Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Understanding the pathology of dystonia by hardware emulation
(USC Thesis Other)
Understanding the pathology of dystonia by hardware emulation
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
Copyright 2015 Won Joon Eric Sohn
BIOMEDICAL ENGINEERING
Understanding the pathology of dystonia by
hardware emulation
By
Won Joon Eric Sohn
submitted to the
Department of Biomedical Engineering.
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the Requirements for the Degree
DOCTOR OF PHILOSOPHY
Principal Investigator: Terence D. Sanger
August 2015
2
To my parents,
For their unwavering support and love.
3
Acknowledgements
I would like to express my gratitude to many people who encouraged and helped me along the
way, by walking beside me, to come to the journey of life this far. In particular, I am immensely
indebted to my advisor Terence Sanger for his generous encouragement and guidance throughout
my Ph.D. studies. His intellectual input and encouragement was sufficient to propel me to bring
this study forward and helped me to defend the thesis. I am also thankful that he provided me a
model for an excellent academic leader. I sincerely thank C. Minos Niu, friend, coworker and a
mentor, for his inspiration and guidance throughout my doctoral studies. My successful thesis
completion cannot be explained without his contribution. I would also like to express my
gratitude to Francisco Valero-Cuevas for his whole-hearted support for this study and his
personal encouragement. I take this opportunity to thank Alice Parker and Viktor Prasanna for
not only serving as my dissertation committee members but also for generously spending their
time to review this study. I wish to acknowledge the collaborative work of this study provided
by Sirish Nandyala. I would like to thank my colleagues Matteo Bertucco, Diana Ferman, Aprille
Tongol, Shanie Liyanagamage, Enrique Arguelles, Amber Dunning, Adam Feinman, Cassie
Borish, Francesca Lunardini, Nasir Bhanpuri, Scott Young for their tremendous support and
contribution to build a friendly and warm working environment, which should not be taken for
granted. My gratitude also goes to Junseob Shin and people in the Light Ministry for their
companionship in walking the narrow road of life. I am grateful for the support from the James S.
McDonnel Foundation and National Institute of Neurologic Disorders and Stroke (R01-
NS069214) and Biomedical Engineering Department from University of Southern California. I
dedicate this work to my family for their endless love and patiently supporting me. I would like
to mention my sincere gratitude for Hyemi Kim for her tremendous patience, love, inspiration
and support. Lastly, I thank God for His providence. From the beginning of the doctoral study to
this moment, I lacked nothing and I confess that His grace was sufficient for me.
4
Table of Contents
1 Introduction ...................................................................................... 11
1.1 Specific aim ...................................................................................................................... 11
1.2 Background ...................................................................................................................... 11
1.3 Pathophysiology of dystonia ............................................................................................ 13
1.4 Impact of an early treatment of dystonia .......................................................................... 14
1.5 The value of emulation in studying injury in nervous system leading to abnormal
motor behavior .......................................................................................................................... 15
2 Technical preparation ..................................................................... 16
2.1 Challenges of studying developmental motor disorders .................................................. 17
2.2 Methodology of multi-scale neural emulation ................................................................. 19
2.2.1 Modularized architecture for multi-scale models ...................................................... 19
2.2.2 Selection of models for emulation ............................................................................. 20
2.2.3 Neuron connectivity with sparse interconnections .................................................... 23
2.2.4 Hardware implementation on FPGA ......................................................................... 25
2.2.5 Floating-point arithmetics in combinational logic ..................................................... 25
2.2.6 Asynchronous spike-based communication between FPGA chips ........................... 26
2.2.7 Serialize neuron evaluations within a homogeneous population ............................... 27
2.3 Results: emulated activities of motor nervous system ..................................................... 27
2.4 Discussion and future work .............................................................................................. 28
3 Study 1: What is the mechanism of childhood secondary
dystonia due to cerebral palsy? ........................................................... 32
3.1 Introduction ...................................................................................................................... 35
3.2 Models for the sensorimotor system and increased LLSR ............................................... 38
3.2.1 Implementation of sensorimotor system on hardware ............................................... 41
3.2.2 Experiments ............................................................................................................... 43
3.3 Results .............................................................................................................................. 46
3.3.1 Non-impaired stretch reflex and hypertonia with increased LLSR ........................... 46
3.3.2 Experiment 1: Involuntary responses to passive joint stretch ................................... 48
5
3.3.3 Experiment 2: Delayed relaxation of muscle force ................................................... 53
3.3.4 Experiment 3: Reduced range of motion in voluntary movements ........................... 56
3.4 Discussion ........................................................................................................................ 57
3.4.1 Relationship between LLSR and secondary dystonia ............................................... 57
3.4.2 Rationale of model validation .................................................................................... 58
3.4.3 Thoughts on the two mechanisms of elevated LLSR ................................................ 59
3.4.4 Factors that may affect the phase angle in Experiment 1 .......................................... 59
3.4.5 EMG response during muscle shortening .................................................................. 60
3.4.6 Advantage and limitation of neuromorphic hardware emulation .............................. 60
3.5 Conclusion ........................................................................................................................ 61
3.6 Author Contributions ....................................................................................................... 61
4 Study 2. What causes motor overflow in focal hand dystonia? ... 62
4.1 Introduction ...................................................................................................................... 65
4.2 Materials and methods ..................................................................................................... 67
4.2.1 Experimental procedure ............................................................................................. 70
4.3 Results .............................................................................................................................. 72
4.3.1 Demo of plasticity effect under STDP ...................................................................... 72
4.3.2 Development and perpetuation of motor overflow .................................................... 74
4.3.3 Enlarged sensory representation and increase in receptive field ............................... 75
4.4 Discussion ........................................................................................................................ 78
4.4.1 Relevance of the simulated neural structure to the biology ....................................... 79
4.4.2 Sensitivity of the result to other synaptic learning models ........................................ 80
4.4.3 Implication of the emulated result in clinical treatment of FHD ............................... 80
4.5 Conclusion ........................................................................................................................ 81
4.6 Author Contributions ....................................................................................................... 81
5 Study 3: How does the constraint-induced therapy work? .......... 82
5.1 Introduction ...................................................................................................................... 85
5.1.1 Postnatal brain damage and its effect in corticospinal projection ............................. 86
5.1.2 Spike-based synaptic learning rule ............................................................................ 87
5.1.3 Benefit of simulating a biological system ................................................................. 88
6
5.2 Materials and methods ..................................................................................................... 89
5.2.1 Neural structure ......................................................................................................... 90
5.2.2 Experimental procedure ............................................................................................. 92
5.2.3 Constraint-induced therapy uses the property of synaptic competition. ................... 94
5.3 Results .............................................................................................................................. 96
5.3.1 Illustration of activity-dependent synaptic competition ............................................ 96
5.3.2 Simulating activity-dependent constraint-induced therapy by STDP ....................... 98
5.3.3 What is the condition that the states will be consolidated? ....................................... 99
5.4 Discussion ...................................................................................................................... 103
5.4.1 Relevance of the simulated neural structure to the biology ..................................... 103
5.4.2 Similarity to the treatment of amblyopia ................................................................. 104
5.4.3 Escaping from the suboptimal stable state .............................................................. 104
5.4.4 Sensitivity of the result to other synaptic learning models ...................................... 105
5.4.5 Caveats in the simulation of biological systems ...................................................... 105
5.5 Conclusion ...................................................................................................................... 106
6 Conclusions and future work ........................................................ 107
6.1 Conclusion ...................................................................................................................... 107
6.2 Future studies ................................................................................................................. 109
7 Bibliography ................................................................................... 111
7
Table of Figures
Figure 1: Illustration of the multi-scale nature of motor nervous system. ................................... 19
Figure 1.5. Functions of neuron population can be described as the combination of linear
operators (A). Therefore the original neural function can be equivalently produced by sparsely
connected neurons formalizing parallel pathways (B). ................................................................. 24
Figure 2: Timing diagram of asynchronous spike-based communication ................................... 26
Figure 3. The neural emulation platform in operation. Left: One working FPGA node. Center:
Two FPGA nodes networked using asynchronous spiking protocol. Right: Software front-end
displaying multi-scale signals. ...................................................................................................... 28
Figure 4: A) Physiological activity emulated by each model when the muscle is sinusoidally
stretched. B) Comparing the emulated motor unit recruitment order with real experimental data
....................................................................................................................................................... 31
Figure 5. A) Components of human sensorimotor system model. The system includes a limb
joint comprising two opposing monoarticular muscles (flexor and extensor), muscle spindles,
spindle afferents, alpha-motoneurons, and associated spinal and supraspinal structures. Our
model includes a monosynaptic reflex arc with a 32ms loop-delay and a supra-spinal
transcortical reflex pathway with a 64ms loop-delay. B) The TONIC model of dystonia. The
dashed box shows the procedure of introducing involuntary tonic activity. Before the disease
onset (unshaded area), the system executes the voluntary commands received from antagonistic
cortical inputs; after the disease onset (shaded area), the voluntary descending commands are
superimposed on a tonic input. C) The HI-GAIN model of dystonia. The synaptic gain associated
with the afferent cortical projection becomes higher after the disease onset (shaded areas), which
directly increases the loop-gain of the transcortical feedback loop. ............................................. 41
Figure 6. Detailed configuration of the motor nervous system on neuromorphic hardware. ...... 43
Figure 7. The emulated stretch reflex in the non-impaired condition and models of increased
LLSR. ............................................................................................................................................ 47
Figure 8. Biceps EMG during arm rotation in a child with hypertonic arm dystonia (A) created
based on the data in van Doornik et al. (2009). When the subject was instructed to rest, the right
arm was rotated manually by the experimenter following approximately a sinusoidal time profile
with frequency varying from 0.2 Hz up to 2 Hz. The biceps showed phasic EMG responses to the
manual stretch. In the emulation, the virtual joint was passively rotated with the identical
waveform under two models of dystonia. The voluntary command is set to zero to represent the
subject being “at rest”. In the TONIC model (B), the EMG is not silent when the joint is
passively rotated and shows similar phasic patterns as the human EMG recording. In the HI-
GAIN model (C), similar non-silent EMG responses are observed with increased magnitude in
phasic response. ............................................................................................................................ 50
Figure 9. The phase angle calculated from emulation and human data. ...................................... 51
8
Figure 10. The sensitivity of phase angle to the intensity of emulated dystonia in each model. . 53
Figure 11. Relaxation of muscle activity from a state of co-contraction in biceps and triceps in a
normal subject and a patient with secondary dystonia (A), created based on the data in Ghez et al.
(1988). Subjects were first required to maximally co-contract the biceps and triceps muscles and
upon a visual cue relax the muscle as fast as possible. The duration for muscle relaxation in the
patient with secondary dystonia (right) was elongated compared to the normal subject (left). In
emulation (B), both TONIC and HI-GAIN models showed delayed EMG relaxation compared
with the non-impaired condition, when the voluntary descending command that co-contracts
biceps and triceps were abruptly shut off. Single representative trials are shown. ...................... 54
Figure 12. Relationship between the rate of relaxation (τ) and the intensity of either the TONIC
or the HI-GAIN model. ................................................................................................................. 56
Figure 13. Dystonia models predict that the range of motion in voluntary arm-swing movement
should be reduced in dystonia; with certain level of compensation by increasing the voluntary
command, the range of movement can be recovered. ................................................................... 57
Figure 14. STDP demo. A) Spike timing dependent plasticity curve implemented on FPGA. ... 70
Figure 15. The experimental design testing whether confused (correlated) sensory inputs may
lead to motor overflow. ................................................................................................................. 72
Figure 16. Validation of STDP implemented using neuromorphic hardware. ........................... 73
Figure 17. Development and perpetuation of the motor overflow. Change of synaptic weight of
the crosstalk is plotted. .................................................................................................................. 75
Figure 18. Visual aid for the enlarged cortical representation, increase in receptive field and
growth in crosstalk. We also see a decrease in spatial specificity due to a growth of crosstalk. . 77
Figure 19. Qualitative comparison of motor overflow between emulated and human data.
Signals only at diagonal windows represent healthy human control. In dystonia, signals in the
non-diagonal windows increases due to motor overflow. Similar pattern is observed between
emulation of control and dystonia (right). .................................................................................... 78
Figure 20. Model neural structure. A) Simplified schematic for descending CST. CSTs initiate
from the motor cortex and terminate on the cervical gray matter in the spinal cord. The bold
descending lines represent contralateral projection (e.g. right hemisphere to left spinal gray
matter) and the dotted line represents ipsilateral projection. The projection strength of the
descending tract is represented by the relative thickness of the lines. In normal development,
contralateral projection dominates the ipsilateral projection. Distribution of CST projection
within the gray matter is not considered for simplification. B) In the simulated neural structure,
two layers of spiking neurons representing cortical neurons (input neurons) and output neurons
(spinal neurons) are connected via synapses (triangles). A strength of the synapse is represented
both by the relative size of the triangle in the structure and by the opacity of the colors in the
color matrix. Red corresponds to the cortical projection from the right hemisphere and black for
the left hemisphere. ....................................................................................................................... 90
Figure 21. STDP model: the model includes standard all-to-all, additive STDP with synaptic
decay and stochastic current input to the neuron as an activity generator.. .................................. 92
9
Figure 22. Simulating activity-dependent constraint-induced therapy by STDP in 4 neurons, 4
synapses neuronal structure. ......................................................................................................... 95
Figure 23. Contingencies for a synaptic competition according to STPD when two input
neurons project to an output neuron. ............................................................................................. 98
Figure 24. Simulating activity-dependent constraint-induced therapy by STDP: a change of
synaptic weights of 4 synapses according to the five stages of input profiles.. .......................... 102
10
Summary
Movement disorders are neurological conditions that affect speed, fluency, quality, and ease of
movement in a negative direction. In that regard, investigating the neurological underpinning of
the cause of the movement disorder is desirable. In case of secondary dystonia due to cerebral
palsy, which causes involuntary movements, postures and prolonged muscle contraction, great
difficulties in understanding the mechanism of the disorder have been posed by physiologically
limited methods to conduct studies investigating the underlying neurological mechanism in
human body. The emulation study presented here is one of the alternative responses to overcome
the limitation posed by human studies. Centered on understanding the dystonia, three emulation
studies were conducted. The first study proposes two plausible neurological mechanisms that
lead to behavioral characteristics of dystonia and the outcomes from the emulation are compared
with available data from subjects with dystonia (chap 3). The second study investigates the
origin and development of motor overflow in focal hand dystonia in the context of spike-based
plasticity mechanism (chap 4). The third study investigates the mechanism of constraint-induced
therapy, a popular rehabilitative method in impaired biological systems with spike-based
plasticity mechanism (chap 5). The studies utilize extremely fast and customizable hardware,
which provides a unique benefit of accelerated emulation of the development of neurological
system under tested circumstances. The first study is published in Journal of Neural Engineering.
