑊𝑊 =
1
𝑀𝑀 � 𝑒𝑒 𝑚𝑚 2
𝑀𝑀 𝑚𝑚 = 1
, 𝑒𝑒 ℎ 𝑒𝑒 𝐺𝐺 𝑒𝑒
(14)
𝑒𝑒 𝑚𝑚 2
=
1
𝑁𝑁 − 1
� ( 𝑋𝑋 �
𝑚𝑚𝑚𝑚
− 𝑋𝑋 �
𝑚𝑚 ∙
)
2
𝐼𝐼 𝑚𝑚 = 1
.
(15)
Define between-chain variance denoted as B, and weighted mean of W and B denoted as 𝑉𝑉 �
.
𝐵𝐵 =
𝑁𝑁 𝑀𝑀 − 1
� ( 𝑋𝑋 �
𝑚𝑚 ∙
− 𝑋𝑋 �
∙ ∙
)
2
𝑀𝑀 𝑚𝑚 = 1

(16)
𝑉𝑉 �
= �
𝑁𝑁 − 1
𝑁𝑁 � 𝑊𝑊 + �
𝑀𝑀 + 1
𝑀𝑀 𝑁𝑁 � 𝐵𝐵 .
(17)
Finally, define 𝑅𝑅 �
.

𝑅𝑅 �
=
�
𝑉𝑉 �
𝑊𝑊
(18)
Therefore, if the distribution of each chain were the same as the distribution of the pooled chains,
𝑅𝑅 �
would be 1. Conventionally, 𝑅𝑅 �
> 1.1 is considered a failure of convergence. If any of the
parameters failed to converge, we used autojags to automatically update the model until the
model reached 𝑅𝑅 �
< 1.1. The final convergence was tested using rank-normalized 𝑅𝑅 �
(Figure A-
2B). Similar to 𝑅𝑅 �
, rank-normalized 𝑅𝑅 �
provides information about whether each MCMC chain
explores a similar sampling space (Vehtari et al. 2021). Traditional 𝑅𝑅 �
fails to detect convergence
failure when variance across chains is different or the chain has a heavy tail. Rank-normalized 𝑅𝑅 �

solves these limitations and provide more accurate convergence diagnostics. It is computed using
equations for 𝑅𝑅 �
, but replacing the parameter Xmn with rank-normalized values denoted as zmn.

143
First, Xmn is replaced by rmn, which is the rank of Xmn within the pooled samples from all chains.
Then, rmn is transformed to a normal score using the inverse normal transformation and a
fractional offset.
𝑧𝑧 𝑚𝑚𝑚𝑚
= Φ
− 1
(
𝐺𝐺 𝑚𝑚𝑚𝑚
− 3/8
𝑆𝑆 − 1/4
)
(19)
Ideally, sampling of all the chains should occur in the same space. Therefore, the normalized
number of sampling in the solution space should be identical across all chains, which makes a
uniform distribution. Consequently, if rank-normalized 𝑅𝑅 �
sampling distribution for all chains is
uniformly distributed, the model is considered to be converged.

Figure A-2. Example convergence results. (A) Trace plot of all chains that visualize 𝑹𝑹 �
for
learning rate. Each color represents a chain. The overlap across the chains and the lack of a
systematic change in in learning rate as a function of iterations are indicative of 𝑹𝑹 �
close to 1.
(B) Rank plots that visualize rank-normalized 𝑹𝑹 �
for learning rate. Each color represents a
chain. All chains show uniform distributions across the ranks, indicating convergence.

Model comparisons

144
We conducted model comparisons with four models that contained a varying number of
free parameters to find the optimal model (Table A-1). Model 1 was the simplest state-space
model that consisted of learning rate and measurement noise for each group. Model 2 included
learning rate and subject-specific measurement noise instead of the group-based measurement
noise. Model 3 added interference with group-based measurement noise. Model 4, finally, had
learning rate, interference, and subject-specific measurement noise. The model selection was
based on the deviance information criterion (DIC), which measures the goodness-of-fit that
penalizes the effective number of parameters (Spiegelhalter et al. 2002). The preferred model
was the one with the lowest DIC. If the difference in DIC between models is higher than 10, this
difference is considered a strongly important difference. Therefore, if the difference in DIC
between models was not strong (< 10), we selected a simpler model (Spiegelhalter et al. 2002).
Goodness-of-fit was further examined with the root mean squared error (RMSE).
Table A-1. Description of computational models tested. p: participant number, g: group number
Name Description and update equation
Parameters of
interest
Model 1 Learning process with group-based measurement noise 𝐵𝐵 ( 𝑝𝑝 ), 𝜀𝜀 ( 𝑔𝑔 ( 𝑝𝑝 ))
𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 + 1)
𝑇𝑇 = 𝐴𝐴 ∙ 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 )
𝑇𝑇 + 𝐵𝐵 ( 𝑝𝑝 ) ∙ 𝑒𝑒 ( 𝑝𝑝 , 𝑛𝑛 )
𝑦𝑦 ( 𝑝𝑝 , 𝑛𝑛 ) = 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵 𝐵𝐵𝐵𝐵 ( 𝑝𝑝 ) + 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 ) + 𝜀𝜀 ( 𝑔𝑔 ( 𝑝𝑝 ))

Model 2 Learning process with subject-specific measurement noise 𝐵𝐵 ( 𝑝𝑝 ), 𝜀𝜀 ( 𝑝𝑝 )
𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 + 1)
𝑇𝑇 = 𝐴𝐴 ( 𝑝𝑝 ) ∙ 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 )
𝑇𝑇 + 𝐵𝐵 ( 𝑝𝑝 ) ∙ 𝑒𝑒 ( 𝑝𝑝 , 𝑛𝑛 )
𝑦𝑦 ( 𝑝𝑝 , 𝑛𝑛 ) = 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵 𝐵𝐵𝐵𝐵 ( 𝑝𝑝 ) + 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 ) + 𝜀𝜀 ( 𝑝𝑝 )

Model 3 Learning process with interference and group-based
measurement noise
𝐵𝐵 ( 𝑝𝑝 ), 𝑞𝑞 ( 𝑝𝑝 ), 𝜀𝜀 ( 𝑔𝑔 ( 𝑝𝑝 ))
𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 + 1)
𝑇𝑇 = 𝐴𝐴 ( 𝑝𝑝 ) ∙ 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 )
𝑇𝑇 + 𝐵𝐵 ( 𝑝𝑝 ) ∙ 𝑒𝑒 ( 𝑝𝑝 , 𝑛𝑛 ) ∙ 𝒄𝒄 ( 𝑝𝑝 )
𝑦𝑦 ( 𝑝𝑝 , 𝑛𝑛 ) = 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵 𝐵𝐵𝐵𝐵 ( 𝑝𝑝 ) + 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 ) ∙ 𝒄𝒄 ( 𝑝𝑝 ) + 𝜀𝜀 ( 𝑔𝑔 ( 𝑝𝑝 ))

145
Model 4 Learning process with interference and subject-specific
measurement noise
𝐵𝐵 ( 𝑝𝑝 ), 𝑞𝑞 ( 𝑝𝑝 ), 𝜀𝜀 ( 𝑝𝑝 )
𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 + 1)
𝑇𝑇 = 𝐴𝐴 ( 𝑝𝑝 ) ∙ 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 )
𝑇𝑇 + 𝐵𝐵 ( 𝑝𝑝 ) ∙ 𝑒𝑒 ( 𝑝𝑝 , 𝑛𝑛 ) ∙ 𝒄𝒄 ( 𝑝𝑝 )
𝑦𝑦 ( 𝑝𝑝 , 𝑛𝑛 ) = 𝐹𝐹𝐹𝐹
𝐵𝐵 𝐵𝐵 𝐵𝐵𝐵𝐵 ( 𝑝𝑝 ) + 𝒙𝒙 ( 𝑝𝑝 , 𝑛𝑛 ) ∙ 𝒄𝒄 ( 𝑝𝑝 ) + 𝜀𝜀 ( 𝑝𝑝 )

After selecting the optimal model, we compared the results from the hierarchical
Bayesian state-space model with MCMC and the true simulation values with parameter estimates
using the least squares algorithms. Here, we estimated parameters of the same artificial data in
Matlab (R2019B) using fmincon where the objective was to minimize the sum of square errors
between the observed and estimated output. We added two additional parameters 𝑥𝑥 𝐿𝐿𝐿𝐿 𝐿𝐿 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 , and
𝑥𝑥 𝐻𝐻 𝐻𝐻 𝐻𝐻𝐻𝐻 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 to account for the initial deviation of foot clearance from the baseline foot clearance.
Therefore, the model estimated A, B, q, 𝑥𝑥 𝐿𝐿𝐿𝐿 𝐿𝐿 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 , and 𝑥𝑥 𝐻𝐻 𝐻𝐻 𝐻𝐻𝐻𝐻 _ 𝐻𝐻 𝐼𝐼 𝐻𝐻 𝑇𝑇 . We produced ten random
initial parameter values within the boundary of each parameter using a random number generator
to prevent from finding local minima.

Results
Model comparisons
The summary of model comparisons is in Table A-2. Parameters for all models
converged to 𝑅𝑅 �
less than 1.1 and the rank-normalized 𝑅𝑅 �
were uniformly distributed. The most
complex model with learning rate, interference, and subject-specific measurement noise (Model
4) had the lowest DIC and the difference in DIC between Model 4 and the model with the second
lowest DIC (Model 2) was larger than 10. RMSE for all models was similar. Therefore, we
selected Model 4 as our best model.

146
Table A-2. Summary of model comparisons.
Model B for PD B for Control q for PD q for Control DIC RMSE
1 0.32 ± 0.17 0.11 ± 0.03 - - -11333.4
0.034
2 0.31 ± 0.15 0.11 ± 0.03 - - -12088.3 0.034
3 0.24 ± 0.08 0.08 ± 0.02 0.25 ± 0.17 0.25 ± 0.14 -11531.5
0.032
4 0.24 ± 0.03 0.07 ± 0.01 0.28 ± 0.16 0.28 ± 0.12 -12852.5 0.032

Learning rate
The posterior distribution of learning rate captured the true simulation learning rate for
most of the simulated data sets (Figure A-3C). When learning rate was low (Group 2), both the
MCMC and least squares algorithms (fmincon) accurately estimated true learning rate. However,
when learning rate was high (Group 1), the learning rate estimated by both approaches tended to
diverge from the true value. However, this divergence was more prominent using the least
squares algorithms. The disagreement in learning rate estimation between MCMC and least
squares algorithms appears when there is greater uncertainty in the posterior probability
distribution (Figure A-3A). Note that MCMC methods demonstrated wider posterior
distributions across artificial participants in Group 1 than Group 2, indicating that the learning
rate estimation was less certain in Group 1 than Group 2 (Figure A-3A-B). Nonetheless, learning
rate using the MCMC estimated true learning rate more accurately than using the least squares
algorithm in both groups.

147

Figure A-3. Posterior distributions of learning rate. Learning rate density plots for (A) Group
1 and (B) Group 2. The red dashed line represents true learning rate for each participant, and
the gray dashed line represents estimated learning rate using the least squares algorithms
(fmincon). The black curve represents the posterior distribution from the MCMC. (C)
Difference between true and estimated learning rate by MCMC (Gray dots) and least squares

148
(empty black dots) algorithms. Each dot represents one participant. For MCMC, 89% highest
density intervals (HDI) are plotted as error bars.

Interference
The posterior distribution of interference had wider distributions (Figure A-4A-B) than
that of learning rate, indicating that this parameter was more difficult to estimate. Both MCMC
and least squares algorithms underestimated true interference (Figure A-4C). Although the
estimation of interference using either the MCMC or least squares algorithms differed from true
interference, model comparisons revealed that adding the interference parameter was preferable.
Therefore, interpretation of interference should be with caution.

149

Figure A-4. Posterior distribution of interference. Interference density plots for (A) Group 1
and (B) Group 2. (C) Difference between true and estimated interference. The description of
the figure is the same as Figure A-3.

150
Discussion
We examined if hierarchical Bayesian estimation accurately estimated free parameters of
the state-space model using simulated foot clearance data compared to parameter estimation
using least squares algorithm. Model comparison results revealed that the state-space model
including forgetting, learning rate, and interference with participant-specific measurement error
was the best model compared to the without interference or participant-specific measurement
error. We found that Bayesian estimation accurately estimated simulated learning rate better than
least squares algorithm did. Especially, Bayesian estimation accurately estimated learning rate
even for participants who had lower initial performance error, which the least squares algorithm
did not estimate well. The individual difference in initial performance error is a common
confounding factor in the analysis of motor learning. Therefore, Bayesian estimation would be
beneficial to estimate learning rate regardless of the difference in initial performance during
motor learning. However, we found that either Bayesian estimation or least squares algorithm
estimated interference well. Bayesian estimation demonstrated that there was a large uncertainty
in interference estimation. In part, this may be due to the small number of trials to estimate
interference accurately. Therefore, although the state-space model should include interference as
a free parameter in analyzing this locomotor learning task, the result of interference should be
interpreted with caution. Overall, Bayesian estimation was superior to least squares algorithm in
estimating learning rate from the state-space model for locomotor learning.

151
JAGS script in R for hierarchical Bayesian state-space model
[01] # Bayesian state-space model for motor learning
[02] model <- function() {
[03]
[04] # Set up the state space model ----------------------------------------------------------------
[05]
[06] for (s in 1:NSubjects) {
[07]
[08] # Update the states to reach to the success target
[09] for (n in 2:TrialEnd) {
[10]
[11] xmu_Low[n, s] <- (A[s] * xmu_Low[n-1, s]) +
[12] (B[s] * e[n-1, s] * C_Low[n-1, s])
[13] xmu_High[n, s] <- (A[s] * xmu_High[n-1, s]) +
[14] (B[s] * e[n-1 ,s] * C_High[n-1 ,s])
[15] }
[16]
[17] for (n in 1:TrialEnd) {
[18]
[19] # Switch interference depending on the height of the obstacle.
[20] # For example, if n is with a low obstacle (height = 0.05), set the
[21] # interference index as [1 q]
T
.
[22] C_Low[n, s] <- ifelse(Obheight[n, s] == 0.05, 1, q[s])
[23] C_High[n, s] <- ifelse(Obheight[n, s] == 0.18, 1, q[s])
[24]
[25] # Separating the distribution and calculation of foot clearance allows NaN
[26] # values in the foot clearance data for collision by treating NaN value to be
[27] # estimated by posterior distribution
[28]
[29] # Foot clearance calculation
[30] yhat[n, s] <- Base_FC[n, s] + xmu_Low[n, s] * C_Low[n, s] +
[31] xmu_High[n, s] * C_High[n, s]
[32] y[n, s] ~ dnorm(yhat[n, s], epsilonprec[s])
[33]
[34] # Calculate error based on the height of obstacles
[35] e[n, s] <- Perturbation[n, s] - y[n, s]
[36]
[37] }
[38]
[39] # Set up prior distributions ----------------------------------------------------------------
[40] # Initialize states based on the baseline data
[41] xmu_Low[1, s] ~ dnorm(xmu_init_Low[s] * Base_Data_Low[s], 1); T(0,)
[42] xmu_High[1, s] ~ dnorm(xmu_init_High[s] * Base_Data_High[s], 1); T(0,)
[43]
[44] # Change A,B,q distribution into the logistic space
[45] logit(A[s]) <- A1[s] # Logit of A1
[46] # Add a group variable
[47] A1[s] ~ dnorm(A1mu + (A1_group * Group[s]), A1prec)
[48]
[49] logit(B[s]) <- B1[s] # Logit of B1

152
[50] B1[s] ~ dnorm(B1mu + (B1_group * Group[s], B1prec)
[51]
[52] logit(q[s]) <- q1[s] # Logit of q1
[53] q1[s] ~ dnorm(q1mu + (q1_group * Group[s], q1prec)
[54]
[55] # Prior for initializing states
[56] xmu_init_Low[s] ~ dnorm(0, 1.0E-3); T(0,)
[57] xmu_init_Low[s] ~ dnorm(0, 1.0E-3); T(0,)
[58]
[59] # Prior for measurement noise as precision
[60] epsilonprec[s] ~ dgamma(1.0E-3, 1.0E-3)
[61]
[62] }
[63]
[64] # Hyper-parameters for A, B, q
[65]
[66] A1mu ~ dnorm(0.0, 1.0E-3) # mean of A1
[67] A1prec ~ dgamma(1.0E-3, 1.0E-3) # precision of A1
[68]
[69] B1mu ~ dnorm(0.0, 1.0E-3) # mean of B1
[70] B1prec ~ dgamma(1.0E-3, 1.0E-3) # precision of B1
[71]
[72] q1mu ~ dnorm(0.0, 1.0E-3) # mean of q1
[73] q1prec ~ dgamma(1.0E-3, 1.0E-3) # precision of q1
[74]
[75] # Priors for group variables
[76] A1_group ~ dnorm(0.0, 1.0E-3)
[77] B1_group ~ dnorm(0.0, 1.0E-3)
[78] q1_group ~ dnorm(0.0, 1.0E-3)
[79] }

153
APPENDIX B
DEVELOPMENT OF A NOVEL LOCOMOTOR LEARNING TASK FOR
REINFORCEMENT LEARNING
Introduction
Learning dynamic walking skills relevant to community walking is a vital element of
rehabilitation. When we learn new skills, we often learn from the difference between what we
intended to achieve and what we actually achieved. This deviation from the expectation to the
outcome provides valuable error signals to us, so that we change our behavior to achieve the
intended skills. This process requires repetitive practice and this is a hallmark of motor learning
(Dayan and Cohen 2011; Schmidt and Lee 2011). Among various processes of motor learning,
this particular process is referred to reinforcement learning (Lohse et al. 2019). The process of
reinforcement learning occurs learning a value function through feedback in a form of reward or
punishment feedback (Haith and Krakauer 2013; Sutton and Barto 1998; Wolpert et al. 2001).
The value function reflects a sum of future rewards that is expected to gain over a long term,
instead of immediate gain, while minimizing punishments (Dayan and Balleine 2002) by
exploring the space of possible actions and consider the history of actions that leads to successful
outcome. Reinforcement learning can be a useful strategy to train dynamic locomotor skills in
the clinics.
The dopaminergic pathway connecting the ventral striatum and the anterior cingulate
cortex is thought to be the neural mechanism of reinforcement learning with positive rewards
(Haber 2011). The dopamine system in the ventral striatum conveys the reinforcement learning
signal to the anterior cingulate cortex (Holroyd and Yeung 2012). An increase in the phasic
dopamine signal following rewarded trials facilitates neuroplasticity and learning that gradually
modifies behavior to achieve more rewards (Wise 2004). Due to the critical role of dopamine in
reinforcement learning, Parkinson disease (PD), which results in depletion of dopamine in the

154
basal ganglia, is an ideal model to investigate reinforcement learning. Moreover, people with PD
often show diminished motivation (Ferrazzoli et al. 2016), which is a component of the reward
system (Cools 2006). This can lead to less effective motor learning for rehabilitation (Marinelli
et al. 2017). Therefore, understanding how the process of reinforcement learning is achieved
during locomotor learning is important to facilitate strategies in gait rehabilitation.
Despite the usefulness of the reinforcement learning strategy for locomotor learning in
the clinical setting, there is only limited research investigating locomotor learning using a
reinforcement learning approach. One study examined locomotor learning using a novel
reinforcement learning approach (Hasson et al. 2015). They provided one of five levels of
categorical feedbacks that indicated positional error between the participant’s ankle angle and the
desired ankle angle. These five categories consisted of “Very Close”, “Close”, “Fair”, “Far
Away”, “Very Far Away” based on pre-determined ankle angles. However, they did not provide
information about the direction of the error. If participants were successful, they received
fictitious monetary reward for each successful trial. They found that young adults could learn a
new gait pattern using reinforcement learning. Although this study provides a useful starting
position to examine reinforcement learning, the visual feedback used in the study was not
conventional, binary reward feedback. The difficulty in creating a reinforcement learning task
during locomotion, relative to upper extremity tasks, is that walking is a continuous task and
typically do not have explicit goals or targets to achieve. Therefore, there is a need to develop an
explicit, goal-based locomotor task to investigate reinforcement learning during locomotion.
A recent study introduced a new, adaptive reinforcement learning approach for a reaching
to a target task (Therrien et al. 2016). This approach introduced a gradual visuomotor rotation
based on participants’ previous reaches. The inner bound of the reward zone was determined by

155
the ‘new’ rotation angle, which was calculated by the moving average of the last 10 reach angles,
or a desired angle if the moving average was greater than the desired angle. Participants received
binary feedback (success or failure) about their movement. If the moving average of reach angle
was within the new rotation angle and the outer bound of the reward zone, the target turned into
green, indicating success. Otherwise, the target disappeared, indicating failure. They found that
healthy adults gradually adapted to the new reach angle, and successfully reached to the desired
reward zone. The adaptation, then, was retained about 85% on the next day. Learning with only
binary feedback can often take considerable time to figure out what the ‘right’ strategy is.
However, this approach shapes and accelerates the learning behavior in a laboratory setting, so
that we can observe the process of reinforcement learning for locomotor learning in a relatively
short timeframe.
Therefore, we performed a pilot study to develop novel precision-based obstacle
negotiation task for reinforcement learning. In Experiment 1, we prescribed desired foot height
as a “reward zone” based on the participant’s baseline performance. Participants received binary
performance feedback about success or failure. In Experiment 2, we used an “adaptive reward
zone” so that participants gradually change their performance to obtain successful trials based on
binary feedback. We performed two experiments to compare the improvement of the task
performance via different reinforcement learning approaches and how these approaches
influenced time-delayed retention or savings.
Methods
Experiment 1
We recruited eight healthy young adults (29 ± 3 years, 4 F) at the University of Southern
California (USC). The Institutional Review Board at USC approved study procedures, and all

156
participants provided written, informed consent before testing began. All aspects of the study
conformed to the principles described in the Declaration of Helsinki.

Experimental Protocol and Procedure
The schematic experimental protocol is in Figure B-1A. On Day 1, participants
performed an obstacle negotiation task in a virtual environment on a treadmill (Bertec Fully
Instrumented Treadmill, USA) at their self-selected walking speed (Figure B-1B-D). Participants
viewed the virtual environment and their interaction with the environment in a head-mounted
display (Vive, HTC, Taiwan) simulated using Vizard (WorldViz, USA). During the baseline
block (BASE), participants stepped over virtual obstacles with the instruction to step over the
obstacle as naturally as possible and avoid collisions. If a collision occurred, participants heard
an unpleasant sound. Based on their natural foot clearance during BASE, we prescribed a
“reward zone” in which the lower and upper threshold of the reward zone was 8 and 12 cm
higher than the mean foot clearance during BASE, respectively. Participants were instructed that
there was a desired range of foot clearance that is invisible to them, and they needed to find the
reward zone and maintain their foot clearance once they found it (REWARD). They were also
instructed that if their foot clearance was within the reward zone, they would hear pleasant
sound. However, if their foot clearance was outside of the reward zone, they would not hear any
sound feedback. We provided the same collision feedback as BASE trials. Each block of
REWARD contained 30 obstacles, and participants performed 6 REWARD blocks. The second
to last block was composed of No Feedback (NFB) trials in which participants received collision
auditory feedback only to test short-term retention. NFB trial data of one participant (Participant
1) was lost due to technical difficulty. On the next day (Day 2), after 24 hours, participants

157
revisited the laboratory to complete retention tests. Participants performed the obstacle
negotiation task without reinforcement feedback. The reward zone was the same as Day 1.

Data Acquisition and Analysis
10-camera Qualisys Oqus motion capture system (Qualisys AB, Sweden) captured
position data from reflective markers that were placed on participants’ feet: the second toe and
the heel at the same height as the toe at 100 Hz. Qualisys streamed marker position data to
Vizard and this was used to create a visual representation of the lower extremity (Figure B-1C).
Foot clearance, defined as the minimum foot distance from obstacle’s height, was calculated in
real-time during experiments.
We used the calculated foot clearance to compute performance error offline in MATLAB
R2019b (The MathWorks, USA). We calculated performance error as the difference between
foot clearance and the upper threshold of the reward zone. Positive error denoted that foot
clearance was higher than the reward zone, and negative error greater than -0.04 m denoted that
foot clearance was lower than the reward zone.
Outcome measures were the amount of learning calculate by the difference in
performance error (PE) between BASE and RET, and success rate during REWARD.
Performance error during BASE and RET was computed using the last and first five trials,
respectively, to capture the final response to the obstacle crossing task. Success rate was
computed as the total number of successful trials divided by the total number of obstacles.

158

Figure B-1. Experimental protocol and setup. (A) Experimental protocol and design. On Day
1, participants performed baseline trials (BASE) with the instruction to step over virtual
obstacles as naturally as possible. Following BASE, participants received an instruction that
there was an invisible “reward zone” that they needed to find based on a binary sound
feedback. The lower threshold of the reward zone was 8 cm higher than mean foot clearance
during BASE. The upper threshold was 4 cm higher than the lower threshold. Participants
performed a total of 6 practice blocks. The second last block on Day 1 was no feedback trials
(NFB) to test a short-term retention. During NFB, there was no sound feedback provided.
After 24 hours (Day 2), participants revisited the lab and performed retention trials (RET).
Participants did not receive sound feedback during RET. (B) Virtual environment. (C) Lower
extremity representation. (D) Experimental setup. Participants walked on a treadmill while
wearing a head-mounted display and holding the handrails lightly.

Experiment 2
We recruited ten people with PD (66 ± 12 years, 7 F) from the neurology clinic at USC.
Inclusion criteria were 1) a Montreal Cognitive Assessment (MoCA) score of 19 or above, which
indicates normal or mild cognitive impairment, 2) ability to provide informed consent, 3)
confirmed diagnosis of PD based on UK Brain Bank criteria, and 4) Hoehn and Yahr (H&Y)

159
stage 1 to 3. Exclusion criteria for individuals with PD were 1) other neurological,
cardiovascular, orthopedic, and psychiatric diagnoses, 2) L-dopa induced hallucinations, and 3)
freezing of gait. Study procedures were approved by the Institutional Review Board at USC, and
all participants provided written, informed consent before testing began. All aspects of the study
conformed to the principles described in the Declaration of Helsinki.

Experimental Protocol and Procedure
In Experiment 1, we identified possible task variables that led to difficulty obtaining
successful trials consistently after 180 trials. These variables included 1) the low number of
obstacles per block and 2) the narrow width of the reward zone. The second source is critical in
reinforcement learning, particularly in people with PD, who have high variability in
spatiotemporal gait parameters (Peterson and Horak 2016). Here, we increased the number of
obstacles for each block from 30 to 40 obstacles and increased the width of the reward zone from
4 to 6 cm. Moreover, a pilot study of Experiment 2 revealed that the 8 cm increase relative to
baseline performance as the lower threshold of reward zone was too difficult for people with PD.
Therefore, we reduced the distance from the mean foot clearance during BASE to the lower
threshold of the reward zone from 8 to 6 cm.
A schematic of the experimental protocol is shown in Figure B-2A-B. Seven participants
with PD (65 ± 10, 4 F) performed the reinforcement learning task (RL group), and three
participants with PD (66 ± 17, 3 F) served as a control group (Control group). The difference
between the two groups was that the RL group performed the reinforcement learning
(REWARD) task while the Control group was instructed to step over obstacles as naturally as

160
possible throughout the experiment (Control). The Control group allowed us to test if changes in
foot clearance in the RL group were due to a spontaneous return to their baseline foot clearance.
On Day 1, participants performed an obstacle negotiation task in a virtual environment on
a treadmill (Bertec Fully Instrumented Treadmill, USA) at their self-selected walking speed
(Figure B-B-D). We adopted an adaptive reward-based learning protocol from a previous study
(Therrien et al. 2016). During the baseline block (BASE), both groups stepped over virtual
obstacles with an instruction to step over the obstacle as naturally as possible and avoid
collisions. If a collision occurred, participants heard an unpleasant sound. During pre-training
(PRE1-FB), we instructed the both groups to minimize the vertical distance between their foot
and the top of the obstacle during crossing, similar to our previous locomotor learning study
(Kim et al. 2019). When participants’ foot clearance was 4 cm and lower, participants heard a
pleasant sound, indicating success. When participants’ foot clearance was higher than 4 cm,
participants heard a monotone sound in which the frequency of the sound was scaled to their foot
clearance. We also provided the same collision feedback as BASE trials. If participants stepped
over at least 80% of obstacles without collision, they performed the same task without sound
feedback (PRE1-NFB). This block ensured that all participants began either REWARD or
Control trials with similar foot clearance. Following PRE1-NFB, participants in the RL group
were instructed that the reward zone has changed, and they needed to find the reward zone and
maintained their foot clearance once they found it (REWARD1). They received binary auditory
feedback about success and failure. If their foot clearance was within the reward zone, they heard
the same pleasant sound as PRE1-FB trials. However, if their foot clearance was outside of the
reward zone, they did not hear any sound feedback. We provided the same collision feedback as
BASE trials. The desired lower and upper threshold of the reward zone was 6 and 12 cm higher

161
than the mean PRE1-NFB, respectively. The average of the last ten successful trials during
PRE1-NFB was used to set the reward zone. The lower threshold of the reward zone started at
the same foot clearance as the average of the last ten successful trials during PRE1-NFB. This
lower threshold gradually increased as participants increased foot clearance to the reward zone.
Once the average of the last ten successful trials became the desired lower threshold of the
reward zone, the reward zone did not change. For Control group, we instructed participants to
step over obstacles as naturally as possible, but to avoid collision (Control 1). Both REWARD
and Control trials consisted of a total of 200 obstacles in 50-trial bouts.
On the next day (Day 2), after approximately 24 hours, participants revisited the
laboratory to complete savings trials. Both groups first performed pre-training (PRE2-FB and
PRE2-NFB), which was the same as PRE1-FB and PRE1-NFB trials on Day 1. Then,
participants in the RL group performed reinforcement learning trials (REWARD2), which were
the same as REWARD1 on Day 1. The reward zone was the same as Day 1. Whereas, the
Control group was instructed to step over obstacles as naturally as possible during savings trials
(Control 2).
During REWARD trials, the lower threshold of the reward zone was dependent on the
participants’ previous trials. The initial lower threshold was calculated as the mean of the last ten
successful trials of PRE1-NFB. Following that, the lower threshold was calculated as the moving
average of participants’ previous ten successful trials or 6 cm higher than the mean of PRE1-
NFB if the moving average was greater than this (Figure B-2A). If the total number of successful
trials was less than ten trials, the lower threshold was calculated as the moving average of
previous successful trials.
Data Acquisition and Analysis

162
Data acquisition and calculations of foot clearance were the same as Experiment 1.
Performance error was calculated as the difference between foot clearance and the upper
threshold of the reward zone. We used the upper threshold here because the lower threshold of
the reward zone changed over the course of locomotor learning. Outcome measures were savings
calculated as the difference in performance error of the first five trials during REWARD1 and 2,
the differences in success rate of REWARD1 and 2, and the number of trials performed to obtain
the desired reward zone for each day. Success rate was computed as the total number of
successful trials divided by the total number of obstacles.