The second and third studies will be submitted to relevant journals (undecided at this point). The
engineering technique and general methodology behind the use of programmable hardware
(chap 2) is published in neural information processing systems (NIPS).
11
1 Introduction
1.1 Specific aim
There is currently no quantitative model of how the functions of neurons affect the specific
abnormalities observed in movement disorders. Although clinical experience has provided
insight in making qualitative prediction of how certain kinds of injury might lead to particular
outcome, qualitative clinical insight itself is of little use if the goal is to make a specific
prediction on the effect of particular impairment–whether they are neuronal or anatomic in
nature, based on specific quantifiable physiological measures in individual patients. In order to
study the causal relationship between a particular neuronal injury and the resultant immediate or
long term biomechanical effect on the movement of a patient, we designed a multi-purpose high-
speed emulation platform in scalable hardware. In this project, the platform is designed to
emulate a subset of human sensorimotor nervous system that is speculated to be responsible for
many movement disorders when it is impaired. Technical preparation section (chap 2) is
dedicated for the methodological considerations. The structures of fundamental building blocks
of the monosynaptic spinal stretch reflex pathway, including spiking neurons, spindle, muscle,
and synapse, etc., as well as how fast computation are achieved in customizable hardware are
described. Although the thesis is centered on understanding the pathology of dystonia, the third
study (chap 5) extends the use of the technology to understand the mechanism for the popular
rehabilitative method called constraint-induced therapy.
1.2 Background
Dystonia is a movement disorder in which involuntary sustained or intermittent muscle
contractions cause twisting and repetitive movements, abnormal postures, or both (Fahn, 1988;
Sanger et al., 2003). Dystonia may involve one body region such as neck, face, leg or hand (focal
dystonia), involve contiguous body regions (segmental dystonia) or involve broad regions of the
body (generalized dystonia). Dystonia is called primary if the origin is known to be genetic or
hereditary, and secondary if it results from structural damages or environmental factor that
provided insult to the brain. Statistics shows that dystonia affects men, women, and children of
all ages and backgrounds. It is estimated that there are 250,000 cases of idiopathic dystonia in the
12
U.S., 1:3,000 ratio, but the true prevalence could rise much above the reported number. Focal
dystonia, 300 per millions, has 9 times the prevalence of generalized dystonia, among them
cervical dystonia (CD) is the most frequent form among focal dystonia, writer’s cramp
(graphospasm) may be the most prevalent form of dystonia in the general population.
Dystonia causes a wide degree of disabilities and pain ranging from mild to severe. Dystonia
can have a devastating impact on the quality of life of a patient and their family in that it
debilitates physical and mental wellbeing as well as social function of a person, not to mention it
can have a stigmatizing effect of living with such a visible disorder. At present there is no known
cure, but only selective treatment options exist.
It is worth mentioning the term cerebral palsy (CP) in relation to dystonia since it is a blanket
term for a group of movement and posture disorders that occur as a consequence of damages in
the brain which is acquired at an early age when the brain is still developing – before birth,
during birth and immediately after birth. Typically, dystonia is one feature or symptom
associated with the syndrome of dyskinetic CP (a type of CP featuring variable movement that is
involuntary) when it involves twisting and repetitive movements with fixed postures. Dyskinetic
movement is called athetosis when it involves slow and ‘stormy’ movements, chorea when it
involves dance-like irregular, unpredictable movements and so on.
In focal task-specific dystonia, a type of dystonia that is characterized by excessive muscle
contraction producing abnormal posture during selective motor activity that often involve highly
skilled, repetitive movements, is best known for focal hand dystonia, writer’s cramp, musician’s
cramp or occupational dystonia because it interferes with the performance of the common tasks
such as writing or playing a musical instrument. The origin and development of focal hand
dystonia is emulated in the second study (chap 4).
In this project, we attempt to emulate some of the well-known characteristic features of dystonia
including:
• hypertonia, an increased resistance to a passive perturbation
• prolonged time required to relax previously-contracted muscles
• reduced active range of motion
• development and perpetuation of motor overflow in focal hand dystonia
13
And also we attempt to emulate the mechanism of constraint-induced therapy in the hemiplegic
CP.
In doing so, physiological evidence of dystonia will be cross-examined with clinical data.
1.3 Pathophysiology of dystonia
Neurophysiological mechanisms that lead to the clinical manifestations of dystonia are largely
unknown. The exact cause of dystonia may be highly heterogeneous due to the variety of injuries
identified in clinical examinations. It is classified as primary dystonia when no abnormality in
the brain is observed and secondary dystonia when observable damage or lesion in the brain is
seen which can be detected in the advanced neuroimaging techniques.
In specific, lesions in the basal ganglia, thalamus, or brain stem is likely to associate with
secondary dystonia (Kahn et al., 1985); and although not common, dystonia can sometimes
appear after trauma or peripheral nerve injuries (Jankovic, 2001). Structural lesion could be a
result from brain trauma, tumor, stroke, oxygen deprivation, infection, drug reactions, poisoning
caused by lead or carbon monoxide (webmd.com). Brain injuries eventually develop into
symptoms of dystonia, perhaps through the ensuing abnormality in the motor cortex. For
example, injuries in basal ganglia have been found to increase the activity of the primary motor
cortex (Playford et al., 1993; Ceballos-Baumann, 1994), which could be attributable to impaired
cortical inhibition (Hallett, 2011; Beck et al., 2008; Ridding et al., 1995). Clinical treatments
also support the linkage between cortical injury and dystonia through the effect of motor cortex:
high frequency stimulation and ablative surgeries in thalamus and GPi, the main basal ganglia
outflow nucleus inhibitory to motor cortex, has been practiced to reduce cortical overactivity in
dystonia patients. Moreover, long-latency stretch reflex (LLSR) has been found hyperactive
during voluntary movements in childhood dystonia. Due to the possible linkage between motor
cortical activity and stretch reflex (Suminski et al., 2007; Morimoto et al., 1984; Evarts and
Tanji, 1976), it is possible that dystonia is not specific to injuries but can be sufficiently caused
by a hyperactive LLSR pathway. Here we focus on testing this possibility using hardware
emulation.
14
The causes of focal task-specific dystonia are unknown, but it is conjectured that the disorder is
likely the combination of genetic and environmental factors. There is an evidence from studies in
monkeys of disorganization of sensory cortical representation. In particular, monkeys with focal
hand dystonia had neurons in the primary sensory cortex responded to tactile stimulation in more
than one finger, in other words with a confused receptive field. It has been shown that the
coupled use of two fingers for a long time could develop a de-differentiation in the cortical
representational map in the primary sensory cortex. As such, the conjecture is that sensory
abnormalities could contribute to the motor abnormality observed in dystonia. A previous study
(Sanger and Merzenich, 2000) provided a computational model that explains how the sensory
abnormalities could lead to motor manifestations of dystonia. Sanger hypothesized that task-
specific dystonia such as writer’s cramp has to do with an abnormally high gain in the
sensorimotor loop that could take years to develop.
1.4 Impact of an early treatment of dystonia
In many cases the causes of CP are unknown and thus it is hard to prevent it. Because early
signs of CP usually appear before a child reaches 3 years of age, physical impairment such as a
problem in walking, lack of muscle coordination, hearing loss, speech and language disorders
can cause severe developmental problem. Often children with CP develop into intellectual
disability. Therefore, the treatment is aimed at helping child’s motor and cognitive development
and to prevent the occurrence of secondary injury (Sanger et al. 2008).
For example, reduction in arm dystonia may permit improved handwriting, which may in
turn allow better participation in school and thereby contribute to improved intellectual
ability later in life. Thus, the neurologist can work with the child, the family, and other
clinicians to facilitate and ensure developmental progress. Just as a physical therapist’s
efforts may be required to obtain maximum benefit from a neurologist’s tone-reducing
medication, a neurologist’s efforts may be required to provide medications that permit
continued progress in physical therapy (Sanger 2008)
Currently, there is no cure for CP, but there are many treatments that can increase the quality of
life for kids with CP. In dystonia, pharmacologic therapies including treating Levodopa,
anticholinergic medication, baclofen, clonazepam as well as nonpharmacological therapies such
15
as desensitization, limb immobilization which aim at restructuring the cortical map with learning
based on the principle of neuroplasticity. Injection of botulinum toxin, which targets local
muscle with a several months of short-lasting period, is currently a mainstay of treatment for
most focal dystonia. Surgical treatment is not often suggested due to its risk factors but targeting
the anatomical source of a disturbance with deep brain stimulation, pallidotomy has been
successfully performed for many decades.
1.5 The value of emulation in studying injury in nervous system
leading to abnormal motor behavior
1. Emulation study allows us to answer questions about the sufficient mechanism
responsible for the movement disorders that have a neurological origin.
2. We can emulate specific disease hypotheses without emulating whole brain by building a
neural circuit around a structure that is relatively well-known, such as stretch reflex
pathway.
3. We can conduct experiment that is generally prohibited in human studies due to practical
and ethical reasons.
4. We can use the emulator to study not only the immediate effects of a change in
physiological parameter to account for any behavior abnormalities but also the long-term
plasticity effect of certain injury or intervention to the system.
5. The only way to test and characterize the high-level behavior of a brain model is to
actually build the closed loop between the artificial nervous system and the body (plant)
acting in an environment and to interrogate the model through a well-designed
experiment.
6. Once the general purpose emulation platform is built, we can identify the causal
mechanism by manipulating specific brain regions, such as simulating the effect of brain
lesion, neuronal injury, cell death, pharmacologic treatment, rehabilitation, etc.
16
2 Technical preparation
Disclaimer: This section is originally published in neural information processing systems (NIPS).
C.M. Niu, S. Nandyala, W.J. Sohn, and T.D. Sanger, Multi-scale Hyper-time Hardware
Emulation of Human Motor Nervous System Based on Spiking Neurons using FPGA.
Advances in Neural Information Processing Systems 25 (2012) 37-45.
Our central goal in building a high-speed emulation platform in hardware is to quantify the long-
term progression of pediatric neurological diseases, such as a typical 10-15 years progression of
child dystonia. To this purpose, quantitative models are convincing only if they can provide
multi-scale details ranging from neuron spikes to limb biomechanics. The models also need to be
evaluated in hyper-time, i.e. significantly faster than real-time, for producing useful predictions.
We designed a platform with digital VLSI hardware for multiscale hyper-time emulations of
human motor nervous systems. The platform is constructed on a scalable, distributed array of
Field Programmable Gate Array (FPGA) devices. All devices operate asynchronously with 1
millisecond time granularity, and the overall system is accelerated to 365x real-time. Each
physiological component is implemented using models from well documented studies and can be
flexibly modified. Thus the validity of emulation can be easily advised by neurophysiologists
and clinicians. The following sections present the methodology of building FPGA modules in
correspondence to components of a monosynaptic spinal loop. Results of emulated activities are
shown. The paper also discusses the rationale of approximating neural circuitry by organizing
neurons with sparse interconnections. In conclusion, our platform allows introducing various
abnormalities into the neural emulation such that the emerging motor symptoms can be analyzed.
It compels us to test the origins of childhood motor disorders and predict their long-term
progressions.
17
2.1 Challenges of studying developmental motor disorders
There is currently no quantitative model of how a neuropathological condition, which mainly
affects the function of neurons, ends up causing the functional abnormalities identified in clinical
examinations. The gap in knowledge is particularly evident for disorders in developing human
nervous systems, i.e. childhood neurological diseases. In these cases, the ultimate clinical effect
of cellular injury is compounded by a complex interplay among the child’s injury, development,
behavior, experience, plasticity, etc. Qualitative insight has been provided by clinical
experiences into the association between particular types of injury and particular types of
outcome. Their quantitative linkages, nevertheless, have yet to be created — neither in clinic nor
in cellular physiological tests. This discrepancy is significantly more prominent for individual
child patients, which makes it very difficult to estimate the efficacy of treatment plans. In order
to understand the consequence of injury and discover new treatments, it is necessary to create a
modeling toolset with certain design guidelines, such that child neurological diseases can be
quantitatively analyzed.
Perhaps more than any other organ, the brain necessarily operates on multiple spatial and
temporal scales. On the one hand, it is the neurons that perform fundamental computations, but
neurons have to interact with large-scale organs (ears, eyes, skeletal muscles, etc.) to achieve
global functions. This multi-scale nature worth more attention in injuries, where the overall
deficits depend on both the cellular effects of injuries and the propagated consequences. On the
other hand, neural processes in developmental diseases usually operate on drastically different
time scales, e.g. spinal reflex in milliseconds versus learning in years. Thus when studying motor
nervous systems, mathematical modeling is convincing only if it can provide multi-scale details,
ranging from neuron spikes to limb biomechanics; also the models should be evaluated with time
granularity as small as 1 millisecond, meanwhile the evaluation needs to continue trillions of
cycles in order to cover years of life.
It is particularly challenging to describe the multi-scale nature of human nervous system
when modeling childhood movement disorders. Note that for a child who suffered brain injury at
birth, the full development of all motor symptoms may easily take more than 10 years. Therefore
the millisecond-based model needs to be evaluated significantly faster than real-time, otherwise
18
the model will fail to produce any useful predictions in time. We have implemented realistic
models for spiking motoneurons, sensory neurons, neural circuitry, muscle fibers and
proprioceptors using VLSI and programmable logic technologies. All models are computed in
Field Programmable Gate Array (FPGA) hardware in 365 times real-time. Therefore one year of
disease progression can be assessed after one day of emulation. This section presents the
methodology of building the emulation platform. The results demonstrate that our platform is
capable of producing physiologically realistic multi-scale signals, which are usually scarce in
experiments. Successful emulations enabled by this platform will be used to verify theories of
neuropathology. New treatment mechanisms and drug effects can also be emulated before animal
experiments or clinical trials.
19
2.2 Methodology of multi-scale neural emulation
Figure 1: Illustration of the multi-scale nature of motor nervous system.
The motor part of human nervous system is responsible for maintaining body postures and
generating voluntary movements. The multi-scale nature of motor nervous system is illustrated
in Error! Reference source not found.. When the elbow (Error! Reference source not
found.A) is maintaining a posture or performing a movement, the involved muscle produces
force based on how much spiking excitation is delivered from its
motoneurons (Error! Reference source not found.B). The motoneurons are regulated by their
own sensory input, which in-turn comes from the proprioceptors residing in the muscle. As the
primary sensory organ found in skeletal muscles, a muscle spindle is another complex system
that has its own microscopic Multiple-Input-Multiple-Output structure (Error! Reference
source not found.C). Spindles continuously provide information about the length and
lengthening speed of the muscle fiber. This section uses the monosynaptic spinal loop as an
example for explaining the methodology of multi-scale hyper-time neural emulation in hardware.
Additional structures can be added to the backbone platform using similar methods described
here.
2.2.1 Modularized architecture for multi-scale models 2.2.1 Modularized architecture for multi-scale models
Decades of studies on neurophysiology provided an abundance of models characterizing
different components of the human motor nervous system. The functional differentiation
between physiological components allowed us to model the motor nervous system as
20
concatenated structures, each of which maps input signals to the output. In particular, in a
monosynaptic spinal loop illustrated in Error! Reference source not found.B, stretching the
uscle will elicit a chain of physiological activities as: muscle stretch ⇒ spindle ⇒ sensory neuron
⇒ synapse ⇒ motoneuron ⇒ muscle contraction. The adjacent components must have compatible
interfaces, and the interfacing variables must also be physiologically realistic. In our design, each
component is mathematically described in Table 1:
As can be seen, all components are modeled as black-box functions that map the inputs to the
outputs. The meanings of these mathematical definitions are explained below. This design allows
existing physiological models to be easily inserted and switched. In all models the input signals
are time-varying, e.g. I = I(t); L = L(t) , etc. The argument of t in input signals are omitted
throughout this paper.
2.2.2 Selection of models for emulation
Models were selected in consideration of their computational cost, physiological verisimilitude,
and whether it can be adapted to the mathematical form defined in Table 1.
Model of Neuron
Neurons take post-synaptic current I as the input, and produce a binary spike train S in the output.
The neuron model adopted in the emulation was developed by Izhikevich (Izhikevich, 2003b):
21
where the output is the action potential v, which directly produces a binary spike train; a; b; c; d
are model parameters that need to be tuned based on the neuron’s firing properties. Note that in
Izhikevich model the action potential v is in millivolts and the time is in milliseconds. Since all
other models require SI units the coefficients in eq.1 need to be adjusted.
Model of Synapse
When a pre-synaptic neuron fires, i.e. S(0) = 1, an excitatory synapse subsequently produces an
Excitatory Post-Synaptic Current (EPSC) that drives the post-synaptic neuron. Neural recording
of hair cells in rats (Glowatzki and Fuchs, 2002) provided evidence that the time profile of EPSC
can be well characterized using the equations below:
The key parameters in a synapse model is the time constants for rising (τr) and decaying (τd). In
our emulation τr = 0.001s and τr = 0.003s.
Model of Muscle force and electromyograph (EMG)
The primary effect of skeletal muscle is converting the motoneuron spikes S into a force T
depending on the instantaneous length L and lengthening speed L
̇ of the muscle itself. We used
Hill’s muscle model in the emulation with parameter tuning described in (Shadmehr and Wise,
2005). Another measurable output of muscle is electroencephalograph (EMG). EMG is the small
skin current polarized by motor unit action potential (MUAP) when it travels along muscle fibers.
Models exist to describe the typical waveform picked by surface EMG electrodes. In this project
we chose to implement the one described in (Fuglevand et al., 1993b). We further implement the
muscle to produce the recruitment order and size principles observed in real physiological data.
It has been well known that when a voluntary motor command is sent to the motoneuron pool,
the motor units are recruited in an order that small ones get recruited first followed by the big
ones (Henneman, 1957). Further detail can be found in Figure 4 and Figure 6.
22
Hill’s muscle model is mathematically described as the differential equation above. Recaptured
in Shadmehr and Arbib (1992).
T Muscle Tension.
A Active tension applied by stimulation of membrane voltage.
K
se
Serial spring constant (SE).
K
pe
Parallel spring constant (PE).
B Damping Coefficient.
X Muscle Length.
Model of Proprioceptor
Spindle is a sensory organ that provides the main source of proprioceptive information. As can
be seen in Fig.1C, a typified spindle model produces two afferent outputs (primary Ia and
secondary II) according to its gamma fusimotor drives (Γ
dynamic
and Γ
static
) and muscle states (L and
L
̇ ). Spindle model needs to account for various types of stretch inputs, which requires complex
dynamic model. The complexity originated from the non-linear nature of muscle fibers and their
coupling with spike generating spots. On representative model that numerically approximates the
spindle dynamics was developed by Mileusnic et al. (Mileusnic et al., 2006b). The model used
differential equations to characterize a typical cat soleus spindle. Eqs.3-9 present a subset of this
model for spindle bag1 fiber:
23
2.2.3 Neuron connectivity with sparse interconnections
Although the number of spinal neurons (~1 billion) is significantly less compared to that in the
brain cortex (~100 billion), a fully connected spinal network still means approximately 2 trillion
synaptic endings (Gelfan et al., 1970). Implementing such a huge number of synapses imposes a
major challenge, if not impossible, given limited hardware resource. In this platform we
approximated the neural connectivity by sparsely connecting sensory neurons to motoneurons as
parallel pathways. We do not attempt to introduce the full connectivity. The rationale is that in a
neural control system, the effect of a single neuron can be considered as mapping current state x
to change in state x through a band-limited channel. Therefore when a collection of neurons are
firing stochastically, the probability of _ x depends on both x and each neuron’s firing behavior s
(s = 1 when spiking, otherwise s = 0), as such:
p(ẋ|x, s) = p(ẋ|s = 1)p(s = 1|x) + p(ẋ|s = 0)p(s = 0|x)
Eq.10 is by definition a master equation that determines a probability flow on the state. From the
Kramers-Moyal expansion we can associate this probability flow with a partial differential
equation for the change in probability density:
24
It has been shown in (Sanger, 2010, 2011b) that when higher order (>2) fluctuations in the
probability density are ignored, according to the Fokker-Planck equation, the probability flow
can be deterministically described using a linear operator L:
Due to the linearity, various Ls can be superimposed to achieve complex system dynamics
(illustrated in Error! Reference source not found.A)
Figure 1.5. Functions of neuron population can be described as the combination of linear
operators (A). Therefore the original neural function can be equivalently produced by sparsely
connected neurons formalizing parallel pathways (B).
As a consequence, the statistical effect of two fully connected neuron populations is
equivalent to ones that are only sparsely connected, as long as the probability flow can be
described by the same L. In particular, in a movement task it is the statistical effect from the
neuron ensemble to skeletal muscles that determines the global behavior. Therefore we argue
that it is feasible to approximate the spinal cord connectivity by sparsely interconnecting sensory
and motor neurons (Fig.1.5B). Here a pool of homogenous sensory neurons projects to another
pool of homogeneous motoneurons. Pseudorandom noise is added to the input of all
homogeneous neurons within a population. It is worth noting that this approximation
significantly reduces the number of synapses that need to be implemented in hardware.
25
2.2.4 Hardware implementation on FPGA
We select FPGA as the implementation device due to its inherent parallelism that resembles the
nervous system. FPGA is favored over GPU or clustered CPUs because it is relatively easy to
network hundreds of nodes with customizable protocols. The platform is distributed on multiple
nodes of Xilinx Spartan-6 devices. The interfacing among FPGAs and computers is created using
OpalKelly development board XEM6010. The dynamic range of variables is tight in the selected
models of Izhikevich neuron, synapse and EMG, which helps maintaining the accuracy even
when evaluated in 32-bit fixed-point arithmetics. The spindle model, in contrast, requires
floating-point arithmetics due to its wide dynamic range and complex calculations (see eq.3-9).
Hyper-time computations with floating-point numbers are resource consuming and therefore
need to be implemented with special attentions.
2.2.5 Floating-point arithmetics in combinational logic
Our arithmetic implementations are compatible with IEEE-754 standard. Typical floating-point
arithmetic IP cores are either pipe-lined or based on iterative algorithms such as CORDIC, all of
which require clocks to schedule the calculation. In our platform, no clock is provided for model
evaluations thus all arithmetics need to be executed in pure combinational logic. Taking
advantage of combinational logic allows all model evaluations to be 1) fast, the evaluation time
depends entirely on the propagating and settling time of signals, which is on the order of
microseconds, and 2) parallel, each model is evaluated on its own circuit without waiting for any
other results.
Our implementations of adder and multiplier are inspired by the open source project “Free
FloatingPoint Madness”, available at http://www.hmc.edu/chips/. Please contact the authors of
this paper if the modified code is needed.
Fast combinational floating-point division
Floating-point division is even more resource demanding than multiplications. We avoided
directly implementing the dividing algorithm by approximating it with additions and
multiplications. Our approach is inspired by an algorithm described in (Lomont, 2003), which
26
provides a good approximation of the inverse square root for any positive number x within one
Newton-Raphson iteration:
Q(x) can be implemented only using floating-point adders and multipliers. Thereby any division
with a positive divisor can be achieved by concatenating two blocks of Q(x):
This algorithm has been adjusted to also work with negative divisors (b < 0).
Numerical integrators for differential equations
Evaluating the instantaneous states of differential equation models require a fixed-step numerical
integrator. Euler’s Method was chosen to balance the numerical error and FPGA usage:
where T is the sampling interval. f(x, t) is the derivative function for state variable x.
2.2.6 Asynchronous spike-based communication between FPGA chips
Figure 2: Timing diagram of asynchronous spike-based communication
27
FPGA nodes are networked by transferring 1-bit binary spikes to each other. Our design allowed
the sender and the receiver to operate on independent clocks without having to synchronize. The
timing diagram of the spike-based communication is shown in fig. 2. The sender issues Spike on
with a pulse width of 1 ⁄ (365 × F
emu
) second. Each Spike then triggers a counting event on the
receiver, meanwhile each Clock first reads the accumulated spike count and cleans the counter
afterwards. Note that the phase difference between Spike and Clock is not predictable due to
asynchronicity. Although it is possible to lose a spike during the setup time of latching the
counter, it only loses a spike per clock cycle at most, which is effectively negligible considering
the nature of robustness in spike-modulated signal transmission.
2.2.7 Serialize neuron evaluations within a homogeneous population
Different neuron populations are instantiated as standalone circuits. Within in each population,
however, homogeneous neurons mentioned in Section 2.5 are evaluated in series in order to
optimize FPGA usage. Within each FPGA node all modules operate synchronously, meaning
that all updating events are triggered by a central clock. Therefore the maximal number of
neurons that can be serialized (N
serial
) per block is restrained by the following relationship:
Here F
fpga
is the fastest clock rate that a FPGA can operate on; C = 4 is the minimal clock cycles
needed for updating and storing each state variable in the block RAM; F
emu
= 1 kHz is the time
granularity of emulation (1 millisecond), and 365 × F
emu
represents 365x real-time. Consider that
Xilinx Spartan-6 FPGA devices peaks at 200MHz central clock frequency, the theoretical
maximum of neurons that can be serialized per neuron block is:
In the current design we choose N
serial
= 128.
2.3 Results: emulated activities of motor nervous system
Figure 3 shows pictures of a working FPGA node, two networked nodes and a screenshot of the
software front-end. Each FPGA node is able to emulate monosynaptic spinal loops consisting of
28
1,024 sensory neurons and 1,024 motor neurons (which is 8 blocks of neurons). The spike-based
asynchronous communication is successful between two FPGA nodes. Note that the emulation
has to be significantly slowed down for on-line plotting. When the emulation is at full speed
(365x real-time) the software front-end is not able to visualize the signals due to limited data
throughput.
Figure 3. The neural emulation platform in operation. Left: One working FPGA node. Center:
Two FPGA nodes networked using asynchronous spiking protocol. Right: Software front-end
displaying multi-scale signals.
The emulation platform successfully created multi-scale information when the muscle is
externally stretched (Figure 4A). We also tested if our emulated motor system is able to produce
the recruitment order and size principles observed in real physiological data. It has been well
known that when a voluntary motor command is sent to the motoneuron pool, the motor units are
recruited in an order that small ones get recruited first followed by the big ones (Henneman,
1957). The comparison between our results and real data are shown in Figure 4B, where the top
panel shows decoded motor unit activities from real human EMG (De Luca and Hostage, 2010),
and the bottom panel shows 20 motor unit activities emulated using our platform. No qualitative
difference was found.
2.4 Discussion and future work
We designed a hardware platform for emulating the multi-scale motor nervous activities in
hypertime. We managed to use one node of single Xilinx Spartan-6 FPGA to emulate
monosynaptic spinal loops consisting of 2,048 neurons, associated muscles and proprioceptors.
29
The neurons are organized as parallel pathways with sparse interconnections. The emulation is
successfully accelerated to 365x real-time. The platform can be scaled by networking multiple
FPGA nodes, which is enabled by an asynchronous spike-based communication protocol. The
emulated monosynaptic spinal loops are capable of producing reflex-like activities in response to
muscle stretch. Our results of motor unit recruitment order are compatible with the physiological
data collected in real human subjects. There is a question of whether this stochastic system turns
out chaotic, especially with accumulated errors from Backward Euler’s integrator. Note that the
firing property of a neuron population is usually stable (Sanger, 2011b) even with explicit noise,
and spindle inputs are updated by measurement continuously so the integrator errors are
corrected at every iteration. To our knowledge, the system is not critically sensitive to the initial
conditions or integrator errors. This question, however, is both interesting and important for in-
depth investigations in the future.