163

Figure B-2. Experimental protocol and raw data on Day 1 for an example participant. (A) RL
group protocol. (B) Control group protocol. (C) An example foot clearance data of a single
participant with the adaptive reward zone on Day 1.

Results
Experiment 1
On Day 1, all participants decreased the magnitude of performance error during
REWARD 6 compared to performance error during REWARD 1 by 0.10 ± 0.04 m (mean and
standard deviation) (Table B-1, Figure B-3). However, success rate during practice trials were
low (mean and SD: 0.28 ± 0.13) and most participants did not consistently perform successful
trials at the end of practice despite practicing stepping over 180 obstacles (
Figure B-4. (A) Success rate for each trials. Each data point represents each participant.
(B) Mean performance error for the last five trials during BASE and for the first five trials during
RET. Each color represents one participant. The dotted horizontal lines represent the lower and
upper threshold of the reward zone.. Participant 8 was much less successful than others and this
participant also had the largest average and highest variability during REWARD 6, even though
they reduced performance error from REWARD 1.

164
Table B-1. Summary statistics. Mean and standard deviation of performance error in meters and
success rate in a percentage during reinforcement learning trials for each participant. Success rate
was calculated only for practice trials. Bold texts indicate average performance error within the
reward zone.
Participant
ID
Last 5 trials
during BASE
First 5 trials during
REWARD 1
Last 5 trials during
REWARD 6
First 5 trials
during RET
Success
rate
1 -0.14 ± 0.02 -0.14 ± 0.05 -0.02 ± 0.02 0 ± 0.02 0.24
2 -0.15 ± 0.03 -0.09 ± 0.10 -0.04 ± 0.03 -0.05 ± 0.02 0.34
3 -0.11 ± 0.01 -0.12 ± 0.03 0.01 ± 0.02 0.01 ± 0.04 0.46
4 -0.07 ± 0.03 -0.19 ± 0.10 -0.06 ± 0.06 -0.08 ± 0.07 0.24
5 -0.11 ± 0.03 -0.11 ± 0.04 0 ± 0.08 -0.05 ± 0.05 0.21
6 -0.13 ± 0.02 -0.15 ± 0.05 -0.02 ± 0.02 0.07 ± 0.06 0.23
7 -0.12 ± 0.02 -0.06 ± 0.06 -0.02 ± 0.02 -0.12 ± 0.02 0.42
8 -0.12 ± 0.03 -0.12 ± 0.10 -0.10 ± 0.09 -0.14 ± 0.03 0.07

165

Figure B-3. Performance error as a function of trials. The dotted horizontal lines indicate the
reward zone. Blue dots indicate trials within the reward zone, and gray dots indicate trials that
were not within the reward zone. Each subplot represents one participant. The order of the
trials was as follows: Baseline (BASE), reinforcement learning practice (REWARD), and
retention (RET). RET was performed after 24 hours.
Although five of eight participants reduced absolute performance error during RET
compared to BASE, only one participant successfully maintained their foot clearance within the
reward zone (Table B-1,
Figure B-4B). Three participants did not change performance during RET compared to BASE.

166

Figure B-4. (A) Success rate for each trials. Each data point represents each participant. (B)
Mean performance error for the last five trials during BASE and for the first five trials during
RET. Each color represents one participant. The dotted horizontal lines represent the lower
and upper threshold of the reward zone.
Experiment 2
On Day 1, a majority of the participants in both groups reduced their foot clearance as
instructed from PRE1-FB to PRE1-NFB (Figure B-5). Participants in the RL group changed foot
clearance from the first five trials during PRE1-FB (0.10 ± 0.02 m) to the last five trials during
PRE1-NFB (0.06 ± 0.06 m) and Control group changed from 0.07 ± 0.01 to 0.05 ± 0.03 m.
However, one participant in each group (first panel in the RL group and third panel in the
Control group of Figure B-5) did not reduce foot clearance during PRE1. During REWARD, all
participants increased the lower boundary of the reward zone to the desired reward zone of 6 cm
higher than the mean foot clearance during PRE1-NFB. Their foot clearance increased from the
first five trials (0.06 ± 0.02 m) to the last five trials (0.12 ± 0.07 m) (Figure B-6A). Participants
in the Control group maintained the foot clearance from the initial trials (0.07 ± 0.03 m) to the
final trials (0.08 ± 0.05 m) (Figure B-6A).

167

Figure B-5. Individual foot clearance during reinforcement learning and control trials. Trials
for the RL group were baseline (BASE), pre-training (PRE), and reward trials (REWARD) on
Day 1 and PRE and REWARD on Day 2. Trials for the Control group were BASE, PRE, and
trials without reward feedback (Control) on Day 1 and PRE and Control on Day 2. The gray
dotted line in PRE trials indicate a change in trials from PRE-FB to PRE-NFB. In the

168

On Day 2, participants in the RL group maintained their foot clearance during PRE2,
similar to PRE1 on Day 1 (Δ (PRE2-NFB – PRE1-NFB): 0 ± 0.02 m for RL and 0 ± 0.04 m).
Participants in the RL group reached the reward zone with fewer trials (Figure B-6B) and
achieved a greater number of successful trials than Day 1 (Figure B-6C). Foot clearance in the
RL group increased from the first five trials (0.08 ± 0.04 m) to the last five trials (0.10 ± 0.05 m)
(Figure B-6A). Importantly, participants in the RL group had higher foot clearance during the
first block of REWARD on Day 2 compared to Day 1, indicating a saving from practice on Day
2. Participants in the Control group maintained their foot clearance on Day 2 similar to Day 1
(initial: 0.06 ± 0.03 m, final: 0.05 ± 0.02 m).

REWARD group figure, the thick black line indicates the changes in reward zones. Blue dots
represent successful trials. Each subplot containing Day 1 and Day 2 represents one
participant.

169

Figure B-6. (A) Mean foot clearance for each block. (B) A number of trials required to adapt
to the desired lower threshold of the reward zone. Each data point represents each participant.
Dotted lines indicate more trials to achieved the desired lower threshold on Day 2 than Day 1.
C) Success rate in each block of reward trials. Each data point represents each participant.

Discussion
For Experiment 1, we developed a novel locomotor learning task to investigate
reinforcement learning. We showed that the baseline performance-based “reward zone” ensured
that all participants had enough room to learn the task. We also demonstrated that the majority of
the participants could reduce their performance error via a binary success or failure sound
feedback during a precision-based obstacle negotiation learning task. However, most participants
failed to consistently obtain successful trials and demonstrated fluctuations in motor performance
at the end of practice. One explanation for the large variability in performance is that the reward
zone we prescribed may be too narrow that it was not feasible for some participants to perform

170
successful trials consistently. However, when we calculated a reward zone of the same size
centered on the mean performance from BASE, mean of 41.25% of BASE trials fall within this
reward zone. This indicates that given their natural variability and width of the reward zone, the
ceiling of their potential success rate was 41.25%. The mean success rate at the end of our
learning task on Day 1 was 33.33%, which was lower than the potential success rate. Therefore,
the width of the reward zone may not be the source of the variable performance. Another
explanation may be that the amount of practice (30 obstacles per block) was not sufficient for
participants to maximize their performance, which we further addressed in Experiment 2.
Overall, although we have demonstrated the feasibility of implementing a reinforcement learning
based obstacle negotiation task, further fine-tuning of task variables such as the reward zone
width and the number of obstacles per block should be considered.
For Experiment 2, we aimed to address the limitations of our previous reinforcement
learning task and develop an appropriate reinforcement learning task for locomotion in people
with PD. In general, most participants gradually reached the reward zone on Day 1 and they
could achieve more successful trials and reached the reward zone with fewer obstacles on Day 2.
However, we found that even with an adaptive reward zone, people with PD showed difficulty
learning precision-based obstacle negotiation using binary reinforcement feedback on Day 1, as
their foot clearance reached the desired reward zone after the third or fourth block on average.
Our experiments had a lower number of trials than previous studies. We had 200 trials on Day 1
while Hasson and colleagues had 250 trials for their reinforcement learning during locomotion
(Hasson et al. 2015) and Therrian and colleagues had 320 trials for their adaptive reinforcement
learning during reaching (Therrien et al. 2016). This suggests that there was indeed learning

171
during our task, but the amount of practices may still be too low to capture this learning on Day
1.
Moreover, the increase in foot clearance with practice was not due to a spontaneous
return to participant’s baseline performance. Participants performed pre-training to reduce their
foot clearance so that all participants had similar foot clearance at the beginning of REWARD or
Control. This may lead to a return to natural foot clearance spontaneously when the instruction to
minimize their foot clearance was removed, resulting to reaching to the prescribed reward zone.
Our results demonstrated that the increase in foot clearance by the REWARD group was not a
result from the spontaneous return to their baseline behavior as reduced foot clearance in the
Control group after PRE trials were maintained throughout the trials when the instruction was
removed. Overall, an adaptive reinforcement learning task for locomotion can be a useful task to
understand the underlying mechanisms of reinforcement learning during locomotion. However,
further research is necessary to find an appropriate number of trial repetitions for people with PD
to learn and consistently perform this motor skill within a single day.
Questions for future research
• What is the role of variability in reinforcement learning during locomotion? Does
variability during reinforcement-based locomotor learning represent motor
exploration?
• What is the trade-off between effort and reward during reinforcement-based locomotor
learning? How does fatigue influence this effort-based reinforcement learning during
locomotion?
• How does reward and punishment differentially affect locomotor learning?
• What are the individual characteristics of learners and non-learners of reinforcement-
based locomotor learning?
In summary, we demonstrated that our precision-based obstacle negotiation task can be
acquired via binary feedback based on reinforcement learning. Moreover, we showed that the
number of practice trials improved consolidation of the locomotor task after a time delay.

172
However, there was a number of individuals who did not learn the task. Future studies can refine
the task to promote locomotor skill acquisition and retention via reinforcement learning. The
development of locomotor tasks to investigate reinforcement learning can broaden our
understanding on the process of reinforcement learning during locomotion and its effectiveness
in neurorehabilitation.

173
REFERENCES
Aarsland D, Bronnick K, Williams-Gray C, Weintraub D, Marder K, Kulisevsky J, Burn D,
Barone P, Pagonabarraga J, Allcock L, Santangelo G, Foltynie T, Janvin C, Larsen JP, Barker
RA, Emre M. Mild cognitive impairment in Parkinson disease: A multicenter pooled analysis.
Neurology 75: 1062–1069, 2010.
Abernethy B. Dual-task methodology and motor skills research: Some applications and
methodological constraints. J Hum Mov Stud , 1988.
Abrahamse E, Ruitenberg M, De Kleine E, Verwey WB. Control of automated behavior: insights
from the discrete sequence production task. Front Hum Neurosci 7, 2013.
Ahmed AA, Wolpert DM. Transfer of Dynamic Learning Across Postures. J Neurophysiol 102:
2816–2824, 2009.
Albert ST, Shadmehr R. Estimating properties of the fast and slow adaptive processes during
sensorimotor adaptation. J Neurophysiol 119: 1367–1393, 2018.
Albin RL, Dauer WT. Parkinson syndrome. Heterogeneity of etiology; heterogeneity of
pathogenesis? Neurology 79: 202–203, 2012.
Albouy G, Sterpenich V, Vandewalle G, Darsaud A, Gais S, Rauchs G, Desseilles M, Boly M,
Dang-Vu T, Balteau E, Degueldre C, Phillips C, Luxen A, Maquet P. Interaction between
Hippocampal and Striatal Systems Predicts Subsequent Consolidation of Motor Sequence
Memory. PLOS ONE 8: e59490, 2013.
Alexander GE, Crutcher MD, DeLong MR. Basal ganglia-thalamocortical circuits: parallel
substrates for motor, oculomotor, “prefrontal” and “limbic” functions. Prog Brain Res 85: 119–
146, 1990.
Alexander GE, DeLong MR, Strick PL. Parallel organization of functionally segregated circuits
linking basal ganglia and cortex. Annu Rev Neurosci 9: 357–381, 1986.
Amir A Mohagheghi RM. The effects of distant and on-line visual information on the control of
approach phase and step over an obstacle during locomotion. Exp Brain Res 155: 459–68, 2004.
Anderkova L, Barton M, Rektorova I. Striato-cortical connections in Parkinson’s and
Alzheimer’s diseases: Relation to cognition. Mov Disord 32: 917–922, 2017.
Anglin J, Saldana D, Schmiesing A, Liew S. Transfer of a skilled motor learning task between
virtual and conventional environments. In: IEEE Virtual Reality. 2017, p. 401–402.
Arias P, Robles-García V, Sanmartín G, Flores J, Cudeiro J. Virtual Reality as a Tool for
Evaluation of Repetitive Rhythmic Movements in the Elderly and Parkinson’s Disease Patients.
PLOS ONE 7: e30021, 2012.