It has been shown (Raphael et al., 2010) that replicating classic types of spinal interneurons
(propriospinal, Ia-excitatory, Ia-inhibitory, Renshaw, etc.) is sufficient to produce stabilizing
responses and rapid reaching movement in a wrist. Our platform will introduce those
interneurons to describe the known spinal circuitry in further details. Physiological models will
also be refined as needed. For the purpose of modeling movement behavior or diseases,
Izhikevich model is a good balance between verisimilitude and computational cost. When testing
drug effects along disease progression, however, neuron models are expected to cover sufficient
molecular details including how neurotransmitters affect various ion channels. With the
advancing of programmable semiconductor technology, it is expected to upgrade our neuron
model to Hodgkin-Huxley’s. For the muscle models, Hill’s type of model does not fit the muscle
properties accurately enough when the muscle is being shortened. Alternative models will be
tested.
Other studies showed that the functional dexterity of human limbs – especially in the hands –
is critically enabled by the tendon configurations and joint geometry (Valero-Cuevas et al.,
2007). As a result, if our platform is used to understand whether known neurophysiology and
biomechanics are sufficient to produce able and pathological movements, it will be necessary to
use this platform to control human-like limbs. Since the emulation speed can be flexibly adjusted
30
from arbitrarily slow to 365x real-time, when speeded to exactly 1x real-time the platform will
function as a digital controller with 1kHz refresh rate.
The main purpose of the emulation is to learn how certain motor disorders progress during
childhood development. This first requires the platform to reproduce motor symptoms that are
compatible with clinical observations. For example it has been suggested that muscle spasticity
in rats is associated with decreased soma size of motoneurons (Brashear and Elovic, 2010),
which presumably reduced the firing threshold of neurons. Thus when lower firing threshold is
introduced to the emulated motoneuron pool, similar EMG patterns as in (Levin and Feldman,
1994) should be observed. It is also necessary for the symptoms to evolve with neural plasticity.
In the current version we presume that the structure of each component remains time invariant.
In the future work spike-timing-dependent plasticity (STDP) will be introduced such that all
components are subject to temporal modifications.
31
Figure 4: A) Physiological activity emulated by each model when the muscle is sinusoidally
stretched. B) Comparing the emulated motor unit recruitment order with real experimental data.
32
3 Study 1: What is the mechanism of childhood
secondary dystonia due to cerebral palsy?
Publication Citation:
W.J. Sohn, C.M. Niu, and T.D. Sanger, Increased long-latency reflex activity as a
sufficient explanation for childhood hypertonic dystonia: a neuromorphic emulation
study. Journal of neural engineering 12 (2015) 036010.
33
Increased long-latency reflex activity as a sufficient explanation
for childhood hypertonic dystonia: a neuromorphic emulation
study
Won J. Sohn
1†
, Chuanxin M. Niu
4†
, Terence D. Sanger
1,2,3*
† The authors equally contributed to the study
1
Department of Biomedical Engineering,
2
Biokinesiology, and
3
Neurology,
University of Southern California, 1042 Downey Way, Los Angeles, California,
90089
4
Department of Rehabilitation, Ruijin Hospital, School of Medicine, Shanghai
Jiao Tong University, Shanghai, China
* E-mail: tsanger@usc.edu
Abstract. Objective. Childhood dystonia is a movement disorder that interferes
with daily movements and can have a devastating effect on quality of life for
children and their families. Although injury to basal ganglia is associated with
dystonia, the neurophysiological mechanisms leading to the clinical
manifestations of dystonia are not understood. Previous work suggested that long-
latency stretch reflex (LLSR) is hyperactive in children with hypertonia due to
secondary dystonia. We hypothesize that abnormal activity in motor cortices may
cause an increase in the long-latency stretch reflex leading to hypertonia.
Approach. We modelled two possibilities of hyperactive LLSR by either creating
a tonic involuntary drive to cortex, or increasing the synaptic gain in cortical
neurons. Both models are emulated using programmable Very-Large-Scale-
Integrated-circuit (VLSI) hardware to test their sufficiency for producing dystonic
symptoms. The emulation includes a joint with two Hill-type muscles, realistic
muscle spindles, and 2,304 Izhikevich-type spiking neurons. The muscles are
regulated by a monosynaptic spinal pathway with 32ms delay and a long-latency
pathway with 64ms loop-delay representing transcortical/supra-spinal
connections. Main results. When the limb is passively stretched, both models
produce involuntary resistance with increased antagonist EMG responses similar
to human data; also the muscle relaxation is delayed similar to human data. Both
34
models predict reduced range of motion in voluntary movements. Significance.
Although our model is a highly simplified and limited representation of reflex
pathways, it shows that increased activity of the long-latency stretch reflex is by
itself sufficient to cause many of the features of hypertonic dystonia.
35
3.1 Introduction
Dystonia is an involuntary alteration in the pattern of muscle activation during voluntary
movement or maintenance of posture (Sanger et al., 2003). In secondary dystonia, symptoms are
often caused by injury to cortex, thalamus or basal ganglia (Colton et al., 2002; Sanger et al.,
2003; Breakefield et al., 2008), but the link between injury to these areas and the resulting
clinical symptoms remains unclear. Previous work suggested that the long-latency stretch reflex
(LLSR) is abnormally increased in childhood hypertonia due to secondary dystonia (Kukke and
Sanger, 2011). We do not know whether the elevated LLSR is a cause of dystonia or merely an
associated phenomenon. To explore this question, we test using simulation whether elevation of
LLSR in a highly simplified model is sufficient to cause features of hypertonic dystonia,
including resistance to passive stretch, delayed muscle relaxation, and reduced range of motion
in voluntary movement.
The eventual manifestation of secondary dystonia may be attributable to increased activity in
the motor cortex. Brain imaging provides the direct evidence of increased motor cortical activity
in secondary dystonia (Ceballos-Baumann, 1994); other studies using transcranial magnetic
stimulation (TMS) over motor cortex also show increased corticospinal excitability (Trompetto
et al., 2012; Kojovic et al., 2013). In other forms of dystonia, it was found that patients exhibit
reduced intracortical inhibition (Ridding et al., 1995; Edwards et al., 2003; Quartarone et al.,
2003; Prescott et al., 2013) and increased cortical plasticity (Quartarone et al., 2003; Edwards et
al., 2006; Weise et al., 2006; Prescott et al., 2013). The association between increased motor
cortex activity and secondary dystonia could be due to an inhibitory effect of basal ganglia over
motor cortex, possibly through thalamocortical pathways (DeLong and Wichmann, 2007;
Hallett, 2011). This association is supported by clinical treatments for patients of dystonia, where
dystonic symptoms were alleviated after ablative surgeries (Imer et al., 2005; Hashimoto et al.,
2010) or deep brain stimulations (McIntyre et al., 2004; Vidailhet et al., 2005; Krauss, 2010; Air
et al., 2011) in the globus pallidus internus (GPi). Therefore in this emulation study we focus on
how different cortical parameters lead to abnormal reflex behaviors similar to dystonia.
One possible outcome of increased motor cortex excitability is to elevate LLSR due to the
role of primary motor cortex in reflex modulation (Evarts and Tanji, 1976; Lee et al., 1983;
Morimoto et al., 1984; Capaday et al., 1991; Matthews, 1991; Palmer and Ashby, 1992;
36
Pruszynski et al., 2011). Therefore loss of inhibition in the motor cortex may leak an
uncontrolled drive that either lowers the threshold of cortical cells or amplifies their response to
afferent input. In both cases the activity of LLSR is expected to increase. In our previous studies,
children with secondary dystonia showed increased reflex activity (van Doornik et al., 2009)
with long-latency responses (Kukke and Sanger, 2011), which could potentially cause hypertonia
in this population. Taken together, existing evidence suggests that the hypertonic manifestation
of secondary dystonia may be directly caused by elevated LLSR, which potentially results from
many insults including increased motor cortex excitability after injuries in basal ganglia,
cerebellum, thalamus, or sensory cortices.
To determine whether this is a plausible mechanism for hypertonic dystonia and not merely
an epiphenomenon, we emulate the effect of increased cortical drive or increased afferent input
to cortex. The term “emulation” is used to disambiguate from numerical simulations (usually
slower than real-time) in software. We do not include a model of the basal ganglia, because we
seek to test whether any structure projecting to and causing uncontrolled firing of motor cortical
areas could potentially be a cause of dystonia. Other possible areas with oligosynaptic
connectivity to motor cortices include prefrontal cortex, primary sensory cortex, thalamus, and
cerebellum. We choose the synthetic analysis approach primarily because it allows flexible ways
of introducing abnormalities that are physiologically plausible but difficult to obtain from human
subjects in laboratory. It also provides information about physiological components at
drastically different physical and temporal scales, including millisecond time-scale action
potentials in microscopic neurons, and seconds-long contractions in whole muscles. We
leverage the recently available technology of programmable Very-Large-Scale-Integrated-circuit
(VLSI), which allows us to create emulations of neurons that communicate using spikes, with the
ability to increase the number of emulated neurons without sacrificing speed. With VLSI, the
neural circuitry and connectivity is also easily modifiable. In this study, we first built a small set
of structures to create a non-impaired system with functioning reflexes including a short-latency
loop representing the spinal monosynaptic reflex pathway, and a long-latency loop representing
the supra-spinal/transcortical pathway. Due to the aforementioned contribution of motor cortex
to long-latency response, we selectively increase the activity of the transcortical pathway to
emulate a hyperactive LLSR. We hypothesize two possible causes of hyperactive LLSR. First,
the cortical neurons may receive a tonic drive that is either sub- or supra-threshold but overall
37
depolarizing. This TONIC model makes the cortical neurons easier to fire or achieve high firing
rate, even when receiving the same level of excitatory post-synaptic current (EPSC) from
sensory feedback. Second, the synaptic gain of cortical neurons may uniformly increase, which
augments the excitability of cortical neurons. This HI-GAIN model amplifies the EPSC provided
by ascending sensory feedback, which eventually elevates the overall activity of the transcortical
pathway. These two mechanisms are the major categories of abnormality that can lead to
increases in LLSR at the cellular level, so we modeled both. We argue it is important to test if
different mechanisms are both sufficient to produce dystonia, which may eventually help sub-
categorizing secondary dystonia. In both models, the spinal pathway remains intact and therefore
only the long-latency component in the reflex pathway is elevated.
We focus on changes in EMG or movement kinematics by comparing both TONIC and HI-
GAIN models with the non-impaired condition. There are three experiments in this study:
1. passive back-and-forth stretch
2. voluntary relaxation of force
3. voluntary back-and-forth movement
In the first two experiments, data from human subjects are available and thus compared to
verify the sufficiency of our model for producing dystonia; human data are not yet available for
the last experiment, therefore the results can be used as testable hypotheses for future
experiments.
3.2.Materials and methods
We focus on using spike-based emulation to determine the functional role of sensorimotor
components, especially their sufficiency for causing clinical symptoms in abnormal conditions.
The hardware emulation of the spiking neurons and sensorimotor components is constructed
using Field Programmable Gate Arrays (FPGA, Xilinx Spartan-6), a programmable version of
VLSI electronic chips. We favor FPGAs over pipelined hardware such as GPUs (Graphic
Processing Units) or clustered CPUs (Central Processing Units) due to their inherent parallelism
that resembles neural circuitry. We also found that when networking multiple units for large-
scale disease emulation, FPGAs allow significantly more flexibility than custom-built hardware
38
for communication protocols such as neuromorphic transmission protocols that directly transmit
neuron-like spikes.
The activity of the emulated sensorimotor system is recorded using a dedicated data-logging
computer. The FPGA communicates with the data-logging computer through a high-speed USB
channel and the OpalKelly development kit and interface software (XEM6010, OpalKelly Inc.).
The technical details can be located in the previous study (Niu et al., 2014). The biomechanics of
the limb is simulated in software. The biomechanical simulation updates its states by first polling
muscle force from the FPGAs, followed by sending muscle kinematic variables (length and
velocity of lengthening) back to the FPGAs. This hybrid setup containing both hardware and
software is slower than pure hardware emulation, but it simplifies the coordination between
flexor and extensor FPGA chips.
3.2 Models for the sensorimotor system and increased LLSR
We study a sensorimotor system that includes a limb joint comprising two opposing
monoarticular muscles (flexor and extensor), muscle spindles, spindle afferents, alpha-
motoneurons and associated supraspinal structures (Figure 5A). The hardware emulates parallel
proprioceptive pathways including monosynaptic connections with 32ms loop-delay representing
the spinal proprioceptive pathway, and oligosynaptic pathways with 64ms loop-delay
representing the supra-spinal/transcortical components of the stretch reflex loop. The delays were
chosen both to approximate known conduction delays and also for efficiency of hardware
emulation. Our previous work suggests that in sensorimotor systems, the statistical effect of two
fully connected neuron populations is equivalent to ones that are only sparsely connected
(Sanger, 2011a), therefore we connect the spindle afferents to motoneurons using parallel
connections instead of implementing the full connectivity among neurons. In this study we do
not introduce inhibitory mechanisms such as Renshaw cells or reciprocal inhibition. Both gamma
dynamic and gamma static drive are set to 80Hz, which is a moderate intensity of fusimotor
stimulus in the classic experiment chosen for modeling the muscle spindle (Emonet-Denand et
al., 1977). No alpha-gamma coactivation is introduced. This provides a baseline system for the
non-impaired behavior before introducing the disease model.
On top of the non-impaired system, we model two possible mechanisms to increase LLSR:
the TONIC model and HI-GAIN model. In the TONIC model (Figure 5B), we superimpose a
39
tonic input on the voluntary commands via a depolarizing synapse. The tonic input lowers the
threshold of the cortical neuron pool and therefore facilitates the transcortical pathway when
stimulated by proprioceptive feedback. In the HI-GAIN model (Figure 5C), the synaptic gain of
afferent inputs to the cortical neuron pool is increased by augmenting the excitatory postsynaptic
potential (EPSC) in response to afferent input. Therefore the TONIC model is an additive
excitatory drive that is present even in the absence of afferent input, while the HI-GAIN model is
a multiplicative drive that is present only when afferent input is present. Both TONIC and HI-
GAIN models increase the excitability of cortical neuron pools to afferent input, either by
lowering threshold or by increasing the effect of the input. We test both models because either
or both mechanisms could be active in childhood dystonia.