174
Ashburn A, Stack E, Pickering RM, Ward CD. A community-dwelling sample of people with
Parkinson’s disease: characteristics of fallers and non-fallers. Age Ageing 30: 47–52, 2001.
Asjad NS, Adams H, Paris R, Bodenheimer B. Perception of Height in Virtual Reality: A Study
of Climbing Stairs. In: Proceedings of the 15th ACM Symposium on Applied Perception. ACM,
p. 4:1-4:8.
Bastian AJ. Understanding sensorimotor adaptation and learning for rehabilitation. Curr Opin
Neurol 21: 628–633, 2008.
Behrman AL, Cauraugh JH, Light KE. Practice as an intervention to improve speeded motor
performance and motor learning in Parkinson’s disease. J Neurol Sci 174: 127–136, 2000.
Benedict RHB. Brief Visuospatial Memory Test–Revised
TM
. Lutz, FL: Psychological
Assessment Resources, Inc1997.
Berman BD, Smucny J, Wylie KP, Shelton E, Kronberg E, Leehey M, Tregellas JR. Levodopa
modulates small-world architecture of functional brain networks in Parkinson’s disease. Mov
Disord Off J Mov Disord Soc 31: 1676–1684, 2016.
Biswal B, Yetkin FZ, Haughton VM, Hyde JS. Functional connectivity in the motor cortex of
resting human brain using echo-planar MRI. Magn Reson Med 34: 537–541, 1995.
Bloem BR, Grimbergen YA, Cramer M, Willemsen M, Zwinderman AH. Prospective
assessment of falls in Parkinson’s disease. J Neurol 248: 950–958, 2001.
Bobrownicki R, MacPherson AC, Coleman SGS, Collins D, Sproule J. Re-examining the effects
of verbal instructional type on early stage motor learning. Hum Mov Sci 44: 168–181, 2015.
Bodenheimer B, Fu Q. The Effect of Avatar Model in Stepping off a Ledge in an Immersive
Virtual Environment. In: Proceedings of the ACM SIGGRAPH Symposium on Applied
Perception. ACM, p. 115–118.
Bodenheimer B, Meng J, Wu H, Narasimham G, Rump B, McNamara TP, Carr TH, Rieser JJ.
Distance Estimation in Virtual and Real Environments Using Bisection. In: Proceedings of the
4th Symposium on Applied Perception in Graphics and Visualization. ACM, p. 35–40.
Bonzano L, Palmaro E, Teodorescu R, Fleysher L, Inglese M, Bove M. Functional connectivity
in the resting-state motor networks influences the kinematic processes during motor sequence
learning. Eur J Neurosci 41: 243–253, 2015.
Bostan AC, Strick PL. The basal ganglia and the cerebellum: nodes in an integrated network. Nat
Rev Neurosci 19: 338–350, 2018.
Brockett AT, Tennyson SS, deBettencourt CA, Gaye F, Roesch MR. Anterior cingulate cortex is
necessary for adaptation of action plans. Proc Natl Acad Sci 117: 6196–6204, 2020.

175
Browne WJ, Draper D. A comparison of Bayesian and likelihood-based methods for fitting
multilevel models. Bayesian Anal 1: 473–514, 2006.
Burdet E, Osu R, Franklin DW, Milner TE, Kawato M. The central nervous system stabilizes
unstable dynamics by learning optimal impedance. Nature 414: 446–449, 2001.
Cabeza R. Cognitive neuroscience of aging: contributions of functional neuroimaging. Scand J
Psychol 42: 277–286, 2001.
Cacciola A, Calamuneri A, Milardi D, Mormina E, Chillemi G, Marino S, Naro A, Rizzo G,
Anastasi G, Quartarone A. A Connectomic Analysis of the Human Basal Ganglia Network.
Front Neuroanat 11, 2017.
Canning CG, Allen NE, Nackaerts E, Paul SS, Nieuwboer A, Gilat M. Virtual reality in research
and rehabilitation of gait and balance in Parkinson disease. Nat Rev Neurol , 2020.
doi:10.1038/s41582-020-0370-2.
Cassady K, Ruitenberg M, Koppelmans V, Reuter-Lorenz P, De Dios Y, Gadd N, Wood S,
Riascos Castenada R, Kofman I, Bloomberg J, Mulavara A, Seidler R. Neural predictors of
sensorimotor adaptation rate and savings. Hum Brain Mapp 39: 1516–1531, 2018.
Ceballos-Baumann AO. Functional imaging in Parkinson’s disease: activation studies with PET,
fMRI and SPECT. J Neurol 250 Suppl 1: I15-23, 2003.
Chase C, Seidler R. Degree of handedness affects intermanual transfer of skill learning. Exp
Brain Res 190: 317–328, 2008.
Chaudhuri KR, Healy DG, Schapira AH. Non-motor symptoms of Parkinson’s disease: diagnosis
and management. Lancet Neurol 5: 235–245, 2006.
Chen HC, Ashton-Miller JA, Alexander NB, Schultz AB. Stepping over obstacles: gait patterns
of healthy young and old adults. J Gerontol 46: M196-203, 1991.
Chib S, Greenberg E. Understanding the Metropolis-Hastings Algorithm. Am Stat 49: 327–335,
1995.
Chien JH, Post J, Siu K-C. Effects of Aging on the Obstacle Negotiation Strategy while Stepping
over Multiple Obstacles. Sci Rep 8: 8576, 2018.
Choi JT, Bastian AJ. Adaptation reveals independent control networks for human walking. Nat
Neurosci 10: 1055–1062, 2007.
Choi Y, Shin EY, Kim S. Spatiotemporal dissociation of fMRI activity in the caudate nucleus
underlies human de novo motor skill learning. Proc Natl Acad Sci 117: 23886–23897, 2020.
Chou L-S, Draganich LF. Placing the trailing foot closer to an obstacle reduces flexion of the
hip, knee, and ankle to increase the risk of tripping. J Biomech 31: 685–691, 1998a.

176
Chou L-S, Draganich LF. Increasing obstacle height and decreasing toe-obstacle distance affect
the joint moments of the stance limb differently when stepping over an obstacle. Gait Posture 8:
186–204, 1998b.
Cohen J. Statistical Power Analysis for the Behavioral Sciences. Routledge.
Cole DM, Smith SM, Beckmann CF. Advances and Pitfalls in the Analysis and Interpretation of
Resting-State FMRI Data. Front Syst Neurosci 4, 2010.
Contreras-Vidal JL, Stelmach GE. A neural model of basal ganglia-thalamocortical relations in
normal and parkinsonian movement. Biol Cybern 73: 467–476, 1995.
Cools R. Dopaminergic modulation of cognitive function-implications for L-DOPA treatment in
Parkinson’s disease. Neurosci Biobehav Rev 30: 1–23, 2006.
Cothros N, Wong JD, Gribble PL. Are there distinct neural representations of object and limb
dynamics? Exp Brain Res 173: 689–697, 2006.
Crosbie J, Ko V. Changes in the temporal and distance parameters of gait evoked by negotiation
of curbs. Aust J Physiother 46: 103–112, 2000.
Cullen KE, Brooks JX. Neural correlates of sensory prediction errors in monkeys: evidence for
internal models of voluntary self-motion in the cerebellum. Cerebellum Lond Engl 14: 31–34,
2015.
Cumming G. The new statistics: why and how. Psychol Sci 25: 7–29, 2014.
D’Ardenne K, Eshel N, Luka J, Lenartowicz A, Nystrom LE, Cohen JD. Role of prefrontal
cortex and the midbrain dopamine system in working memory updating. Proc Natl Acad Sci U S
A 109: 19900–19909, 2012.
Davidsdottir S, Cronin-Golomb A, Lee A. Visual and spatial symptoms in Parkinson’s disease.
Vision Res 45: 1285–1296, 2005.
Day KA, Leech KA, Roemmich RT, Bastian AJ. Accelerating locomotor savings in learning:
compressing four training days to one. J Neurophysiol 119: 2100–2113, 2018.
Dayan E, Cohen LG. Neuroplasticity subserving motor skill learning. Neuron 72: 443–454,
2011.
Dayan P, Balleine BW. Reward, Motivation, and Reinforcement Learning. Neuron 36: 285–298,
2002.
Delis DC, Kaplan E, Kramer JH. Delis-Kaplan Executive Function System: Examiner’s Manual.
San Antionio, TX: The Psychological Corporation2001.
Delis DC, Kramer JH, Kaplan E, Ober BA. Manual for the California Verbal Learning Test,
(CVLT-II). San Antonio, TX: The Psychological Corporation2000.

177
Della-Maggiore V, Malfait N, Ostry DJ, Paus T. Stimulation of the posterior parietal cortex
interferes with arm trajectory adjustments during the learning of new dynamics. J Neurosci Off J
Soc Neurosci 24: 9971–9976, 2004.
Della-Maggiore V, McIntosh AR. Time course of changes in brain activity and functional
connectivity associated with long-term adaptation to a rotational transformation. J Neurophysiol
93: 2254–2262, 2005.
Deroost N, Kerckhofs E, Coene M, Wijnants G, Soetens E. Learning sequence movements in a
homogenous sample of patients with Parkinson’s disease. Neuropsychologia 44: 1653–1662,
2006.
Destrebecqz A, Peigneux P, Laureys S, Degueldre C, Del Fiore G, Aerts J, Luxen A, Van Der
Linden M, Cleeremans A, Maquet P. The neural correlates of implicit and explicit sequence
learning: Interacting networks revealed by the process dissociation procedure. Learn Mem 12:
480–490, 2005.
Di Martino A, Scheres A, Margulies DS, Kelly AMC, Uddin LQ, Shehzad Z, Biswal B, Walters
JR, Castellanos FX, Milham MP. Functional connectivity of human striatum: a resting state
FMRI study. Cereb Cortex N Y N 1991 18: 2735–2747, 2008.
Di X, Gohel S, Kim EH, Biswal BB. Task vs. rest—different network configurations between
the coactivation and the resting-state brain networks. Front Hum Neurosci 7, 2013.
Dietz V, Zijlstra W, Duysens J. Human neuronal interlimb coordination during split-belt
locomotion. Exp Brain Res 101: 513–520, 1994.
Ding L, Gold JI. The basal ganglia’s contributions to perceptual decision making. Neuron 79:
640–649, 2013.
Dominey PF, Boussaoud D. Encoding behavioral context in recurrent networks of the fronto-
striatal system: a simulation study. Brain Res Cogn Brain Res 6: 53–65, 1997.
Doyon J, Bellec P, Amsel R, Penhune V, Monchi O, Carrier J, Lehéricy S, Benali H.
Contributions of the basal ganglia and functionally related brain structures to motor learning.
Behav Brain Res 199: 61–75, 2009.
Doyon J, Benali H. Reorganization and plasticity in the adult brain during learning of motor
skills. Curr Opin Neurobiol 15: 161–167, 2005.
Doyon J, Penhune V, Ungerleider LG. Distinct contribution of the cortico-striatal and cortico-
cerebellar systems to motor skill learning. Neuropsychologia 41: 252–262, 2003.
Doyon J, Song AW, Karni A, Lalonde F, Adams MM, Ungerleider LG. Experience-dependent
changes in cerebellar contributions to motor sequence learning. Proc Natl Acad Sci 99: 1017–
1022, 2002.

178
Du Y, Clark JE. New insights into statistical learning and chunk learning in implicit sequence
acquisition. Psychon Bull Rev 24: 1225–1233, 2017.
Du Y, Clark JE. The “Motor” in Implicit Motor Sequence Learning: A Foot-stepping Serial
Reaction Time Task. J Vis Exp JoVE , 2018. doi:10.3791/56483.
Duarte JE, Reinkensmeyer DJ. Effects of robotically modulating kinematic variability on motor
skill learning and motivation. J Neurophysiol 113: 2682–2691, 2015.
Ehringer H, Hornykiewicz O. Distribution of noradrenaline and dopamine (3-hydroxytyramine)
in the human brain and their behavior in diseases of the extrapyramidal system1This article was
originally published in German in the journal Klinische Wochenschrift, volume 38, issue 24, 15
December 1960, pp. 1236–1239.1. Parkinsonism Relat Disord 4: 53–57, 1998.
Elfmarková N, Gajdoš M, Mračková M, Mekyska J, Mikl M, Rektorová I. Impact of Parkinson’s
disease and levodopa on resting state functional connectivity related to speech prosody control.
Parkinsonism Relat Disord 22 Suppl 1: S52-55, 2016.
Emken JL, Benitez R, Reinkensmeyer DJ. Human-robot cooperative movement training:
Learning a novel sensory motor transformation during walking with robotic assistance-as-
needed. J NeuroEngineering Rehabil 4: 8, 2007.
Erni T, Dietz V. Obstacle avoidance during human walking: learning rate and cross-modal
transfer. J Physiol 534: 303–312, 2001.
Erro R, Vitale C, Amboni M, Picillo M, Moccia M, Longo K, Santangelo G, Rosa AD, Allocca
R, Giordano F, Orefice G, Michele GD, Santoro L, Pellecchia MT, Barone P. The Heterogeneity
of Early Parkinson’s Disease: A Cluster Analysis on Newly Diagnosed Untreated Patients. PLOS
ONE 8: e70244, 2013.
Faiman I, Pizzamiglio S, Turner DL. Resting-state functional connectivity predicts the ability to
adapt arm reaching in a robot-mediated force field. NeuroImage 174: 494–503, 2018.
Ferrazzoli D, Carter A, Ustun FS, Palamara G, Ortelli P, Maestri R, Yücel M, Frazzitta G.
Dopamine Replacement Therapy, Learning and Reward Prediction in Parkinson’s Disease:
Implications for Rehabilitation. Front Behav Neurosci 10, 2016.
Ferrazzoli D, Ortelli P, Madeo G, Giladi N, Petzinger GM, Frazzitta G. Basal ganglia and
beyond: The interplay between motor and cognitive aspects in Parkinson’s disease rehabilitation.
Neurosci Biobehav Rev 90: 294–308, 2018.
Fisher BE, Li Q, Nacca A, Salem GJ, Song J, Yip J, Hui JS, Jakowec MW, Petzinger GM.
Treadmill exercise elevates striatal dopamine D2 receptor binding potential in patients with early
Parkinson’s disease: NeuroReport 24: 509–514, 2013.
Fitts PM. Perceptual-Motor Skill Learning. In: Categories of Human Learning, edited by Melton
AW. Academic Press, p. 243–285.