40
A) Physiology
B) TONIC model
C) HI-GAIN model
flexor
extensor
spindle
CN
αMN
αMN
involuntary
cortical
drive
afferented flexor
afferented extensor
excessive
LLSR gain
CN = Cortical Neuron Pool
α MN = alpha-motoneuron
CN
CN
αMN
αMN
voluntary
command
afferented flexor
afferented extensor
CN
G
G
cortical region
cortical region
voluntary
command
motoneuron
41
Figure 5. A) Components of human sensorimotor system model. The system includes a limb
joint comprising two opposing monoarticular muscles (flexor and extensor), muscle spindles,
spindle afferents, alpha-motoneurons, and associated spinal and supraspinal structures. Our
model includes a monosynaptic reflex arc with a 32ms loop-delay and a supra-spinal
transcortical reflex pathway with a 64ms loop-delay. B) The TONIC model of dystonia. The
dashed box shows the procedure of introducing involuntary tonic activity. Before the disease
onset (unshaded area), the system executes the voluntary commands received from
antagonistic cortical inputs; after the disease onset (shaded area), the voluntary descending
commands are superimposed on a tonic input. C) The HI-GAIN model of dystonia. The
synaptic gain associated with the afferent cortical projection becomes higher after the disease
onset (shaded areas), which directly increases the loop-gain of the transcortical feedback loop.
3.2.1 Implementation of sensorimotor system on hardware
An adult elbow joint is simulated in software as a 34cm beam freely rotating around a single
D.O.F. axis, the mass of the beam is 1.52kg as an approximation to a human fore arm (Scheidt
and Ghez, 2007); the moment arm of muscle is simplified as constant 30mm, which is in the
middle range of moment arm measured from human biceps (Murray et al., 1995). The joint is
driven by a pair of antagonistic muscles following Hill-type model (Hill, 1938). The
sensorimotor components that interact with the simulated joint are arranged on FPGA hardware
as shown in Figure 6. Each muscle is controlled by 128 spindles as modeled by Mileusnic et al.
(2006a), simple spiking neurons developed by Izhikevich (Izhikevich, 2003a), and a motor-unit
action potential (MUAP) model similar as (Fuglevand et al., 1993a; Rodriguez et al., 2007;
Krutki et al., 2008) producing a surface electromyogram. Izhikevich neurons are used because
they permit use of biologically realistic variables including transmembrane currents, yet they can
be implemented much more efficiently in hardware than the more complex Hodgkin-Huxley
equations that they approximate. A total of 768 alpha-motoneurons were divided into 6 groups
representing motor units with various sizes, so that the size principle (Henneman et al., 1965) is
present when motor units are recruited. The parameters are tuned such that at maximal spiking
rate of the alpha-motoneuron pool, the muscle exerts approximately 5N tangential force at the tip
of the joint. In the transcortical loop, the spindle afferent information travels through a
42
population of 128 cortical neurons representing the primary sensory and motor cortices. The
main focus of modeling cortex is to enable a longer loop-delay compared to spinal pathways,
therefore we accept a pool of 128 neurons as a model of cortex even though it is a clear
oversimplification of the real biology. Due to the limited capacity of each FPGA unit, the system
must be distributed on multiple FPGAs as indicated by the blocks. The entire system uses 6
interconnected FPGA boards to emulate 2 muscles and 2,304 spiking neurons.
No. 1
2
3
2
3
No. 1
Motor FPGA
Spinal-cord FPGA
Cortex FPGA
2
3
...
128
Spikes
Digital signals
Izhikevich neuron
Ƞ Noise
Ia
II
No. 1
...
128
...
128
Ƞ
Σ
Spindle
gamma
static
gamma
dynamic
Σ Σ
Flexor
Limb
Coupling
Ƞ
Ƞ
Ƞ
Ƞ
Ƞ
Ƞ
Ƞ
Ƞ
Flexor length and velocity
TONIC
HIGAIN
43
Figure 6. Detailed configuration of the motor nervous system on neuromorphic hardware.
The components are implemented separately for the flexor and extensor, which
simultaneously drive the joint modeled as a beam freely rotating around the endpoint
(simulated in software outside FPGAs). For each muscle (flexor or extensor), the muscle
force is calculated from a Hill-type muscle model activated by 6 motoneuron pools with 6
different motoneuron sizes, each pool comprising of 128 identical motoneurons modeled by
Izhikevich (Izhikevich, 2003a). The motoneuron pools receive excitatory input from both
the spinal loop and transcortical loop. In the spinal loop, the sensory feedback is provided
by muscle spindles implemented as Mileusnic and colleagues did (Mileusnic et al., 2006a),
which include both the Primary (Ia) and Secondary (II) afferents to provide the dynamic and
static proprioceptive information about the muscle. A total of 128 spindles are implemented
for each muscle, thus providing 256 independently spiking afferents. In the transcortical
loop, the spindle afferents synapse on a population of 128 cortical neurons representing the
primary sensory and motor cortex. The cortical part is clearly an oversimplification of the
cortex but it enables an additional 32ms conduction delay in the proprioceptive feedback
loop, which is the main focus of this study rather than modeling a full cortex. In the cortex,
spindle afferents may join additional inputs modeling the voluntary motor command. The
TONIC model is implemented by adding to the cortical drive a tonic component that is
independent from the voluntary drive or the sensory feedback. The HI-GAIN model is
implemented by increasing the synaptic gain of the spindle afferents prior to activating the
cortical neuron pool. Due to the limited capacity of each FPGA unit, the system must be
distributed on multiple FPGAs as indicated by the blocks. The entire system uses 6 FPGA
boards enabling 2 muscles with 2,304 neurons. Only half of the system (flexor) is shown.
3.2.2 Experiments
We have verified the sufficiency of both the TONIC and HI-GAIN models for dystonic
symptoms in the following three experiments:
44
3.2.2.1 Experiment 1: Passive back-and-forth stretch
We replicate the experiment where a sinusoidal perturbation was applied to a joint of a child with
hypertonic arm dystonia (van Doornik et al., 2009). In the original study, the subject was
instructed to rest, and the right arm was rotated manually by the experimenter following an
approximately sinusoidal time profile with frequency varying from 0.2 Hz up to 2 Hz.
We use the same waveform as in the original study to stretch the emulated system. A pre-
recorded waveform of joint displacement is applied to the software-simulated joint. We keep the
voluntary command at zero during the passive stretch to capture the fact the subject was at rest. It
was found that patients with dystonia failed to relax the muscles and produced cyclic EMG
responses to the stretch. The models of dystonia are validated based on their abilities to
qualitatively explain human data. To this purpose, we calculate the phase angle between the
model-generated EMG and the joint angle using the same approach as in van Doornik et al.
(2009). The range of phase angle reflects which kinematic variable could be the main cause of
the EMG response. For example, a phase angle of 0° means the EMG response is in line with the
joint angle, thus making the EMG position-dependent; while -90° means the EMG response is
velocity-dependent; any phase angle between -180° and -90° suggests that the EMG response has
acceleration-dependent components; any phase angle between 0° and 180° means that the EMG
is active during muscle shortening.
During model validation, we first introduce a small TONIC input (TONIC model) to the
cortical neuron pool or a small cortical gain (HI-GAIN model) sufficient to produce dystonic
symptoms, i.e. an observable phase angle. Then we progressively increase the intensity of
dystonia until the phase angle stopped changing significantly, mainly due to the saturation of
neurons. We consider a model sufficient to explain human data if the parameter-scan produces a
wide range of phase angle that includes those from human patients.
3.2.2.2 Experiment 2: Voluntary relaxation of muscle
One of the common manifestations of childhood dystonia is that the voluntary relaxation of
muscle contraction is delayed (Sanger et al., 2003). We replicate the experiment that
documented this delay by comparing normal subjects to patients with secondary dystonia (Ghez
et al., 1988).
45
Both muscles were first activated at 50% maximum voluntary contraction (MVC), followed
by a step decrease in the voluntary command. This experiment simulated when the subject
attempts to rapidly relax the co-contracted muscles following a cue signal (Ghez et al., 1988). It
was shown that the EMG decreased more slowly in patients with secondary dystonia, and thus
the overall duration of muscle relaxation was delayed.
We quantify the rate of muscle relaxation by fitting the filtered EMG to a sigmoid function
defined below:
𝐸𝑀𝐺
!"#$%"&
𝑡 =𝐸𝑀𝐺
!"!
+
(𝐸𝑀𝐺
!"#
−𝐸𝑀𝐺
!"#
)
1+𝑒
!"
where the free parameter 𝜏 denotes the rate at which the EMG decreases during muscle
relaxation. The relationship between 𝜏 and the severity of dystonia was tested using a similar
parameter-scan as in Experiment 1.
3.2.2.3 Experiment 3: Reduced range of motion in voluntary movements
In clinic, it is commonly observed that patients with dystonia apply great efforts in order to
achieve a normal range of movement; otherwise the range of motion in voluntary movements is
reduced. These phenomena are, however, not well documented in experimental studies. It is
reported that the range of motion is reduced in neck and knee for patients with primary dystonia
(Carpaneto et al., 2004; Lebiedowska et al., 2004). We hypothesize that similar reduction in
range of motion is likely to occur in secondary dystonia. In addition, we test if the reduced range
of motion can be improved by amplifying the voluntary command, which is a straightforward
engineering approach to compensate a less responsive system.
The two cortical neuron pools regulating flexor and extensor receive half-wave rectified
sinusoidal waveforms (180 degrees out of phase), which produces a back-and-forth joint
swinging movement. The peak-to-peak amplitudes of joint angle resulting from these inputs are
analyzed. If the amplitude of joint angle is reduced in dystonia models, we linearly increase the
sinusoidal voluntary commands to test if the reduction can be compensated.
46
3.2.2.4 Data acquisition and processing
All data are sampled at 1kHz. The EMG signals are first high-pass filtered using 10Hz cut-off
frequency (Butterworth, 3
rd
order), followed by rectification and a low-pass filter at 120Hz cut-
off frequency (Butterworth, 3
rd
order). Phase angle analysis and the nonlinear fitting of the
sigmoid function (for quantifying muscle relaxation, explained below) are performed using
Matlab (Mathworks, Inc.).
3.3 Results
3.3.1 Non-impaired stretch reflex and hypertonia with increased LLSR
We first verify the quality of emulation by passively stretching the joint and monitoring the
reflex behavior elicited by the stretch. The joint was passively extended by 45 degrees within 0.2
seconds from the software interface. The sensorimotor information recorded in response to a
virtual stretch in the non-impaired condition is shown in Figure 7A, including the spindle
afferents (group Ia and II), motoneuron rasters, muscle force and EMG. As can be seen, the
hardware emulation is capable of producing concurrent multi-scale information during a virtual
behavior.
We further compared the stretch reflex between the non-impaired condition and our two
dystonia models. The joint was briefly stretched using a short torque pulse (5N for 20ms, Figure
7B), which extended the flexor by approximately 40% of L
0
(the resting length of muscle).
Hypertonia can be seen from the increased flexor force in the dystonia conditions compared to
the non-impaired condition (Figure 7B, Muscle force). The EMG response is divided into R1
(30-50ms), R2 (50-80ms), and R3 (80-100ms) regions representing the short-, long-, and longer-
latency responses. By overlapping the non-impaired condition and our models of dystonia
(Figure 7B, EMG), it can be seen that R2 responses are increased by both TONIC and HI-GAIN
model, i.e. LLSR is increased in both models.
47
Figure 7. The emulated stretch reflex in the non-impaired condition and models of
increased LLSR. A) EMG responses to a stretch-and-hold perturbation in the non-
impaired condition. The joint flexor was stretched by approximately 40% of L
0
. The
emulated EMG showed a burst in response to the stretch. The motoneuron raster showed
patterns compatible with both the early burst in EMG response and the subsequent tonic
components. B) Emulated hypertonia and increased LLSR. The muscle force increases
from baseline in both TONIC and HI-GAIN model, which implemented the increased
muscle resistance to passive muscle stretch commonly observed in hypertonic dystonia.
EMG responses to a torque pulse perturbation (5N for 20ms) in both the non-impaired
condition and the models of dystonia. R1 (30-50ms), R2 (50-80ms), and R3 (80-100ms)
regions represent the short-, medium-, and long-latency responses. In both TONIC and
HI-GAIN model the R2 response was increased compared with the non-impaired
condition.
A)
Muscle length
EMG (a.u.)
Muscle force
Motoneuron
raster
B)
0 0.05 0. 1 0.15
Perturbation
Torque
EMG (a.u.)
R1 R1 R2 R2 R3 R3
EMG (a.u.)
Muscle force
Muscle force
TONIC model TONIC model
Non-impaired Non-impaired
R1 R1 R2 R2 R3 R3
sec
0.2 sec
0.3L
0
5N
50% F
max
Spindle afferents Ia
Spindle afferents II
HI-GAIN model HI-GAIN model
Non-impaired Non-impaired
48
3.3.2 Experiment 1: Involuntary responses to passive joint stretch
The dystonic subjects showed phasic EMG responses to the manual sinusoidal stretch in their
biceps (Fig.8A). In the emulation, the virtual arm was passively rotated with the identical
waveform as shown in Fig.8 A. Our emulated result showed that in both TONIC (Fig.8B) and
HI-GAIN (Fig.8C) models, the flexor EMG activity modulates with the joint angle similarly to
the original human experiment.
49
Elbow Angle
Elbow Angle
Emulated Flexor EMG
Biceps EMG
Emulated Flexor EMG
Time (s)
Angle[°]
0 80 20 40 60
Angle [°]
Amplitude Angle[°] Amplitude
Amplitude
50
-50
50
-50
0
1
1
A) Human data
B) TONIC model
0
C) HI-GAIN model
Elbow Angle
50
Figure 8. Biceps EMG during arm rotation in a child with hypertonic arm dystonia (A)
created based on the data in van Doornik et al. (2009). When the subject was instructed to
rest, the right arm was rotated manually by the experimenter following approximately a
sinusoidal time profile with frequency varying from 0.2 Hz up to 2 Hz. The biceps
showed phasic EMG responses to the manual stretch. In the emulation, the virtual joint
was passively rotated with the identical waveform under two models of dystonia. The
voluntary command is set to zero to represent the subject being “at rest”. In the TONIC
model (B), the EMG is not silent when the joint is passively rotated and shows similar
phasic patterns as the human EMG recording. In the HI-GAIN model (C), similar non-
silent EMG responses are observed with increased magnitude in phasic response.
The 8 patients with dystonia from human experiments (van Doornik et al., 2009) showed a
wide range of phase angle between -101° and 17° (Table 2, van Doornik et al., 2009). There
were 2 patients (Subjects 4 and 5) with dyskinetic cerebral palsy (CP) who showed positive
phase angles, suggesting EMG responses during muscle shortening (a “shortening reaction”
more frequently seen in Parkinsonism). These 2 patients were excluded from the model
validation, since their mechanisms were unlikely the same as the other 6 patients who were non-
dyskinetic. As can be seen from Fig.9, 5 out of 6 patients showed phase angles between -90° and
0° (combination of position- and velocity-dependent); 1 out of 6 patients showed a joint angle
below -90° (with acceleration-dependent components).
We validate our models of dystonia by testing whether the reported phase angles in human
can be achieved in emulation. In the TONIC model, we scan the tonic EPSC with fixed
increments from 40pA up to 280pA. We stop at 280pA since further increases did not
significantly change the phase angle mainly due to the saturation of cortical neurons. The scan of
tonic EPSC produces phase angles from -32°to -53°, which explains only 1 out of 6 patients
(dark gray area, Fig.9A). In the HI-GAIN model, we start the scan by selecting a unit gain that
amplifies the non-impaired EPSC by 4 times, then we progressively increase the gain to 19 times
of a unit. We stop at the gain of 19 for the same reason of saturation. This process produces
phase angles from -29° to -147°, which covers the reported phase angles of all 6 subjects (dark
gray area, Fig.9B).
51
We infer the upper and lower limits of phase angle by setting the emulation to boundary
conditions (light gray area in Fig.9). The upper limit of phase angle is obtained by de-afferenting
the secondary fibers (group II) and keeping the loop gain high; and the lower limit is obtained by
de-afferenting the primary fibers (group Ia) and keeping the loop gain low. The boundary
conditions are constructed based on the factors that may affect the phase shift, e.g. loop delay,
loop gain, and the relative contribution between position- and velocity-dependent components,
etc. See Discussion for further considerations. The overall range of phase angle (-9° to -169°) is
larger than both TONIC and HI-GAIN models.
Figure 9. The phase angle calculated from emulation and human data. The thick lines
show the phase angles of 6 patients with hypertonic arm dystonia. (A) The TONIC model
produces phase angles from -32° to -53° (dark gray area), which includes 1 out of 6
patients. (B) The HI-GAIN model produces phase angles from -29° to -147° (dark gray
area), which includes all 6 patients. The boundary of emulation is tested by de-afferenting
either the primary (group Ia) or the secondary (group II) spindle fiber, resulting in the
minimum phase angle of -169° and the maximum of -9° (light gray area).
The sensitivity of phase angle to the intensity of emulated dystonia in each model is shown in
Fig.10. In the TONIC model, the phase angle decreases with increased tonic current input
A) TONIC model
S7: −101°
S3: -79°
S8: −68°
S1: −57°
S6: −51°
S2: −30°
0°
180°
-180°
-90°
90°
0°
180°
-180°
-90°
90°
B) HI-GAIN model
S7: −101°
S3: -79°
S8: −68°
S1: −57°
S6: −51°
S2: −30°
Model coverage
Boundary of emulation
Individual patients
(van Doornik et al. 2009)
52
following a significant linear correlation (Fig.10A, slope=−0.074,p< 0.0001,r
!
= 0.59, 4
repetition for each level of TONIC input). In the HI-GAIN model, similar linear correlation is
observed (Fig.10B, slope=−6.50,p< 0.0001,r
!
= 0.78, 4 repetitions for each level of HI-
GAIN). Overall, the HI-GAIN model can capture more variance in phase angle than could the
TONIC model. The results suggest that in terms of phase angle, the HI-GAIN model can explain
more data than the TONIC model, although both models produce results qualitatively similar to
the human data.
A) TONIC model B) HI-GAIN model
TONIC input (pA)
40 80 120 160 200 240 280 1 4 7 10 13 16 19
GAIN
Phase angle (deg)
-10
-180
-10
-180
53
Figure 10. The sensitivity of phase angle to the intensity of emulated dystonia in
each model. In the TONIC model, we scan the TONIC input from 40pA to 280pA
with fixed increments. The average phase angles produced by these TONIC inputs
range from -32° to -53°, where a higher TONIC input significantly reduces the
phase angle (slope=−0.074,p< 0.0001,r
!
= 0.59, 4 repetition for each level of
TONIC input). The linear trend tends to saturate when the TONIC input exceeds
240pA. In the HI-GAIN model, we start the scan of cortical gain by selecting a unit
gain that amplifies the non-impaired EPSC by 4 times. The gain is increased to 19
times of a unit with fixed increments. The average phase angles produced by the
HI-GAIN model range from -29° to -147°, where a higher cortical gain
significantly reduces the phase angle (slope=−6.50,p< 0.0001,r
!
= 0.78, 4
repetitions for each level of HI-GAIN). The linear trend tends to saturate when the
TONIC input exceeds 16 units. These results suggest that in terms of phase angle,
the HI-GAIN model can explain more human data than the TONIC model.
3.3.3 Experiment 2: Delayed relaxation of muscle force
Fig.11 shows that both the TONIC and HI-GAIN models suffice to delay muscle relaxation
compared with the non-impaired condition. The phenomenon is qualitatively similar to human
data.
54
Figure 11. Relaxation of muscle activity from a state of co-
contraction in biceps and triceps in a normal subject and a
patient with secondary dystonia (A), created based on the data in
Ghez et al. (1988). Subjects were first required to maximally co-
contract the biceps and triceps muscles and upon a visual cue
relax the muscle as fast as possible. The duration for muscle
relaxation in the patient with secondary dystonia (right) was
elongated compared to the normal subject (left). In emulation
(B), both TONIC and HI-GAIN models showed delayed EMG
relaxation compared with the non-impaired condition, when the
voluntary descending command that co-contracts biceps and
triceps were abruptly shut off. Single representative trials are
shown.
HI-GAIN
TONIC
200 ms
Biceps EMG
Triceps EMG
Biceps EMG
Triceps EMG
Non-impaired
Secondary Dystonia Normal
A) Human data
B) TONIC and HI-GAIN model
55
The rate of relaxation (𝜏) was calculated both from human data and emulated results for
model validation. Due to the limited human data, we could only obtain two values of 𝜏 based on
the original EMG time series in Ghez et al. (1988). The 𝜏 calculated from the normal subject
almost tripled that from the patient with secondary dystonia (𝜏
!"#$%&
= 0.044,𝜏
!"#$%&'(
=
0.015), which is compatible with the visual pattern of delayed relaxation (Fig.11A). These two
values of 𝜏 are plotted in Fig.12 (dotted and dashed lines) for comparison with emulated 𝜏.
In the TONIC model, the level of TONIC input was scanned using the same set-up as in
Experiment 1. The emulated 𝜏s were all lower than 𝜏
!"#$%!
(Fig.12A), and 𝜏
!"#$%&'(
was
included in the emulated range of 𝜏. When the TONIC input increases, the delay of relaxation is
significantly longer (𝑠𝑙𝑜𝑝𝑒=−0.00012, p< 0.00001, r
!
= 0.65, 4 repetitions for each level
of TONIC input). In the HI-GAIN model, the emulated 𝜏s were also lower than 𝜏
!"#$%&
(Fig.12B). But the linear correlation between 𝜏 and model intensity is much weaker, represented
by modest significance and lower r
!
value (Fig.12B, 𝑠𝑙𝑜𝑝𝑒=−0.00051, p= 0.046, r
!
= 0.11,
4 repetitions for each level of HI-GAIN). In contrast to Experiment 1, results from Experiment 2
suggest that the TONIC model can explain more data than the HI-GAIN model.
A) TONIC model B) HI-GAIN model
TONIC input (pA) GAIN
0.00 0.02 0.04 0.06
τ
Normal
τ
τ
Dystonia
147 10 13 16 19 40 80 120 160 200 240 280
56
Figure 12. Relationship between the rate of relaxation (𝜏) and the intensity of either the
TONIC or the HI-GAIN model. In both models, the intensity of dystonia is scanned the
same as in Experiment 1. Two reference αs calculated from human data are shown, the
rate in the normal subject (𝜏
!"#$%&
= 0.044, dashed line) represents faster relaxation
compared to the patient with secondary dystonia (𝜏
!"#$%&'(
= 0.015, dotted line). In the
TONIC model (A), the emulated 𝜏s are all lower than 𝜏
!"#$%&
, and their range includes
𝜏
!"#$%&'(
. When the TONIC input increases, the delay of relaxation is significantly
longer (slope=−0.00012, p< 0.00001, r
!
= 0.65, 4 repetitions for each level of
TONIC input). In the HI-GAIN model, the emulated 𝜏s are also lower than 𝜏
!"#$%&
(B).
But the linear correlation between 𝜏 and model intensity is much weaker, represented by
modest significance and lower r
!
value (slope=−0.00051, p= 0.046, r
!
= 0.11, 4
repetitions for each level of HI-GAIN).
3.3.4 Experiment 3: Reduced range of motion in voluntary movements
In this experiment, we aimed to predict how dystonia affects movement kinematics. In the
TONIC model, just adding a middle level TONIC input is sufficient to reduce the peak-to-peak
amplitude of voluntary movement (Fig.13). The reduced range of motion can almost fully
recover if we amplify the voluntary commands, suggesting that patients can compensate the
reduced range of motion by increasing voluntary effort. Similar results are produced using the
HI-GAIN model. In this experiment, both models predicted similar kinematic consequences
using anecdotal parameters. More human data are required to distinguish between these two
models.
joint angle
(deg)
1 sec
0
30
-30
flexor EMG
(a.u.)
extensor EMG
(a.u.)
21% decrease
Non-impaired TONIC
No compensation
TONIC
67% compensation
HIGAIN
No compensation
HIGAIN
100% compensation
56% decrease
57
Figure 13. Dystonia models predict that the range of motion in voluntary arm-swing
movement should be reduced in dystonia; with certain level of compensation by
increasing the voluntary command, the range of movement can be recovered. In the non-
impaired condition, the joint was voluntarily swinging at 1Hz with approximately 50°
range of movement; the muscles are activated using alternating descending commands at
1Hz as depicted in Fig. and thus generate alternating EMG bursts. In the TONIC model,
we maintained the level and frequency of the 1Hz voluntary command, but added a tonic
input with 80% of the peak-to-peak magnitude of the voluntary command. This reduces
the movement range by 21%, which can be recovered by compensating the voluntary
command by 67% of its original magnitude. In the HI-GAIN model, a 4.7 times increase
in the synaptic gain of cortical neuron pool decreased the movement range by 56%,
which can be recovered by compensating the voluntary command by 100%.
3.4 Discussion
Using our recently developed technique of neuromorphic emulation in hardware, we tested the
hypothesis that increased long-latency stretch reflexes, created by excessive activity in the motor
cortex, are sufficient to induce forces and EMGs with similar patterns to those seen in patients
with secondary dystonia. In particular, we verified that when the limb is passively stretched as in
van Doornik et al. (2009), both an additive tonic input in the cortical neuron pool (TONIC
model) and an elevated synaptic gain in the motor cortex (HI-GAIN model) suffice to induce
EMG responses in the absence of a voluntary command. The HI-GAIN model explains a wider
range of phase angle from human data than does the TONIC model. Furthermore, we verified
that both TONIC and HI-GAIN models suffice to delay muscle relaxation similar to the results
of Ghez et al. (1988). The rate of relaxation could be better explained by TONIC model than the
HI-GAIN model. Our models also predict that the range of movement should be reduced if the
magnitude of the voluntary command remains the same in dystonia. Alternatively, we predict
that dystonia increases the voluntary effort required to make movements of the same magnitude.
3.4.1 Relationship between LLSR and secondary dystonia
Our results suggest that increased LLSR may be an intermediate mechanism linking brain
injuries and the clinical manifestations of hypertonic dystonia. That is, even though the injuries
58
may occur in various parts of the brain (e.g. thalamus, basal ganglia, cortex, etc.), all these
injuries could produce hypertonic dystonia due to excess activity in the motor cortex via elevated
LLSR. We show that both TONIC and HI-GAIN models are sufficient for explaining the
dystonic symptoms in the demonstrated cases. Nevertheless, the manifestation of these two
models could differ when the severity of dystonia increases (Fig.10); it is also suggested that
hypertonia due to tonic input may have a stronger effect than high synaptic gain on the delay in
muscle relaxation (Fig.12). In theory, the TONIC model can be interpreted as a non-intrinsic
abnormal drive to the motor cortex, originated from other parts of the brain; while the HI-GAIN
model suggests that there is some intrinsic deficit in the motor cortex that amplifies its overall
activity. These differences suggest testable hypotheses in future studies to distinguish the
heterogeneous causes of dystonia.
Elevated LLSR is not exclusively observed in dystonia but also reported in other disorders
such as rigidity. It is possible that dystonia and rigidity may share similarities in mechanism,
although due to very different causes. It was not our intent to study rigidity, which is rare in
children, but from our prior work (van Doornik et al., 2009) we have noted that even within
dystonia there appears to be a continuum between spring-like hypertonia and viscous hypertonia.
Rigidity is primarily viscous and it is tempting to conjecture that rigidity may result from
increased LLSR involving primarily velocity-dependent (rapidly adapting) afferents, whereas
dystonia may result from elevated LLSR involving primarily position-dependent (slowly
adapting) afferents.
3.4.2 Rationale of model validation
When validating our models against human data, we focused on whether varying a single
parameter in the model could produce a wide range of outcome measures compatible with the
human data. We did not, however, focus on the exact matching between emulated time series and
human data. This is because dystonia has been known as a disease with high inter-patient
variability, and data from humans are limited. As a result, we argue it is not that useful if the
model aims at matching exactly to the physiological signals of human patients. Take Experiment
1 as an example, the power of the HI-GAIN model comes from its ability to explain the phase
angle of any of the 6 patients using the same disease structure. Given more data from patients
59
with secondary dystonia, the range of outcome measures also provides an experimentally testable
criterion to distinguish different models.