179
Fitts PM, Posner MI. Human performance. Brooks/Cole Publishing Company, 1967.
Floyer-Lea A, Matthews PM. Changing Brain Networks for Visuomotor Control With Increased
Movement Automaticity. J Neurophysiol 92: 2405–2412, 2004.
Fortin K, Blanchette A, McFadyen BJ, Bouyer LJ. Effects of walking in a force field for varying
durations on aftereffects and on next day performance. Exp Brain Res 199: 145, 2009.
Fox SH. Non-dopaminergic treatments for motor control in Parkinson’s disease. Drugs 73:
1405–1415, 2013.
Franchak JM, Adolph KE. Visually guided navigation: head-mounted eye-tracking of natural
locomotion in children and adults. Vision Res 50: 2766–2774, 2010.
Frank MJ, Loughry B, O’Reilly RC. Interactions between frontal cortex and basal ganglia in
working memory: a computational model. Cogn Affect Behav Neurosci 1: 137–160, 2001.
Franklin DW, So U, Burdet E, Kawato M. Visual Feedback Is Not Necessary for the Learning of
Novel Dynamics. PLOS ONE 2: e1336, 2007.
Franzen MD. Reliability and Validity in Neuropsychological Assessment. Springer Science &
Business Media, 2000.
Fung J, Richards CL, Malouin F, McFadyen BJ, Lamontagne A. A treadmill and motion coupled
virtual reality system for gait training post-stroke. Cyberpsychology Behav Impact Internet
Multimed Virtual Real Behav Soc 9: 157–162, 2006.
Gelman A, Carlin JB, Stern HS, Dunson DB, Vehtari A, Rubin DB. Bayesian Data Analysis. 3rd
edition. Boca Raton: Chapman and Hall/CRC, 2013.
Gelman A, Rubin DB. Inference from Iterative Simulation Using Multiple Sequences. Stat Sci 7:
457–472, 1992.
Gersch TM, Kowler E, Dosher B. Dynamic allocation of visual attention during the execution of
sequences of saccades. Vision Res 44: 1469–1483, 2004.
Ghanbarzadeh R, Ghapanchi AH, Blumenstein M, Talaei-Khoei A. A Decade of Research on the
Use of Three-Dimensional Virtual Worlds in Health Care: A Systematic Literature Review. J
Med Internet Res 16, 2014.
Ghilardi M-F, Eidelberg D, Silvestri G, Ghez C. The differential effect of PD and normal aging
on early explicit sequence learning. Neurology 60: 1313–1319, 2003.
Ghilardi MF, Moisello C, Silvestri G, Ghez C, Krakauer JW. Learning of a sequential motor skill
comprises explicit and implicit components that consolidate differently. J Neurophysiol 101:
2218–2229, 2009.

180
Gobron SC, Zannini N, Wenk N, Schmitt C, Charrotton Y, Fauquex A, Lauria M, Degache F,
Frischknecht R. Serious Games for Rehabilitation Using Head-Mounted Display and Haptic
Devices. In: Augmented and Virtual Reality. International Conference on Augmented and Virtual
Reality. Springer, Cham, p. 199–219.
Goetz CG, Tilley BC, Shaftman SR, Stebbins GT, Fahn S, Martinez-Martin P, Poewe W,
Sampaio C, Stern MB, Dodel R, Dubois B, Holloway R, Jankovic J, Kulisevsky J, Lang AE,
Lees A, Leurgans S, LeWitt PA, Nyenhuis D, Olanow CW, Rascol O, Schrag A, Teresi JA, van
Hilten JJ, LaPelle N. Movement Disorder Society-sponsored revision of the Unified Parkinson’s
Disease Rating Scale (MDS-UPDRS): Scale presentation and clinimetric testing results. Mov
Disord 23: 2129–2170, 2008.
Goldman JG, Vernaleo BA, Camicioli R, Dahodwala N, Dobkin RD, Ellis T, Galvin JE, Marras
C, Edwards J, Fields J, Golden R, Karlawish J, Levin B, Shulman L, Smith G, Tangney C,
Thomas CA, Tröster AI, Uc EY, Coyan N, Ellman C, Ellman M, Hoffman C, Hoffman S,
Simmonds D. Cognitive impairment in Parkinson’s disease: a report from a multidisciplinary
symposium on unmet needs and future directions to maintain cognitive health. Npj Park Dis 4:
1–11, 2018.
Gordon KE, Ferris DP. Learning to walk with a robotic ankle exoskeleton. J Biomech 40: 2636–
2644, 2007.
Gorges M, Müller H-P, Lulé D, Pinkhardt EH, Ludolph AC, Kassubek J. To rise and to fall:
functional connectivity in cognitively normal and cognitively impaired patients with Parkinson’s
disease. Neurobiol Aging 36: 1727–1735, 2015.
Grabli D, Karachi C, Welter M-L, Lau B, Hirsch EC, Vidailhet M, François C. Normal and
pathological gait: what we learn from Parkinson’s disease. J Neurol Neurosurg Psychiatry jnnp-
2012-302263, 2012.
Graci V, Elliott DB, Buckley JG. Utility of peripheral visual cues in planning and controlling
adaptive gait. Optom Vis Sci Off Publ Am Acad Optom 87: 21–27, 2010.
Graff-Radford J, Williams L, Jones DT, Benarroch EE. Caudate nucleus as a component of
networks controlling behavior. Neurology 89: 2192–2197, 2017.
Gray P, Hildebrand K. Fall risk factors in Parkinson’s disease. J Neurosci Nurs J Am Assoc
Neurosci Nurses 32: 222–228, 2000.
Greenland JC, Williams‐Gray CH, Barker RA. The clinical heterogeneity of Parkinson’s disease
and its therapeutic implications. Eur J Neurosci 49: 328–338, 2019.
Guadagnoli MA, Lee TD. Challenge point: a framework for conceptualizing the effects of
various practice conditions in motor learning. J Mot Behav 36: 212–224, 2004.
Haber SN. Neuroanatomy of Reward: A View from the Ventral Striatum. In: Neurobiology of
Sensation and Reward, edited by Gottfried JA. Boca Raton (FL): CRC Press/Taylor & Francis,
2011.

181
Haber SN. Corticostriatal circuitry. Dialogues Clin Neurosci 18: 7–21, 2016.
Hacker CD, Perlmutter JS, Criswell SR, Ances BM, Snyder AZ. Resting state functional
connectivity of the striatum in Parkinson’s disease. Brain J Neurol 135: 3699–3711, 2012.
Haith AM, Krakauer JW. Model-Based and Model-Free Mechanisms of Human Motor Learning.
In: Progress in Motor Control, edited by Richardson MJ, Riley MA, Shockley K. New York,
NY: Springer, 2013, p. 1–21.
Ham JH, Cha J, Lee JJ, Baek G-M, Sunwoo MK, Hong JY, Shin N-Y, Sohn YH, Lee J-M, Lee
PH. Nigrostriatal dopamine-independent resting-state functional networks in Parkinson’s disease.
NeuroImage 119: 296–304, 2015.
Hamann JM, Dayan E, Hummel FC, Cohen LG. Baseline Frontostriatal-Limbic Connectivity
Predicts Reward-Based Memory Formation. Hum Brain Mapp 35: 5921–5931, 2014.
Hanajima R, Shadmehr R, Ohminami S, Tsutsumi R, Shirota Y, Shimizu T, Tanaka N, Terao Y,
Tsuji S, Ugawa Y, Uchimura M, Inoue M, Kitazawa S. Modulation of error-sensitivity during a
prism adaptation task in people with cerebellar degeneration. J Neurophysiol 114: 2460–2471,
2015.
Harms MP, Somerville LH, Ances BM, Andersson J, Barch DM, Bastiani M, Bookheimer SY,
Brown TB, Buckner RL, Burgess GC, Coalson TS, Chappell MA, Dapretto M, Douaud G, Fischl
B, Glasser MF, Greve DN, Hodge C, Jamison KW, Jbabdi S, Kandala S, Li X, Mair RW,
Mangia S, Marcus D, Mascali D, Moeller S, Nichols TE, Robinson EC, Salat DH, Smith SM,
Sotiropoulos SN, Terpstra M, Thomas KM, Tisdall MD, Ugurbil K, van der Kouwe A, Woods
RP, Zöllei L, Van Essen DC, Yacoub E. Extending the Human Connectome Project across ages:
Imaging protocols for the Lifespan Development and Aging projects. NeuroImage 183: 972–984,
2018.
Harrington DO. Visual Fields: Textbook and Atlas of Clinical Perimetry. 5th edition. St. Louis:
Mosby, 1981.
Harsay HA, Cohen MX, Oosterhof NN, Forstmann BU, Mars RB, Ridderinkhof KR. Functional
Connectivity of the Striatum Links Motivation to Action Control in Humans. J Neurosci 31:
10701–10711, 2011.
Haslinger B, Erhard P, Altenmüller E, Hennenlotter A, Schwaiger M, Gräfin von Einsiedel H,
Rummeny E, Conrad B, Ceballos-Baumann AO. Reduced recruitment of motor association areas
during bimanual coordination in concert pianists. Hum Brain Mapp 22: 206–215, 2004.
Hasson CJ, Manczurowsky J, Yen S-C. A reinforcement learning approach to gait training
improves retention. Front Hum Neurosci 9, 2015.
Hayhoe M, Ballard D. Eye movements in natural behavior. Trends Cogn Sci 9: 188–194, 2005.
Heaton RK. A Manual for the Wisconsin Card Sorting Test. 1st edition. Psychological
Assessment Resources, Inc., 1981.

182
van Hedel HJA, Dietz V. The influence of age on learning a locomotor task. Clin Neurophysiol
Off J Int Fed Clin Neurophysiol 115: 2134–2143, 2004.
van Hedel HJA, Wirth B, Dietz V. Limits of locomotor ability in subjects with a spinal cord
injury. Spinal Cord 43: 593, 2005.
Helmich RC, Derikx LC, Bakker M, Scheeringa R, Bloem BR, Toni I. Spatial Remapping of
Cortico-striatal Connectivity in Parkinson’s Disease. Cereb Cortex 20: 1175–1186, 2010.
van den Heuvel MP, Hulshoff Pol HE. Exploring the brain network: A review on resting-state
fMRI functional connectivity. Eur Neuropsychopharmacol 20: 519–534, 2010.
Hikosaka O, Nakamura K, Sakai K, Nakahara H. Central mechanisms of motor skill learning.
Curr Opin Neurobiol 12: 217–222, 2002.
Hikosaka O, Takikawa Y, Kawagoe R. Role of the Basal Ganglia in the Control of Purposive
Saccadic Eye Movements. Physiol Rev 80: 953–978, 2000.
Holiga Š, Mueller K, Möller HE, Sieger T, Schroeter ML, Vymazal J, Růžička E, Jech R. Motor
Matters: Tackling Heterogeneity of Parkinson’s Disease in Functional MRI Studies. PLOS ONE
8: e56133, 2013.
Holroyd CB, Yeung N. Motivation of extended behaviors by anterior cingulate cortex. Trends
Cogn Sci 16: 122–128, 2012.
Hooper E. Manual for the Hooper Visual Organization Test. Los Angeles: Western
Psychological Services1983.
Hoort B van der, Ehrsson HH. Illusions of having small or large invisible bodies influence visual
perception of object size. Sci Rep 6: 34530, 2016.
van der Hoort B, Guterstam A, Ehrsson HH. Being Barbie: the size of one’s own body
determines the perceived size of the world. PloS One 6: e20195, 2011.
Hoshi E, Tremblay L, Féger J, Carras PL, Strick PL. The cerebellum communicates with the
basal ganglia. Nat Neurosci 8: 1491–1493, 2005.
Hughes AJ, Daniel SE, Kilford L, Lees AJ. Accuracy of clinical diagnosis of idiopathic
Parkinson’s disease: a clinico-pathological study of 100 cases. J Neurol Neurosurg Psychiatry
55: 181–184, 1992.
Jaffe DL, Brown DA, Pierson-Carey CD, Buckley EL, Lew HL. Stepping over obstacles to
improve walking in individuals with poststroke hemiplegia. J Rehabil Res Dev 41: 283–292,
2004.
James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statistical Learning: with
Applications in R. New York: Springer-Verlag, 2013.

183
Jankovic J. Parkinson’s disease: clinical features and diagnosis. J Neurol Neurosurg Psychiatry
79: 368–376, 2008.
Jankovic J, McDermott M, Carter J, Gauthier S, Goetz C, Golbe L, Huber S, Koller W, Olanow
C, Shoulson I. Variable expression of Parkinson’s disease: a base-line analysis of the DATATOP
cohort. The Parkinson Study Group. Neurology 40: 1529–1534, 1990.
Jansen SEM, Toet A, Werkhoven PJ. Obstacle Crossing With Lower Visual Field Restriction:
Shifts in Strategy. J Mot Behav 43: 55–62, 2010.
Jenkinson M, Bannister P, Brady M, Smith S. Improved optimization for the robust and accurate
linear registration and motion correction of brain images. NeuroImage 17: 825–841, 2002.
Kalia LV, Lang AE. Parkinson’s disease. The Lancet 386: 896–912, 2015.
Kaplan E, Goodglass H, Weintraub S. Boston naming test. Philadelphia: Lea & Febiger, 1983.
Karni A, Meyer G, Rey-Hipolito C, Jezzard P, Adams MM, Turner R, Ungerleider LG. The
acquisition of skilled motor performance: Fast and slow experience-driven changes in primary
motor cortex. Proc Natl Acad Sci 95: 861–868, 1998.
Kim A, Darakjian N, Finley JM. Walking in fully immersive virtual environments: an evaluation
of potential adverse effects in older adults and individuals with Parkinson’s disease. J
NeuroEngineering Rehabil 14: 16, 2017a.
Kim A, Kretch KS, Zhou Z, Finley JM. The quality of visual information about the lower
extremities influences visuomotor coordination during virtual obstacle negotiation. J
Neurophysiol , 2018. doi:10.1152/jn.00931.2017.
Kim A, Zhou Z, Kretch KS, Finley JM. Manipulating the fidelity of lower extremity visual
feedback to identify obstacle negotiation strategies in immersive virtual reality. In: 2017 39th
Annual International Conference of the IEEE Engineering in Medicine and Biology Society
(EMBC). 2017 39th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC). 2017b, p. 4491–4494.
Kim J, Criaud M, Cho SS, Díez-Cirarda M, Mihaescu A, Coakeley S, Ghadery C, Valli M,
Jacobs MF, Houle S, Strafella AP. Abnormal intrinsic brain functional network dynamics in
Parkinson’s disease. Brain J Neurol 140: 2955–2967, 2017c.
Kim S, Oh Y, Schweighofer N. Between-Trial Forgetting Due to Interference and Time in Motor
Adaptation. PLOS ONE 10: e0142963, 2015.
Kleim JA, Jones TA. Principles of experience-dependent neural plasticity: implications for
rehabilitation after brain damage. J Speech Lang Hear Res JSLHR 51: S225-239, 2008.
Kluzik J, Diedrichsen J, Shadmehr R, Bastian AJ. Reach adaptation: what determines whether
we learn an internal model of the tool or adapt the model of our arm? J Neurophysiol 100: 1455–
1464, 2008.