3.4.3 Thoughts on the two mechanisms of elevated LLSR
Either abnormality modelled by the two models can lead to elevated LLSR. This is important
because intuitively it might appear that only the HI-GAIN model would increase the gain of the
LLSR. Our results suggest that changes in threshold due to tonic drive can also achieve similar
effects. It is possible that there is more than one mechanism of dystonia, and this points out that
there may be multiple causes of similar phenomenology. For example, disorders of intrinsic
excitability of motor areas (genetic or chemical) can produce dystonia, but so could disorders of
other areas that project to primary motor areas. Therefore dystonia could result either from
disorders of primary motor areas or from disorders of upstream areas, yet the manifestations will
be difficult to distinguish on clinical grounds (although there are some differences).
Another utility of the modeling TONIC and HI-GAIN models is that they potentially
represent different etiologies. In particular, a multiplicative increase in the loop gain (HI-GAIN
model) cannot be implemented by an additive component added to the system, such as a leak in
the trans-membrane current. In consequence, if the experimental data favors one model over the
other, it suggests that the clinical diagnosis of dystonia should reflect the characteristics of the
favored model, e.g. progressively increased phase angles or delayed relaxation.
3.4.4 Factors that may affect the phase angle in Experiment 1
Since phase angle is determined by the nonlinearity of a system and also its intrinsic delay,
several factors in our neuromorphic emulation could directly affect the phase angle, including 1)
the relative contribution between the primary (group Ia, both velocity- and position-dependent)
and secondary (group II, mainly position-dependent) afferents of spindle, 2) the relative
contribution between short- and long-latency pathway, and 3) the overall loop delay. Within the
range from -180° and 0°, higher activity in the primary spindle fiber makes the system more
velocity-dependent, resulting in lower phase angles; while higher activity in the secondary
spindle fiber increases the phase angle. When scanning the intensity of dystonia we keep the
activity similar between primary and secondary spindle afferents. When constructing the
boundary condition of emulation, the secondary spindle fibers are removed such that the system
60
is the least position-dependent, which theoretically provides the smallest phase angle, and vise
versa for the largest phase angle. The effect of loop delay on the phase angle is much weaker
compared to fiber type, therefore loop delay is not considered when building the boundary
conditions.
3.4.5 EMG response during muscle shortening
Two patients (Subject 4 and 5, van Doornik et al., 2009) are excluded from the model since their
positive phase angles suggest their muscles were activated by muscle shortening. This
“shortening reaction” is mostly seen in Parkinsonism, but it is uncommon in normal afferented
muscles since most of the proprioceptive spinal pathways are excitatory. Inhibitory pathways
(e.g. Renshaw cells or reciprocal inhibition connections) are capable of reducing the EMG but
usually not producing shortening reactions. One possible explanation for these two subjects is
that there exists an overflow from the opposing muscle. Our models do not include inhibitory
pathways or motor overflow. More experimental and modeling work is required to explain the
positive joint angles attested in these two patients.
3.4.6 Advantage and limitation of neuromorphic hardware emulation
The use of hardware emulation provides a powerful tool for understanding the minimal model
complexity for producing normal and pathological behaviors. It is clear that the hardware
emulation could not accommodate all the physiological details involved in real behavior. For
example, we emulated the closed-loop reflex of a joint with spinal pathways and simplified
transcortical pathways. The cortical neuron pool only includes a single layer of 128 neurons that
are far from the realistic anatomy in the primary sensory and motor cortices. Nevertheless, this
simplified reconstruction of the cortex allows us to adjust only the long-latency component of the
proprioceptive feedback independent from the functioning short-latency loop. Compared to
studying the real biological system, hardware emulation is advantageous since it allows a precise
and physiologically tenable intervention on the non-impaired system.
It is also an important issue whether our simplification of the sensorimotor system would lead
to a locally correct conclusion that, however, cannot be generalized. It has been found that reflex
of shoulder joint is concurrently modulated by the elbow joint (Pruszynski et al., 2011). This
suggests that a standalone reflex loop of a single joint, as modeled in this study, may overlook
61
the contribution from adjacent loops that would increase the LLSR and eventually worsen
dystonia. On the contrary, our model did not include inhibitory pathways such as reciprocal
inhibition, Renshaw cells, etc., which suggests that a more complex model may decrease LLSR
and thus alleviate dystonia. In short, the model is sufficient to create some features of dystonia,
but there could certainly be other contributors. It is unlikely that the inclusion of more joints
would prevent the occurrence of dystonia, but it is likely to change its nature. It would be
important to model dystonia in a multi-joint scenario.
3.5 Conclusion
In summary, we have shown through neuromorphic emulation that there are at least two possible
mechanisms of cortical abnormality that could cause increased long-latency stretch reflexes and
result in the clinical phenomenon of dystonia. While an emulation of this type does not prove
that this is the mechanism of dystonia, it does provide evidence that a mechanism of this type is
sufficient to cause the clinical features of dystonia, and is therefore worthy of further study. We
hope that these results provide both insight and guidance for future clinical studies to test
whether reducing the long-latency stretch reflex may alleviate symptoms of dystonia.
3.6 Author Contributions
Designed the experiment: WS CN. Performed the experiments: WS. Developed the hardware
and software environment: WS CN. Analyzed the data: WJ CN. Wrote the paper: WS CN.
Reviewed manuscript: WS CN TS.
62
4 Study 2. What causes motor overflow in focal
hand dystonia?
63
Motor overflow in focal hand dystonia develops and perpetuates
under correlated sensory inputs in neuromorphic emulation
Won J. Sohn
1†
, Chuanxin M. Niu
4†
, Terence D. Sanger
1,2,3*
† The authors equally contributed to the study
1
Department of Biomedical Engineering,
2
Biokinesiology, and
3
Neurology,
University of Southern California, 1042 Downey Way, Los Angeles, California,
90089
4
Department of Rehabilitation, Ruijin Hospital, School of Medicine, Shanghai
Jiao Tong University, Shanghai, China
* E-mail: tsanger@usc.edu
Abstract. Objective. Motor overflow is a common and obstinate symptom of
dystonia, manifested as unintentional muscle contraction that occurs during an
intended voluntary movement. Although it is suspected that motor overflow is due
to sensory cortical disorganization in some types of dystonia (e.g. focal hand
dystonia), it remains elusive which mechanisms could initiate and, more
importantly, perpetuate motor overflow. We hypothesize that distinct motor units
have low risk of motor overflow if their inputs remain statistically independent.
But when provided with confused sensory inputs, pre-existing crosstalk among
motor units will grow under spike-timing-dependent-plasticity (STDP) and
eventually produce irreversible motor overflow. Approach. We emulated a
simplified neuromuscular system comprising 2 anatomically distinct muscles
innervated by 2 layers of spiking neurons with STDP. The synaptic connections
between layers included crosstalk connections. The motor units received either
independent or correlated inputs during 4 days of continuous excitation. The
emulation is critically enabled and accelerated by our neuromorphic hardware
created in previous work. Main results. When driven by correlated inputs, the
crosstalk synapses gained weight and produced prominent motor overflow; the
growth of crosstalk synapses resulted in enlarged sensory representation reflecting
cortical reorganization. The overflow failed to recede when the motor inputs
resumed their original uncorrelated statistics. In the control group, no motor
overflow was observed. Significance. Although our model is a highly simplified
64
and limited representation of the human neuromuscular system, it shows that
correlated input to anatomically distinct muscles is by itself sufficient to cause
motor overflow. Among many possible abnormal patterns of neural activity in
patients with dystonia, our results suggest that the statistical independence of
motor drives may be crucial for the specificity of voluntary movements.
Keywords: Overflow, Dystonia, Motor Control, Spike-timing-dependent
plasticity (STDP), Electromyography (EMG)
65
4.1 Introduction
Motor overflow reflects a deficit in the specificity of motor commands. It describes a
phenomenon with a leak in motor signals from one part of the body to other parts, producing
extraneous movements, which accompany, but are incidental to, the voluntary action (Soska et
al., 2012). Motor overflow is commonly observed in dystonia (Gordon et al., 2006; Albanese et
al., 2013), including cervical dystonia (REF), childhood dystonia due to brain injury (Young et
al., 2011), and commonly in focal hand dystonia (FHD) (Sitburana and Jankovic, 2008). Being
one of the characteristic features of FHD, motor overflow activity has been thought to have
bearing on the loss of inhibition at multiple levels in the central nervous system, abnormal
plasticity and abnormal sensory function (Hallett, 2006), but the exact pathogenesis and
pathophysiology are far from being well understood.
Although the neural basis of motor overflow in FHD is not well understood, there is evidence
that sensory deficits, especially a decreased precision of tactile and proprioceptive perception,
could be an important cause (Bara-Jimenez et al., 2000a; Bara-Jimenez et al., 2000b; Sanger and
Merzenich, 2000). It was originally thought that once the sensory cortical map is established no
further change occurs. However, several studies in the past decades have shown that the
organization of somatotopic sensory cortex is plastic such that the map can be reorganized in
response to sensory experience, and dramatic shifts can happen due to stroke or injury (Kaas and
Florence, 1997; Nelles et al., 1999; Chen et al., 2002; Nudo, 2013). In FHD, the somatotopic
map of hand region is believed to be altered through a process of neuroplasticity. Multiple
studies suggest that there are occurrences of severe degradation in finger representation in
somatosensory cortex (Elbert et al., 1998; Bara-Jimenez et al., 1998). In primate models of focal
hand dystonia, de-differentiation in somatosensory cortex is created by excessive repetitive
associative stimulation of a skin region on different digits not normally stimulated conjointly
(Byl et al., 1996). There were also a “smearing” of receptive fields in FHD (Byl et al., 1996). In
patients with Writer’s cramp and Musician’s cramp, enlarged and overlapping receptive fields in
the somatosensory and primary motor cortex were observed. Supporting the sensory dysfunction
in the pathology of dystonia, a decreased performance in spatial and temporal discrimination
tests was consistently reported in patients with focal hand dystonia (Tinazzi et al., 1999; Bara-
Jimenez et al., 2000a; Bara-Jimenez et al., 2000b; Sanger et al., 2001; Scontrini et al., 2009)
66
Despite the physiological evidence of sensory deficits, developmental mechanisms of motor
overflow in FHD have yet to be clarified. The prolonged, repetitive use of the hand may be a
significant contributing factor in the development of the sensory abnormality in FHD. Studies
with primates show that temporally correlated sensory input activity plays a key role in the
establishment of modification in receptive fields and sensory topography through a hebbian-like
plasticity (Bara-Jimenez et al., 1998; Blake and Merzenich, 2002). In humans, associative
pairing of tactile stimuli to the digits induces organizational changes in the sensory cortex
(Godde et al., 1996). In patients with writer’s cramp, stimulation-induced reorganization in the
corticospinal motor system is more rapid and pronounced compared to that in healthy controls
and the brain's response to paired associative stimulation (PAS) was exaggerated and its spatial
specificity was reduced in writer's cramp (Quartarone et al., 2003). Considering the fact that
somatosensory dysfunction has a direct link to motor disorder (Konczak and Abbruzzese, 2013)
the loss of spatial specificity in sensory regions could lead to the loss of specificity in motor
regions. Furthermore, it is known that altered sensorimotor connections to motor cortex may
impair motor function because the somatosensory cortex has an important influence on the motor
system (Hoon et al., 2009; Sanger and Merzenich, 2000).
This evidence supports the temporal correlation hypothesis of cortical dynamic processing—
the temporal coincidence of neural events from various sensory modalities induces plastic
changes in cortical topography. Thus, we postulate that if there are correlated movements among
fingers that generate sensory input statistics that are dependent on each other, it is likely to cause
confusion in the sensory topography and eventually produce irreversible motor overflow.
However, understanding the underlying neural mechanisms linking the characteristic sensory
input statistics and the manifestation of motor overflow symptoms has been limited by
methodological availability of in-vivo neurophysiology. Therefore, we attempt to overcome the
limitation by leveraging a neuromorphic high-speed hardware emulation platform.
The rate-based computational model of the development of focal had dystonia was previously
developed (Sanger and Merzenich, 2000). As predicted by Sanger and Merzenich(Sanger and
Merzenich, 2000), and demonstrated by Merzenich and Byl (Byl et al., 1996), correlated sensory
inputs lead to sensory disorganization and have the potential to produce motor dysfunction. The
innovation of this study is the use of spike-based simulation and high-speed emulation with a
more physiologically accurate model of synaptic plasticity. In the motor overflow in FHD, we
67
speculate that temporally correlated sensory inputs between two digits could lead to the growth
in synaptic connections that are not normally strengthened such that they cause the loss of spatial
specificity and motor overflow between the two digits. In this study, we test a hypothesis that
correlated input activity in two adjacent sensory regions is sufficient to develop and perpetuate
motor overflow if spike-timing-dependent-plasticity (STDP) is simulated. The perpetuation of
motor overflow in FHD is an emergent phenomenon when the crosstalk synapses gain weight
due to the correlated inputs and produce prominent motor overflow; the overflow failed to recede
even when the motor inputs subsequently became uncorrelated. We used programmable Very-
Large-Scale-Integrated (VLSI) hardware to simulate four days of intensified development of
synaptic strength at high speed. Our model demonstrates (1) the enlargement of the cortical
finger representation and increase in receptive field of the digits by growth of crosstalk synapses
under correlated sensory inputs, and (2) the perpetuation of motor overflow after the
synchronized input activity has ended. Although our model is a highly simplified and limited
representation of sensorimotor cortex, it allows us to explain how synchronized sensory inputs
could lead to the development and perpetuation of overflow of motor commands due to
plasticity.