184
Koshimori Y, Cho S-S, Criaud M, Christopher L, Jacobs M, Ghadery C, Coakeley S, Harris M,
Mizrahi R, Hamani C, Lang AE, Houle S, Strafella AP. Disrupted Nodal and Hub Organization
Account for Brain Network Abnormalities in Parkinson’s Disease. Front Aging Neurosci 8: 259,
2016.
Koziol LF, Budding D, Andreasen N, D’Arrigo S, Bulgheroni S, Imamizu H, Ito M, Manto M,
Marvel C, Parker K, Pezzulo G, Ramnani N, Riva D, Schmahmann J, Vandervert L, Yamazaki
T. Consensus Paper: The Cerebellum’s Role in Movement and Cognition. Cerebellum Lond Engl
13: 151–177, 2014.
Krakauer JW, Ghilardi M-F, Mentis M, Barnes A, Veytsman M, Eidelberg D, Ghez C.
Differential cortical and subcortical activations in learning rotations and gains for reaching: a
PET study. J Neurophysiol 91: 924–933, 2004.
Krakauer JW, Hadjiosif AM, Xu J, Wong AL, Haith AM. Motor Learning. In: Comprehensive
Physiology. American Cancer Society, p. 613–663.
Krakauer JW, Pine ZM, Ghilardi MF, Ghez C. Learning of visuomotor transformations for
vectorial planning of reaching trajectories. J Neurosci Off J Soc Neurosci 20: 8916–8924, 2000.
Krell J, Patla AE. The influence of multiple obstacles in the travel path on avoidance strategy.
Gait Posture 16: 15–19, 2002.
Krishnan C, Washabaugh EP, Reid CE, Althoen MM, Ranganathan R. Learning new gait
patterns: Age-related differences in skill acquisition and interlimb transfer. Exp Gerontol 111:
45–52, 2018.
Kruschke J. Doing Bayesian Data Analysis: A Tutorial with R, JAGS, and Stan. 2nd edition.
Boston: Academic Press, 2014.
Kwak Y, Peltier S, Bohnen NI, Müller MLTM, Dayalu P, Seidler RD. Altered Resting State
Cortico-Striatal Connectivity in Mild to Moderate Stage Parkinson’s Disease. Front Syst
Neurosci 4, 2010.
Landi SM, Baguear F, Della-Maggiore V. One Week of Motor Adaptation Induces Structural
Changes in Primary Motor Cortex That Predict Long-Term Memory One Year Later. J Neurosci
31: 11808–11813, 2011.
Lashley KS. The problem of serial order in behavior. In: Cerebral mechanisms in behavior; the
Hixon Symposium. Oxford, England: Wiley, 1951, p. 112–146.
Le Heron C, Plant O, Manohar S, Ang Y-S, Jackson M, Lennox G, Hu MT, Husain M. Distinct
effects of apathy and dopamine on effort-based decision-making in Parkinson’s disease. Brain
141: 1455–1469, 2018.
Lee J-Y, Schweighofer N. Dual Adaptation Supports a Parallel Architecture of Motor Memory. J
Neurosci 29: 10396–10404, 2009.

185
Lee TD, Swinnen SP, Serrien DJ. Cognitive Effort and Motor Learning. Quest 46: 328–344,
1994.
Lee Y-Y, Fisher BE. Use of low frequency repetitive transcranial magnetic stimulation to reduce
context-dependent learning in people with Parkinson’s disease. Eur J Phys Rehabil Med , 2017.
doi:10.23736/S1973-9087.17.04723-2.
Lee Y-Y, Tai C-H, Fisher BE. Context-Dependent Behavior in Parkinson’s Disease With
Freezing of Gait. Neurorehabil Neural Repair 33: 1040–1049, 2019.
Lee Y-Y, Winstein CJ, Fisher BE. Role of the dorsolateral prefrontal cortex in context-dependent
motor performance. Eur J Neurosci 43: 954–960, 2016.
Lee Y-Y, Winstein CJ, Gordon J, Petzinger GM, Zelinski EM, Fisher BE. Context-Dependent
Learning in People With Parkinson’s Disease. J Mot Behav 1–9, 2015.
Leech KA, Day KA, Roemmich RT, Bastian AJ. Movement and perception recalibrate
differently across multiple days of locomotor learning. J Neurophysiol 120: 2130–2137, 2018.
Lefumat HZ, Vercher J-L, Miall RC, Cole J, Buloup F, Bringoux L, Bourdin C, Sarlegna FR. To
transfer or not to transfer? Kinematics and laterality quotient predict interlimb transfer of motor
learning. J Neurophysiol 114: 2764–2774, 2015.
Lehéricy S, Benali H, Van de Moortele P-F, Pélégrini-Issac M, Waechter T, Ugurbil K, Doyon J.
Distinct basal ganglia territories are engaged in early and advanced motor sequence learning.
Proc Natl Acad Sci U S A 102: 12566–12571, 2005.
Lemos J, Pereira D, Almendra L, Rebelo D, Patrício M, Castelhano J, Cunha G, Januário C,
Cunha L, Freire A, Castelo-Branco M. Distinct functional properties of the vertical and
horizontal saccadic network in Health and Parkinson’s disease: An eye-tracking and fMRI study.
Brain Res 1648: 469–484, 2016.
Levy DJ, Glimcher PW. The root of all value: a neural common currency for choice. Curr Opin
Neurobiol 22: 1027–1038, 2012.
Leyrer M, Linkenauger SA, Bülthoff HH, Kloos U, Mohler B. The Influence of Eye Height and
Avatars on Egocentric Distance Estimates in Immersive Virtual Environments. In: Proceedings
of the ACM SIGGRAPH Symposium on Applied Perception in Graphics and Visualization.
ACM, p. 67–74.
Leyrer M, Linkenauger SA, Bülthoff HH, Mohler BJ. Eye Height Manipulations: A Possible
Solution to Reduce Underestimation of Egocentric Distances in Head-Mounted Displays. ACM
Trans Appl Percept 12: 1:1-1:23, 2015.
Li CS, Padoa-Schioppa C, Bizzi E. Neuronal correlates of motor performance and motor learning
in the primary motor cortex of monkeys adapting to an external force field. Neuron 30: 593–607,
2001.

186
Lin Q. People’s perception and action in immersive virtual environments (IVEs). Vanderbilt
University2014.
Lin Q, Rieser J, Bodenheimer B. Affordance Judgments in HMD-Based Virtual Environments:
Stepping over a Pole and Stepping off a Ledge. ACM Trans Appl Percept 12: 6:1-6:21, 2015.
Lingo VanGilder J, Hengge CR, Duff K, Schaefer SY. Visuospatial function predicts one-week
motor skill retention in cognitively intact older adults. Neurosci Lett 664: 139–143, 2018.
Lipton DM, Gonzales BJ, Citri A. Dorsal Striatal Circuits for Habits, Compulsions and
Addictions. Front Syst Neurosci 13, 2019.
Lohse K, Miller M, Bacelar M, Krigolson O. Errors, rewards, and reinforcement in motor skill
learning. Routledge.
Luo C, Song W, Chen Q, Zheng Z, Chen K, Cao B, Yang J, Li J, Huang X, Gong Q, Shang H-F.
Reduced functional connectivity in early-stage drug-naive Parkinson’s disease: a resting-state
fMRI study. Neurobiol Aging 35: 431–441, 2014.
Luo CY, Guo XY, Song W, Chen Q, Cao B, Yang J, Gong QY, Shang H-F. Functional
connectome assessed using graph theory in drug-naive Parkinson’s disease. J Neurol 262: 1557–
1567, 2015.
Ma L-Y, Chan P, Gu Z-Q, Li F-F, Feng T. Heterogeneity among patients with Parkinson’s
disease: Cluster analysis and genetic association. J Neurol Sci 351: 41–45, 2015.
Maagaard M, Sorensen JL, Oestergaard J, Dalsgaard T, Grantcharov TP, Ottesen BS, Larsen CR.
Retention of laparoscopic procedural skills acquired on a virtual-reality surgical trainer. Surg
Endosc 25: 722–727, 2011.
Maidan I, Eyal S, Kurz I, Geffen N, Gazit E, Ravid L, Giladi N, Mirelman A, Hausdorff JM.
Age-associated changes in obstacle negotiation strategies: Does size and timing matter? Gait
Posture 59: 242–247, 2018.
Makino H, Hwang EJ, Hedrick NG, Komiyama T. Circuit Mechanisms of Sensorimotor
Learning. Neuron 92: 705–721, 2016.
Makowski D, Ben-Shachar MS, Lüdecke D. bayestestR: Describing Effects and their
Uncertainty, Existence and Significance within the Bayesian Framework. J Open Source Softw 4:
1541, 2019.
Malik RN, Cote R, Lam T. Sensorimotor integration of vision and proprioception for obstacle
crossing in ambulatory individuals with spinal cord injury. J Neurophysiol 117: 36–46, 2017.
Malone LA, Bastian AJ, Torres-Oviedo G. How does the motor system correct for errors in time
and space during locomotor adaptation? J Neurophysiol 108: 672–683, 2012.

187
Malone LA, Vasudevan EVL, Bastian AJ. Motor adaptation training for faster relearning. J
Neurosci Off J Soc Neurosci 31: 15136–15143, 2011.
Marcos E, Nougaret S, Tsujimoto S, Genovesio A. Outcome Modulation Across Tasks in the
Primate Dorsolateral Prefrontal Cortex. Neuroscience 371: 96–105, 2018.
Marigold DS, Andujar J-E, Lajoie K, Drew T. Chapter 6--motor planning of locomotor
adaptations on the basis of vision: the role of the posterior parietal cortex. Prog Brain Res 188:
83–100, 2011.
Marigold DS, Patla AE. Visual information from the lower visual field is important for walking
across multi-surface terrain. Exp Brain Res 188: 23–31, 2008.
Marigold DS, Weerdesteyn V, Patla AE, Duysens J. Keep looking ahead? Re-direction of visual
fixation does not always occur during an unpredictable obstacle avoidance task. Exp Brain Res
176: 32–42, 2007.
Marinelli L, Crupi D, Di Rocco A, Bove M, Eidelberg D, Abbruzzese G, Ghilardi MF. Learning
and consolidation of visuo-motor adaptation in Parkinson’s disease. Parkinsonism Relat Disord
15: 6–11, 2009.
Marinelli L, Quartarone A, Hallett M, Frazzitta G, Ghilardi MF. The many facets of motor
learning and their relevance for Parkinson’s disease. Clin Neurophysiol Off J Int Fed Clin
Neurophysiol 128: 1127–1141, 2017.
Mary A, Wens V, Op de Beeck M, Leproult R, De Tiège X, Peigneux P. Resting-state Functional
Connectivity is an Age-dependent Predictor of Motor Learning Abilities. Cereb Cortex 27:
4923–4932, 2017.
Matthis JS, Barton SL, Fajen BR. The biomechanics of walking shape the use of visual
information during locomotion over complex terrain. J Vis 15, 2015.
Matthis JS, Fajen BR. Visual control of foot placement when walking over complex terrain. J
Exp Psychol Hum Percept Perform 40: 106–115, 2014a.
Matthis JS, Fajen BR. Visual control of foot placement when walking over complex terrain. J
Exp Psychol Hum Percept Perform 40: 106–115, 2014b.
McManus EA, Bodenheimer B, Streuber S, de la Rosa S, Bülthoff HH, Mohler BJ. The
Influence of Avatar (Self and Character) Animations on Distance Estimation, Object Interaction
and Locomotion in Immersive Virtual Environments. In: Proceedings of the ACM SIGGRAPH
Symposium on Applied Perception in Graphics and Visualization. ACM, p. 37–44.
McSorley E, Gilchrist ID, McCloy R. The programming of sequences of saccades. Exp Brain
Res 237: 1009–1018, 2019.
Mendes FA dos S, Pompeu JE, Lobo AM, da Silva KG, Oliveira T de P, Zomignani AP,
Piemonte MEP. Motor learning, retention and transfer after virtual-reality-based training in

188
Parkinson’s disease – effect of motor and cognitive demands of games: a longitudinal, controlled
clinical study. Physiotherapy 98: 217–223, 2012.
Meyniel F, Goodwin GM, Deakin JW, Klinge C, MacFadyen C, Milligan H, Mullings E,
Pessiglione M, Gaillard R. A specific role for serotonin in overcoming effort cost. eLife 5:
e17282, 2016.
Michel J, Benninger D, Dietz V, van Hedel HJA. Obstacle stepping in patients with Parkinson’s
disease. Complexity does influence performance. J Neurol 256: 457–463, 2009.
Mirelman A, Maidan I, Herman T, Deutsch JE, Giladi N, Hausdorff JM. Virtual reality for gait
training: can it induce motor learning to enhance complex walking and reduce fall risk in patients
with Parkinson’s disease? J Gerontol A Biol Sci Med Sci 66: 234–240, 2011.
Mirelman A, Rochester L, Maidan I, Del Din S, Alcock L, Nieuwhof F, Rikkert MO, Bloem BR,
Pelosin E, Avanzino L, Abbruzzese G, Dockx K, Bekkers E, Giladi N, Nieuwboer A, Hausdorff
JM. Addition of a non-immersive virtual reality component to treadmill training to reduce fall
risk in older adults (V-TIME): a randomised controlled trial. The Lancet , 2016.
doi:10.1016/S0140-6736(16)31325-3.
Mohler BJ, Bülthoff HH, Thompson WB, Creem-Regehr SH. A Full-body Avatar Improves
Egocentric Distance Judgments in an Immersive Virtual Environment. In: Proceedings of the 5th
Symposium on Applied Perception in Graphics and Visualization. ACM, p. 194-.
Mohler BJ, Creem-Regehr SH, Thompson WB, Bülthoff HH. The Effect of Viewing a Self-
Avatar on Distance Judgments in an HMD-Based Virtual Environment. Presence Teleoperators
Virtual Environ 19: 230–242, 2010.
Mongeon D, Blanchet P, Messier J. Impact of Parkinson’s disease and dopaminergic medication
on adaptation to explicit and implicit visuomotor perturbations. Brain Cogn 81: 271–282, 2013.
Moraes R, Patla AE. Determinants guiding alternate foot placement selection and the behavioral
responses are similar when avoiding a real or a virtual obstacle. Exp Brain Res 171: 497–510,
2006.
Morris ME. Locomotor Training in People With Parkinson Disease. Phys Ther 86: 1426–1435,
2006.
Morris ME, Martin CL, Schenkman ML. Striding Out With Parkinson Disease: Evidence-Based
Physical Therapy for Gait Disorders. Phys Ther 90: 280–288, 2010.
Morton SM, Bastian AJ. Cerebellar contributions to locomotor adaptations during splitbelt
treadmill walking. J Neurosci Off J Soc Neurosci 26: 9107–9116, 2006.
Muslimović D, Post B, Speelman JD, Schmand B. Motor procedural learning in Parkinson’s
disease. Brain 130: 2887–2897, 2007.

189
Nakamura T, Ghilardi MF, Mentis M, Dhawan V, Fukuda M, Hacking A, Moeller JR, Ghez C,
Eidelberg D. Functional networks in motor sequence learning: Abnormal topographies in
Parkinson’s disease. Hum Brain Mapp 12: 42–60, 2001.
Newell KM, Liu YT, Mayer-Kress G. Time scales in motor learning and development. Psychol
Rev 108: 57–82, 2001.
Nieuwboer A, De Weerdt W, Dom R, Truyen M, Janssens L, Kamsma Y. The effect of a home
physiotherapy program for persons with Parkinson’s disease. J Rehabil Med 33: 266–272, 2001.
Nieuwboer A, Rochester L, Müncks L, Swinnen SP. Motor learning in Parkinson’s disease:
limitations and potential for rehabilitation. Parkinsonism Relat Disord 15 Suppl 3: S53-58, 2009.
Nieuwhof F, Helmich RC. Entangled cerebral networks in Parkinson’s disease. Brain 140: 2767–
2769, 2017.
Nunes EJ, Randall PA, Podurgiel S, Correa M, Salamone JD. Nucleus accumbens
neurotransmission and effort-related choice behavior in food motivation: effects of drugs acting
on dopamine, adenosine, and muscarinic acetylcholine receptors. Neurosci Biobehav Rev 37:
2015–2025, 2013.
Olson M, Lockhart TE, Lieberman A. Motor Learning Deficits in Parkinson’s Disease (PD) and
Their Effect on Training Response in Gait and Balance: A Narrative Review. Front Neurol 10,
2019.
Onla-or S, Winstein CJ. Determining the optimal challenge point for motor skill learning in
adults with moderately severe Parkinson’s disease. Neurorehabil Neural Repair 22: 385–395,
2008.
Orban de Xivry J-J, Criscimagna-Hemminger SE, Shadmehr R. Contributions of the motor
cortex to adaptive control of reaching depend on the perturbation schedule. Cereb Cortex N Y N
1991 21: 1475–1484, 2011.
Overduin SA, Richardson AG, Bizzi E. Cortical Processing during Dynamic Motor Adaptation.
In: Progress in Motor Control: A Multidisciplinary Perspective, edited by Sternad D. Springer
US, p. 423–438.
Owens-Walton C, Jakabek D, Li X, Wilkes FA, Walterfang M, Velakoulis D, van Westen D,
Looi JCL, Hansson O. Striatal changes in Parkinson disease: An investigation of morphology,
functional connectivity and their relationship to clinical symptoms. Psychiatry Res
Neuroimaging 275: 5–13, 2018.
Parijat P, Lockhart TE, Liu J. Effects of perturbation-based slip training using a virtual reality
environment on slip-induced falls. Ann Biomed Eng 43: 958–967, 2015.
Park H, Schweighofer N. Nonlinear mixed-effects model reveals a distinction between learning
and performance in intensive reach training post-stroke. J Neuroengineering Rehabil 14: 21,
2017.

190
Park KY, Lee JJ, Dierker D, Marple LM, Hacker CD, Roland JL, Marcus DS, Milchenko M,
Miller-Thomas MM, Benzinger TL, Shimony JS, Snyder AZ, Leuthardt EC. Mapping language
function with task-based vs. resting-state functional MRI. PLOS ONE 15: e0236423, 2020.
Patla AE. How Is Human Gait Controlled by Vision. Ecol Psychol 10: 287–302, 1998.
Patla AE, Greig M. Any way you look at it, successful obstacle negotiation needs visually guided
on-line foot placement regulation during the approach phase. Neurosci Lett 397: 110–114, 2006.
Paus T. Primate anterior cingulate cortex: Where motor control, drive and cognition interface.
Nat Rev Neurosci 2: 417–424, 2001.
Pendt LK, Reuter I, Müller H. Motor Skill Learning, Retention, and Control Deficits in
Parkinson’s Disease. PLOS ONE 6: e21669, 2011.
Peterson DS, Dijkstra BW, Horak FB. Postural motor learning in people with Parkinson’s
disease. J Neurol 263: 1518–1529, 2016.
Peterson DS, Horak FB. Neural Control of Walking in People with Parkinsonism. Physiol
Bethesda Md 31: 95–107, 2016.
Petzinger GM, Fisher BE, McEwen S, Beeler JA, Walsh JP, Jakowec MW. Exercise-enhanced
neuroplasticity targeting motor and cognitive circuitry in Parkinson’s disease. Lancet Neurol 12:
716–726, 2013.
Petzinger GM, Fisher BE, Van Leeuwen J-E, Vukovic M, Akopian G, Meshul CK, Holschneider
DP, Nacca A, Walsh JP, Jakowec MW. Enhancing neuroplasticity in the basal ganglia: the role
of exercise in Parkinson’s disease. Mov Disord Off J Mov Disord Soc 25 Suppl 1: S141-145,
2010.
Phillips L, Ries B, Kaeding M, Interrante V. Avatar Self-embodiment Enhances Distance
Perception Accuracy in Non-photorealistic Immersive Virtual Environments. In: Proceedings of
the 2010 IEEE Virtual Reality Conference. IEEE Computer Society, p. 115–1148.
Poewe W, Seppi K, Tanner CM, Halliday GM, Brundin P, Volkmann J, Schrag A-E, Lang AE.
Parkinson disease. Nat Rev Dis Primer 3: 1–21, 2017.
Price A, Shin JC. The impact of Parkinson’s disease on sequence learning: perceptual pattern
learning and executive function. Brain Cogn 69: 252–261, 2009.
Pruim RHR, Mennes M, Buitelaar JK, Beckmann CF. Evaluation of ICA-AROMA and
alternative strategies for motion artifact removal in resting state fMRI. NeuroImage 112: 278–
287, 2015.
Rabe K, Livne O, Gizewski ER, Aurich V, Beck A, Timmann D, Donchin O. Adaptation to
Visuomotor Rotation and Force Field Perturbation Is Correlated to Different Brain Areas in
Patients With Cerebellar Degeneration. J Neurophysiol 101: 1961–1971, 2009.

191
van Ravenzwaaij D, Cassey P, Brown SD. A simple introduction to Markov Chain Monte–Carlo
sampling. Psychon Bull Rev 25: 143–154, 2018.
Rebelo D, Oliveira F, Abrunhosa A, Januário C, Lemos J, Castelo-Branco M. A link between
synaptic plasticity and reorganization of brain activity in Parkinson’s disease. Proc Natl Acad Sci
U S A 118, 2021.
Redding GM, Rossetti Y, Wallace B. Applications of prism adaptation: a tutorial in theory and
method. Neurosci Biobehav Rev 29: 431–444, 2005.
Redgrave P, Rodriguez M, Smith Y, Rodriguez-Oroz MC, Lehericy S, Bergman H, Agid Y,
DeLong MR, Obeso JA. Goal-directed and habitual control in the basal ganglia: implications for
Parkinson’s disease. Nat Rev Neurosci 11: 760–772, 2010.
Reinkensmeyer DJ, Patton JL. Can Robots Help the Learning of Skilled Actions? Exerc Sport
Sci Rev 37: 43–51, 2009.
Reisman DS, Block HJ, Bastian AJ. Interlimb coordination during locomotion: what can be
adapted and stored? J Neurophysiol 94: 2403–2415, 2005.
Renner RS, Velichkovsky BM, Helmert JR. The Perception of Egocentric Distances in Virtual
Environments - A Review. ACM Comput Surv 46: 23:1-23:40, 2013.
Rhea CK, Rietdyk S. Visual exteroceptive information provided during obstacle crossing did not
modify the lower limb trajectory. Neurosci Lett 418: 60–65, 2007.
Richardson AG, Overduin SA, Valero-Cabré A, Padoa-Schioppa C, Pascual-Leone A, Bizzi E,
Press DZ. Disruption of Primary Motor Cortex before Learning Impairs Memory of Movement
Dynamics. J Neurosci 26: 12466–12470, 2006.
Ries B, Interrante V, Kaeding M, Phillips L. Analyzing the Effect of a Virtual Avatar’s
Geometric and Motion Fidelity on Ego-centric Spatial Perception in Immersive Virtual
Environments. In: Proceedings of the 16th ACM Symposium on Virtual Reality Software and
Technology. ACM, p. 59–66.
Rietdyk S, Rhea CK. Control of adaptive locomotion: effect of visual obstruction and visual cues
in the environment. Exp Brain Res 169: 272–278, 2006a.
Rietdyk S, Rhea CK. Control of adaptive locomotion: effect of visual obstruction and visual cues
in the environment. Exp Brain Res 169: 272–278, 2006b.
Rietdyk S, Rhea CK. The effect of the visual characteristics of obstacles on risk of tripping and
gait parameters during locomotion. Ophthalmic Physiol Opt J Br Coll Ophthalmic Opt Optom
31: 302–310, 2011.
Rizzo A, Kim GJ. A SWOT Analysis of the Field of Virtual Reality Rehabilitation and Therapy.
Presence Teleoper Virtual Env 14: 119–146, 2005.

192
Robertson EM. The Serial Reaction Time Task: Implicit Motor Skill Learning? J Neurosci 27:
10073–10075, 2007.
Robinson JL, Laird AR, Glahn DC, Blangero J, Sanghera MK, Pessoa L, Fox PM, Uecker A,
Friehs G, Young KA, Griffin JL, Lovallo WR, Fox PT. The functional connectivity of the human
caudate: an application of meta-analytic connectivity modeling with behavioral filtering.
NeuroImage 60: 117–129, 2012.
Roemmich RT, Nocera JR, Stegemöller EL, Hassan A, Okun MS, Hass CJ. Locomotor
adaptation and locomotor adaptive learning in Parkinson’s disease and normal aging. Clin
Neurophysiol 125: 313–319, 2014.
Rosenbaum DA, Cohen RG, Jax SA, Weiss DJ, van der Wel R. The problem of serial order in
behavior: Lashley’s legacy. Hum Mov Sci 26: 525–554, 2007.
Rosenbaum DA, Kenny SB, Derr MA. Hierarchical control of rapid movement sequences. J Exp
Psychol Hum Percept Perform 9: 86–102, 1983.
Ross GW, Abbott RD. Living and dying with Parkinson’s disease. Mov Disord 29: 1571–1573,
2014.
Ruitenberg MFL, Kleine ED, Lubbe RHJV der, Verwey WB, Abrahamse EL. Context-dependent
motor skill and the role of practice. Psychol Res 76: 812–820, 2012.
Ruitenberg MFL, Koppelmans V, De Dios YE, Gadd NE, Wood SJ, Reuter-Lorenz PA, Kofman
I, Bloomberg JJ, Mulavara AP, Seidler RD. Neural correlates of multi-day learning and savings
in sensorimotor adaptation. Sci Rep 8, 2018.
Sakai K, Kitaguchi K, Hikosaka O. Chunking during human visuomotor sequence learning. Exp
Brain Res 152: 229–242, 2003.
Salamone JD, Correa M, Yang J-H, Rotolo R, Presby R. Dopamine, Effort-Based Choice, and
Behavioral Economics: Basic and Translational Research. Front Behav Neurosci 12, 2018.
Sami S, Robertson EM, Miall RC. The Time Course of Task-Specific Memory Consolidation
Effects in Resting State Networks. J Neurosci 34: 3982–3992, 2014.
Sánchez N, Simha SN, Donelan JM, Finley JM. Using asymmetry to your advantage: learning to
acquire and accept external assistance during prolonged split-belt walking. J Neurophysiol 125:
344–357, 2020.
Sang L, Zhang J, Wang L, Zhang J, Zhang Y, Li P, Wang J, Qiu M. Alteration of Brain
Functional Networks in Early-Stage Parkinson’s Disease: A Resting-State fMRI Study. PLOS
ONE 10: e0141815, 2015.
Schaefer SY, Duff K. Rapid Responsiveness to Practice Predicts Longer-Term Retention of
Upper Extremity Motor Skill in Non-Demented Older Adults. Front Aging Neurosci 7, 2015.

193
de Schipper LJ, Hafkemeijer A, van der Grond J, Marinus J, Henselmans JML, van Hilten JJ.
Altered Whole-Brain and Network-Based Functional Connectivity in Parkinson’s Disease. Front
Neurol 9: 419, 2018.
Schlerf JE, Xu J, Klemfuss NM, Griffiths TL, Ivry RB. Individuals with cerebellar degeneration
show similar adaptation deficits with large and small visuomotor errors. J Neurophysiol 109:
1164–1173, 2013.
Schmidt R, Lee T. Motor Control and Learning: A Behavioral Emphasis. 5th edition.
Champaign, IL: Human Kinetics, 2011.
Schmidt RA, Bjork RA. New conceptualizations of practice: Common principles in three
paradigms suggest new concepts for training. Psychol Sci 3: 207–217, 1992.
Seger CA, Miller EK. Category Learning in the Brain. Annu Rev Neurosci 33: 203–219, 2010.
Seidler RD. Neural Correlates of Motor Learning, Transfer of Learning, and Learning to Learn.
Exerc Sport Sci Rev 38: 3–9, 2010.
Seidler RD, Noll DC, Chintalapati P. Bilateral basal ganglia activation associated with
sensorimotor adaptation. Exp Brain Res 175: 544–555, 2006.
Seltman HJ. Experimental Design and Analysis. Carnegie Mellon University, 2012.
Shadmehr R, Mussa-Ivaldi FA. Adaptive representation of dynamics during learning of a motor
task. J Neurosci 14: 3208–3224, 1994.
Shadmehr R, Smith MA, Krakauer JW. Error correction, sensory prediction, and adaptation in
motor control. Annu Rev Neurosci 33: 89–108, 2010.
Shema SR, Brozgol M, Dorfman M, Maidan I, Sharaby-Yeshayahu L, Malik-Kozuch H,
Wachsler Yannai O, Giladi N, Hausdorff JM, Mirelman A. Clinical experience using a 5-week
treadmill training program with virtual reality to enhance gait in an ambulatory physical therapy
service. Phys Ther 94: 1319–1326, 2014.
da Silva JJ, Barbieri FA, Gobbi L TB. Adaptive locomotion for crossing a moving obstacle.
Motor Control 15: 419–433, 2011.
Siu K-C, Best BJ, Kim JW, Oleynikov D, Ritter FE. Adaptive Virtual Reality Training to
Optimize Military Medical Skills Acquisition and Retention. Mil Med 181: 214–220, 2016.
Smiley-Oyen AL, Lowry KA, Emerson QR. Learning and retention of movement sequences in
Parkinson’s disease. Mov Disord Off J Mov Disord Soc 21: 1078–1087, 2006.
Smiley‐Oyen AL, Worringham CJ, Cross CL. Practice effects in three-dimensional sequential
rapid aiming in Parkinson’s disease. Mov Disord 17: 1196–1204, 2002.

194
Smiley-Oyen AL, Worringham CJ, Cross CL. Motor learning processes in a movement-scaling
task in olivopontocerebellar atrophy and Parkinson’s disease. Exp Brain Res 152: 453–465,
2003.
Smith JD, Martin PE. Walking patterns change rapidly following asymmetrical lower extremity
loading. Hum Mov Sci 26: 412–425, 2007.
Smith MA, Ghazizadeh A, Shadmehr R. Interacting Adaptive Processes with Different
Timescales Underlie Short-Term Motor Learning. PLOS Biol 4: e179, 2006.
Smith MA, Shadmehr R. Intact ability to learn internal models of arm dynamics in Huntington’s
disease but not cerebellar degeneration. J Neurophysiol 93: 2809–2821, 2005.
Smith SM, Fox PT, Miller KL, Glahn DC, Fox PM, Mackay CE, Filippini N, Watkins KE, Toro
R, Laird AR, Beckmann CF. Correspondence of the brain’s functional architecture during
activation and rest. Proc Natl Acad Sci U S A 106: 13040–13045, 2009.
Spiegelhalter DJ, Best NG, Carlin BP, Van Der Linde A. Bayesian measures of model
complexity and fit. J R Stat Soc Ser B 64: 583–639, 2002.
Stack E, Ashburn A. Fall events described by people with Parkinson’s disease: implications for
clinical interviewing and the research agenda. Physiother Res Int J Res Clin Phys Ther 4: 190–
200, 1999.
Statton MA, Vazquez A, Morton SM, Vasudevan EVL, Bastian AJ. Making Sense of Cerebellar
Contributions to Perceptual and Motor Adaptation. Cerebellum Lond Engl 17: 111–121, 2018.
Steiger JH. Beyond the F test: Effect size confidence intervals and tests of close fit in the
analysis of variance and contrast analysis. Psychol Methods 9: 164–182, 2004.
Sternberg S, Monsell S, Knoll RL, Wright CE. 6 - The Latency and Duration of Rapid
Movement Sequences: Comparisons of Speech and Typewriting. In: Information Processing in
Motor Control and Learning, edited by Stelmach GE. Academic Press, p. 117–152.
Stillman CM, Gordon EM, Simon JR, Vaidya CJ, Howard DV, Howard JH. Caudate resting
connectivity predicts implicit probabilistic sequence learning. Brain Connect 3: 601–610, 2013.
Stolze H, Klebe S, Zechlin C, Baecker C, Friege L, Deuschl G. Falls in frequent neurological
diseases--prevalence, risk factors and aetiology. J Neurol 251: 79–84, 2004.
Sutton RS, Barto AG. Reinforcement Learning: An Introduction. second edition. Cambridge,
Mass: A Bradford Book, 1998.
Swinnen SP, Steyvers M, Van Den Bergh L, Stelmach GE. Motor learning and Parkinson’s
disease: refinement of within-limb and between-limb coordination as a result of practice. Behav
Brain Res 111: 45–59, 2000.

195
Tahmasian M, Bettray LM, van Eimeren T, Drzezga A, Timmermann L, Eickhoff CR, Eickhoff
SB, Eggers C. A systematic review on the applications of resting-state fMRI in Parkinson’s
disease: Does dopamine replacement therapy play a role? Cortex J Devoted Study Nerv Syst
Behav 73: 80–105, 2015.
Taylor JA, Ivry RB. The role of strategies in motor learning. Ann N Y Acad Sci 1251: 1–12,
2012.
Taylor JA, Ivry RB. Cerebellar and prefrontal cortex contributions to adaptation, strategies, and
reinforcement learning. Prog Brain Res 210: 217–253, 2014.
Tessitore A, Cirillo M, De Micco R. Functional Connectivity Signatures of Parkinson’s Disease.
J Park Dis 9: 637–652, 2019.
Therrien AS, Wolpert DM, Bastian AJ. Effective reinforcement learning following cerebellar
damage requires a balance between exploration and motor noise. Brain 139: 101–114, 2016.
Thompson WB, Willemsen P, Gooch AA, Creem-Regehr SH, Loomis JM, Beall AC. Does the
Quality of the Computer Graphics Matter when Judging Distances in Visually Immersive
Environments? Presence 13: 560–571, 2004.
Tian Y, Margulies DS, Breakspear M, Zalesky A. Topographic organization of the human
subcortex unveiled with functional connectivity gradients. Nat Neurosci 23: 1421–1432, 2020.
Timmis MA, Buckley JG. Obstacle crossing during locomotion: visual exproprioceptive
information is used in an online mode to update foot placement before the obstacle but not swing
trajectory over it. Gait Posture 36: 160–162, 2012.
Timmis MA, Scarfe AC, Pardhan S. How does the extent of central visual field loss affect
adaptive gait? Gait Posture 44: 55–60, 2016.
Tinaz S, Lauro PM, Ghosh P, Lungu C, Horovitz SG. Changes in functional organization and
white matter integrity in the connectome in Parkinson’s disease. NeuroImage Clin 13: 395–404,
2017.
Toro R, Fox PT, Paus T. Functional Coactivation Map of the Human Brain. Cereb Cortex N Y
NY 18: 2553–2559, 2008.
Torres-Oviedo G, Bastian AJ. Natural error patterns enable transfer of motor learning to novel
contexts. J Neurophysiol 107: 346–356, 2012.
Tseng Y, Diedrichsen J, Krakauer JW, Shadmehr R, Bastian AJ. Sensory Prediction Errors Drive
Cerebellum-Dependent Adaptation of Reaching. J Neurophysiol 98: 54–62, 2007.
Tzvi E, Münte TF, Krämer UM. Delineating the cortico-striatal-cerebellar network in implicit
motor sequence learning. NeuroImage 94: 222–230, 2014.

196
Van Ooteghem K, Frank JS, Horak FB. Postural motor learning in Parkinson’s disease: The
effect of practice on continuous compensatory postural regulation. Gait Posture 57: 299–304,
2017.
Vandenbossche J, Deroost N, Soetens E, Kerckhofs E. Does implicit learning in non-demented
Parkinson’s disease depend on the level of cognitive functioning? Brain Cogn 69: 194–199,
2009.
Vaughan N, Gabrys B, Dubey VN. An overview of self-adaptive technologies within virtual
reality training. Comput Sci Rev 22: 65–87, 2016.
Vehtari A, Gelman A, Simpson D, Carpenter B, Bürkner P-C. Rank-Normalization, Folding, and
Localization: An Improved Rˆ for Assessing Convergence of MCMC. Bayesian Anal 1: 1–28,
2021.
Verschueren SM, Swinnen SP, Dom R, De Weerdt W. Interlimb coordination in patients with
Parkinson’s disease: motor learning deficits and the importance of augmented information
feedback. Exp Brain Res 113: 497–508, 1997.
Voelcker-Rehage C. Motor-skill learning in older adults—a review of studies on age-related
differences. Eur Rev Aging Phys Act 5: 5–16, 2008.
Wadden KP, Asis KD, Mang CS, Neva JL, Peters S, Lakhani B, Boyd LA. Predicting Motor
Sequence Learning in Individuals With Chronic Stroke. Neurorehabil Neural Repair 31: 95–104,
2017.
Watson GS, Leverenz JB. Profile of Cognitive Impairment in Parkinson Disease. Brain Pathol
Zurich Switz 20: 640–645, 2010.
Wei L, Zhang J, Long Z, Wu G-R, Hu X, Zhang Y, Wang J. Reduced Topological Efficiency in
Cortical-Basal Ganglia Motor Network of Parkinson’s Disease: A Resting State fMRI Study.
PLOS ONE 9: e108124, 2014.
Weiner MJ, Hallett M, Funkenstein HH. Adaptation to lateral displacement of vision in patients
with lesions of the central nervous system. Neurology 33: 766–772, 1983.
Weintraub D, Burn DJ. Parkinson’s Disease: The Quintessential Neuropsychiatric Disorder. Mov
Disord Off J Mov Disord Soc 26: 1022–1031, 2011.
Weintraub D, Comella CL, Horn S. Parkinson’s disease--Part 1: Pathophysiology, symptoms,
burden, diagnosis, and assessment. Am J Manag Care 14: S40-48, 2008.
Werheid K, Hoppe C, Thöne A, Müller U, Müngersdorf M, von Cramon DY. The Adaptive
Digit Ordering Test: clinical application, reliability, and validity of a verbal working memory
test. Arch Clin Neuropsychol Off J Natl Acad Neuropsychol 17: 547–565, 2002.
Willingham DB. A neuropsychological theory of motor skill learning. Psychol Rev 105: 558–
584, 1998.

197
Winstein C, Lewthwaite R, Blanton SR, Wolf LB, Wishart L. Infusing motor learning research
into neurorehabilitation practice: a historical perspective with case exemplar from the accelerated
skill acquisition program. J Neurol Phys Ther JNPT 38: 190–200, 2014.
Winter B, Wieling M. How to analyze linguistic change using mixed models, Growth Curve
Analysis and Generalized Additive Modeling. J Lang Evol 1: 7–18, 2016.
Wise RA. Dopamine, learning and motivation. Nat Rev Neurosci 5: 483–494, 2004.
Wise SP, Moody SL, Blomstrom KJ, Mitz AR. Changes in motor cortical activity during
visuomotor adaptation. Exp Brain Res 121: 285–299, 1998.
Wise SP, Murray EA, Gerfen CR. The frontal cortex-basal ganglia system in primates. Crit Rev
Neurobiol 10: 317–356, 1996.
Wolpert DM, Ghahramani Z, Flanagan JR. Perspectives and problems in motor learning. Trends
Cogn Sci 5: 487–494, 2001.
Wood B, Bilclough J, Bowron A, Walker R. Incidence and prediction of falls in Parkinson’s
disease: a prospective multidisciplinary study. J Neurol Neurosurg Psychiatry 72: 721–725,
2002.
Wright DL, Shea CH. Contextual dependencies in motor skills. Mem Cognit 19: 361–370, 1991.
Wu J, Knapp F, Cramer SC, Srinivasan R. Electroencephalographic connectivity measures
predict learning of a motor sequencing task. J Neurophysiol 119: 490–498, 2017.
Wu J, Srinivasan R, Kaur A, Cramer SC. Resting-state cortical connectivity predicts motor skill
acquisition. NeuroImage 91: 84–90, 2014.
Wu T, Hallett M, Chan P. Motor automaticity in Parkinson’s disease. Neurobiol Dis 82: 226–
234, 2015a.
Wu T, Kansaku K, Hallett M. How Self-Initiated Memorized Movements Become Automatic: A
Functional MRI Study. J Neurophysiol 91: 1690–1698, 2004.
Wu T, Liu J, Zhang H, Hallett M, Zheng Z, Chan P. Attention to Automatic Movements in
Parkinson’s Disease: Modified Automatic Mode in the Striatum. Cereb Cortex N Y N 1991 25:
3330–3342, 2015b.
Wu T, Long X, Wang L, Hallett M, Zang Y, Li K, Chan P. Functional connectivity of cortical
motor areas in the resting state in Parkinson’s disease. Hum Brain Mapp 32: 1443–1457, 2011.
Wulf G, Lewthwaite R. Optimizing performance through intrinsic motivation and attention for
learning: The OPTIMAL theory of motor learning. Psychon Bull Rev 23: 1382–1414, 2016.

198
Yan C-G, Cheung B, Kelly C, Colcombe S, Craddock RC, Di Martino A, Li Q, Zuo X-N,
Castellanos FX, Milham MP. A comprehensive assessment of regional variation in the impact of
head micromovements on functional connectomics. NeuroImage 76: 183–201, 2013.
Yan L, Wu Y, Zeng X, Gao L. Dysfunctional putamen modulation during bimanual finger-to-
thumb movement in patients with Parkinson’s disease. Front Hum Neurosci 9, 2015.
Yang H, Zhou XJ, Zhang M, Zheng X, Zhao Y, Wang J. Changes in spontaneous brain activity
in early Parkinson’s disease. Neurosci Lett 549: 24–28, 2013.
Yang W, Liu B, Huang B, Huang R, Wang L, Zhang Y, Zhang X, Wu K. Altered Resting-State
Functional Connectivity of the Striatum in Parkinson’s Disease after Levodopa Administration.
PLOS ONE 11: e0161935, 2016.
Yang Y-R, Tsai M-P, Chuang T-Y, Sung W-H, Wang R-Y. Virtual reality-based training
improves community ambulation in individuals with stroke: a randomized controlled trial. Gait
Posture 28: 201–206, 2008.
Yokoi A, Diedrichsen J. Neural Organization of Hierarchical Motor Sequence Representations in
the Human Neocortex. Neuron 103: 1178-1190.e7, 2019.
Yokoyama H, Sato K, Ogawa T, Yamamoto S-I, Nakazawa K, Kawashima N. Characteristics of
the gait adaptation process due to split-belt treadmill walking under a wide range of right-left
speed ratios in humans. PLOS ONE 13: e0194875, 2018.
Yu R, Liu B, Wang L, Chen J, Liu X. Enhanced Functional Connectivity between Putamen and
Supplementary Motor Area in Parkinson’s Disease Patients. PLOS ONE 8: e59717, 2013.
Zhang D, Liu X, Chen J, Liu B, Wang J. Widespread Increase of Functional Connectivity in
Parkinson’s Disease with Tremor: A Resting-State fMRI Study. Front Aging Neurosci 7, 2015.

Abstract (if available)

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

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

PDF

Impact of enhanced efficacy expectation on motor learning in individuals with Parkinson’s disease

PDF

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

PDF

Behavioral and neurophysiological studies of hand dexterity in health and Parkinson's disease

PDF

Neural substrates associated with context-dependent learning

PDF

Exploring the gut-brain axis in Parkinson’s disease and the influence of physical fitness to restore gut health

PDF

Dynamic postural control during simple and complex locomotor tasks in persons with early stage Parkinson’s disease

PDF

The brain and behavior of motor learning: the what, how and where

PDF

Relationship between brain structure and motor behavior in chronic stroke survivors

PDF

Computational models and model-based fMRI studies in motor learning

PDF

Learning self-agency during a contingency paradigm in infants at commnity and elevated risk of autism

PDF

Learning reaching skills in non-disabled and post-stroke individuals

PDF

Experience-dependent neuroplasticity of the dorsal striatum and prefrontal cortex in the MPTP-lesioned mouse model of Parkinson’s disease

PDF

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

PDF

A virtual reality exergaming system to enhance brain health in older adults at risk for Alzheimer’s disease

PDF

Exploiting mechanical properties of bipedal robots for proprioception and learning of walking

PDF

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

PDF

Infant exploratory learning: influence on leg coordination

PDF

Quantify human experience: integrating virtual reality, biometric sensors, and machine learning

PDF

Relationship between interhemispheric inhibition and bimanual coordination in musicians

Asset Metadata

Creator Kim, Aram (author)

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

School School of Dentistry

Degree Doctor of Philosophy

Degree Program Biokinesiology

Degree Conferral Date 2021-08

Publication Date 07/14/2021

Defense Date 04/13/2021

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

Tag gait,locomotor learning,OAI-PMH Harvest,Parkinson's disease,virtual reality

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Finley, James M. (committee chair), Fisher, Beth (committee member), Kaplan, Jonas (committee member), Liew, Sook-Lei (committee member), Schweighofer, Nicolas (committee member)

Creator Email amy.rmkm@gmail.com,aramkim@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC15491152

Unique identifier UC15491152

Legacy Identifier etd-KimAram-9742

Document Type Dissertation

Format application/pdf (imt)

Rights Kim, Aram

Internet Media Type application/pdf

Type texts

Source 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. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

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