4.2 Materials and methods
We emulated a two-layer neuronal network using our recently developed neuromorphic
hardware. The emulation included 4 spiking neurons using Izhikevich’s approximation to the
classic Hodgkin-Huxley neurons (Izhikevich, 2003a). Izhikevich neurons are used because they
permit use of biologically realistic variables including transmembrane currents, yet they can be
implemented much more efficiently in hardware than the more complex Hodgkin-Huxley
equations that they approximate (Izhikevich, 2003a). Our emulation using Very-Large-Scale-
Integrated-circuit (VLSI) technology allows us to describe the change of synaptic weight under
the influence of each individual spike, especially their relative timing compared to adjacent
spikes. This is critical for evaluating whether the correlation between sensory inputs will have an
effect on the output. We were capable of accelerating the emulation to 190x real-time. Therefore
we are able to test hypotheses about the spatial specificity in adjacent sensorimotor system due to
correlated or uncorrelated sensory input over long-term period. In this study, we did not make
use of the parallel implementation of neuron populations because the major aim was to
68
investigate the phenomena arising from the fundamental properties of simplified neuron structure
described in the next section.
Neural structure
We modeled a subset of sensorimotor system as a layer of adjacent sensory neurons projecting to
the layer of neurons in the sensory cortex (Fig.15B). This is a simplified representation of the
sensory-motor connections in cerebral cortex. The two input neurons (n0, n2) in the input layer,
representing sensory neurons, project to two output neurons (n1, n3) in the output layer,
representing sensory cortex. Among the four synaptic connections present in the structure, the
strength of the connection is represented by the size of the weight in the connecting synapse. In
the healthy state, the horizontal synaptic connection has a dominant connection and the crosstalk
(diagonal) connection bears minimum weight to ensure high spatial specificity. In the network,
all signals are encoded by neurons as spikes, which travel to the next level of neurons via
synapses. The timing of the presynaptic and postsynaptic spikes determines the strengthening
and weakening of the connection according to the spike timing dependent plasticity (STDP) rule
(Bi and Poo, 2001; Froemke and Dan, 2002) (Fig.14A). The standard additive update with all-to-
all scheme (Pfister and Gerstner, 2006) where all pairwise combinations of presynaptic and
postsynaptic spikes contribute the update (Fig.14B). In this STDP implementation, maximum
time difference between spike pairs is 64ms due to hardware restrictions. The number of points
in the curve is adjusted to preserve the ratio of area under LTP and LTD. The STDP
implementation does not include naturally occurring synaptic decay which is not a necessary
requirement to demonstrate the result in this study but to additionally demonstrate that
homosynaptic LTD is not an inevitable property of synapses under certain condition, as
exemplified in the homeostatic synaptic level in the STPD demo (Fig.16C). In case of
presynaptic spikes causing postsynaptic spikes to fire, causally correlated pre-post firing activity
facilitates long-term potentiation (LTP) in the synapse under certain presynaptic frequency range
(<20Hz) in the simulation. Stochastic current input to the neuron is used as an activity generator.
Input current to the neuron ([0.6*I_th < I_input < 1.7*I_th ], I_th: neuron firing threshold) has
random variation with fixed mean value that causes presynaptic firing rate to be about 15Hz.
This represents constant sensory stimulation over time, and is the combined current due to all
sensory inputs to each neuron.
69
The equations that govern the update of the postsynaptic current and the synaptic weight are as
follows:
𝜏
!"#
𝑑𝐼(𝑡)
𝑑𝑡
= 𝐼 𝑡 +𝛿 𝑡− 𝑡
!
𝑔(𝑡)
𝑔 𝑡+1 =𝑔 𝑡 +𝑓 𝑡
!"#$
− 𝑡
!"#
Equation 1
First equation is a rule for updating the postsynaptic current. The postsynaptic current has an
exponential decay (𝜏
!"#
≅ 15𝑚𝑠) and incoming presynaptic spike adds to a current the size of
the strength of the synapse. Synaptic strength (𝑔 𝑡 ) is updated according to the STDP kernel (𝑓)
additively (Fig. 14). The area under the STDP curve is larger in LTD curve than LTP curve by
approximately 1.4 fold. In our implementation, discretized curve has a time resolution of 1ms.
Fig. 14B is an illustration of the all-to-all scheme in STDP. All pairwise combination of
presynaptic and postsynaptic spikes contributes to plasticity. In this implementation, a maximum
time difference between of spike pairs is 64ms due to hardware restriction. The number of points
in the curve is adjusted to preserve the ratio of area under LTP and LTD.
70
Figure 14. STDP demo. A) Spike timing dependent plasticity curve implemented on FPGA.
Synapse potentiates when postsynaptic spike arrives a few milliseconds after presynaptic spike
arrives and depresses if the order is reversed. The parameters A+= 103%, A- = -51%, tau+ =
0.014 sec, tau- = 0.034 sec are taken from Froemke and Dan (2002). The area under the curve
is larger in LTD curve than LTP curve by approximately 1.4 fold. In our implementation,
discretized curve has time resolution of 1ms. B) Illustration of all-to-all algorithm, a common
method in STDP. All pairwise combination of presynaptic and postsynaptic spikes contributes
to plasticity. In this implementation, maximum time difference between of spike pairs is 64ms
due to hardware restriction. The number of points in the curve is adjusted to preserve the ratio
of area under LTP and LTD.
4.2.1 Experimental procedure
The experiment is designed to test whether transient correlated sensory inputs may lead to the
development of motor overflow. Experiment is divided into three phases according to the type
of current profile to the input neurons in order to single out the effect of correlated sensory input
as an inserted intervention in the middle of the experiment. In the control, there is no inserted
intervention phase. In the first phase, two input neurons receive two different and statistically
100
0
0
time (ms)
t
post
< t
pre
A) STDP implemented on FPGA B) All-to-all learning rule
C) Test of STDP in neuromorphic emulation
t
post
> t
pre
-100 100
-100
71
independent currents, which represents two adjacent sensory neurons receiving distinct pattern of
inputs from each other. Two different stochastic currents guarantee that the two are uncorrelated.
In the second phase, current from an identical source is drawn to both input neurons in
synchronous way such that inputs are fully correlated. In the third phase the input profile is the
same as the first phase in order to observe how the hysteresis effect by second phase is carried on
further. The hardware acceleration allows simulation in 30 minutes in real time to represent 95
hours worth continuous sensory stimulation.
Baseline: Indepdent sensory inputs
4 days
Evaluation: Independent sensory inputs Intervention: Correlated sensory inputs
Baseline, Intervention, Evaluation: All independent sensory inputs
B) Control
A) Intervention
input
neurons
output
neurons
w1
w2
w0
w3
0
3
1
2
input
neurons
output
neurons
w1
w2
w0
w3
0
3
1
2
input
neurons
output
neurons
w1
w2
w0
w3
0
3
1
2
input
neurons
output
neurons
w1
w2
w0
w3
0
3
1
2
72
Figure 15. The experimental design testing whether confused (correlated) sensory
inputs may lead to motor overflow. Experiment is divided into three phases according
to the type of current profile to the input neurons. In the first phase, two input neurons
receive different currents, which represents two adjacent sensory neurons receiving
distinct pattern of inputs from each other. Two different stochastic current generators
guarantee that the two are uncorrelated. In the second phase, current from a single
source is drawn to both input neurons in synchronous way such that inputs are fully
correlated. This goal of this phase is to observe the effect of two adjacent sensory
neurons receiving correlated pattern of inputs. In the third phase the input profile is the
same as the first phase in order to observe the perpetuated effect by second phase in
long term. The hardware acceleration allows simulation in 30 minutes (real time) to
represent 96 hours worth continuous sensory stimulation.
4.3 Results
4.3.1 Demo of plasticity effect under STDP
We first demonstrate the functionality of the implemented STDP by tracking the long-term
effects of stochastic current input to the synaptic weight in homosynaptic connection (Fig.16A).
When stochastic current drives the input neuron that generates ~15Hz of presynaptic firing rate,
the postsynaptic spikes are generated from the postsynaptic current. The STDP rule dictates that
causal correlation between pre- and postsynaptic spikes will initiate long-term potentiation (LTP)
of the synapse. As weight increases, presynaptic spike causes a burst of postsynaptic spikes
which gets denser and longer due to the increasing postsynaptic current and therefore works as a
positive feedback to further weight increase. The time course of synaptic weight change
demonstrated shows the synaptic growth by LTP dominant stage (between day1 and day 16) and
homeostatic stable stage by the balance between LTP and LTD (after day 16). Increased
postsynaptic firing rate due to synaptic growth increases LTD, which prevents the synapse to
grow infinitely by forming a plateau of synaptic strength. The STDP implementation does not
include naturally occurring synaptic decay. It is to demonstrate that homosynaptic LTD is not an
inevitable property of synapses with STDP, and it exemplifies that decay may not be necessary
to implement the homeostatis in the synaptic strength under homosynaptic stimulation.
73
Figure 16. Validation of STDP implemented using neuromorphic hardware. We emulated a
homosynaptic connection between 2 neurons (A), of which the pre-synaptic neuron received a
continuous pseudo-white-noise input for an equivalence of 21 days. The neuromorphic
hardware (B) accelerated the emulation by 190x real-time so the emulation took about 2 hours
and 30 minutes, neuromorphic hardware also allows for onboard measurement of neuron
spikes. The longitudinal change of synaptic weight across 21 days (C) shows a gradual
increase between day 0 and day 14 followed by a plateau. When the pre-synaptic neuron
produced spike trains at a constant rate, the increase and plateauing of post-synaptic spiking
rate confirmed the longitudinal change of synaptic weight. Snapshots from oscilloscope are
shown (D). Increased bursting after a single spike is observed as the synaptic weight
increases. This is because the postsynaptic current has a weight dependent update rule
(Equation 1).
74
4.3.2 Development and perpetuation of motor overflow
The initial condition of the network is set to have weak crosstalk connections and strong healthy
horizontal connections (Fig.17A). This condition represents a healthy state. In the first phase,
non-correlated input to the two input neurons is unable to grow the crosstalk (diagonal) synapses.
This is because presynaptic spikes are weakly correlated with crosstalk postsynaptic spikes such
that LTD dominates over LTP in the STDP curve. The first phase lasts until 18th hour. In the
second phase, synchronized currents drive the two input neurons and the crosstalk gradually
grows so long as the inputs were correlated. This is because presynaptic spikes are fully
correlated with postsynaptic spikes such that LTP dominates in the STDP curve for both
crosstalk and direct connections. The second phase lasted up to the 46th hour. The slight
variation in the duration of this phase comes from the variation in time to grow the synapse due
to the stochastic input current profile. The duration of the second phase is arbitrary but is set to
be long enough to observe the growth of crosstalk synapses. In the third phase, non-correlated
input currents were applied, but the network perpetuated the state of significant crosstalk. This is
because presynaptic spikes from the input neurons are partially, but not negligibly, correlated
with the postsynaptic spikes in the output neurons located diagonally due to the significantly
grown crosstalk synapse such that contribution from LTP and LTD balances out in the crosstalk
synapses. The third phase lasted up to the 4th day (95 hours). The shaded area in the plot
represents the mean ± standard deviation for 6 trials.
75
Figure 17. Development and perpetuation of the motor overflow. Change of synaptic
weight of the crosstalk is plotted. A) Healthy state: the size of the crosstalk-synapses
(w1, w2) are kept minimal and they do not grow under non-correlated input profile to the
input neurons. B) Crosstalk synapses have grown after the correlated phase and the grown
states are perpetuated even after the correlated phase has ended.
4.3.3 Enlarged sensory representation and increase in receptive field
We quantified the growth of crosstalk, enlarged cortical representation, increase in receptive
field size in the following way. At time point A and B in Fig. 17, we sampled the input and
output signals to analyze the change made before and after the intervention in the second phase.
To quantify the growth of crosstalk and enlarged cortical representation, electromyogram (EMG)
is generated from the postsynaptic signals as an output measure to represent the synaptic strength.
Increased EMG response due to the growth of crosstalk represents the sensory deficit being
w1
w2
synaptic weight
(crosstalk)
Cor Non
-cor
Non-cor
time (days)
before after
t
t
w1
w2
w1
w2
input
neurons
output
neurons
input
neurons
output
neurons
A B
AB
5000
20,000
control
variable
1
2
34
76
propagated to the motor abnormality. The size of the cortical representation is measured from the
sum of EMG responses in the two output neurons due to the activity in one input neuron (Fig.
18). For the change in receptive field size, we count the number of input nodes that produce
EMG activity in output neuron D (Fig. 18) over a specified threshold.
77
Figure 18. Visual aid for the enlarged cortical representation, increase in
receptive field and growth in crosstalk. We also see a decrease in spatial
specificity due to a growth of crosstalk.
Growth of crosstalk in response to input in A is represented by the difference
between D and D’ (Fig.18A)
Enlarged cortical representation for part A is represented by the difference
between C and C’ + D’ (Fig.18B).
Increase in receptive field size for neuron D is represented by the A and A + B
(Fig.18C)
Fig.19 shows a comparison between human overflow data and our emulation data. The
human data (left) are from the study of motor overflow in patients with dystonic symptoms at
hand (Young et al., 2011). Human data show representative EMG recordings of two muscles
during 60s trials. The participants were instructed to track the target on the screen that is
controlled by EMG modulated signal from a task finger. EMG from a non-task finger is also
recorded to measure the overflow. In the 2x2 matrices, if the signals are only in the diagonal
window it means there is no overflow. If there is an activity in the non-diagonal window it
signifies the existence of overflow of muscle activity from the muscle in the task finger to the
muscle in the non-task finger. The comparison is made between control and dystonia. Similar to
the human data, our emulation data (60s) sampled before and after the correlated input phase are
plotted in the 2x2 matrices (Fig.19, right). Activity in the non-diagonal window signifies the
overflow of activity due to crosstalk. Growth of crosstalk is measured from the change in mean
EMG level in output node D in response to an input in A (Fig.18A). Normalized crosstalk is
0.08 (=0.0054/0.0720) in human control and 1.78 in human dystonia (+1.70, increased crosstalk),
and 0.20 in emulated control and 0.31 (+0.11, increased crosstalk) in emulated dystonia.
Enlarged cortical representation is measured from the change in sum of mean EMG levels in the
two output nodes C and D (Fig.18B). Normalized cortical representation is 1.08
(=(0.0054+0.0720)/0.0720) in human control and 2.78 in human dystonia (+1.70, enlarged
78
cortical representation), and 1.20 in emulation control and 1.31 in emulation dystonia (+0.11,
enlarged cortical representation). Increase in receptive field size is quantified by recording the
number of input neurons that elicit a response in output node D (Fig.18C). Discrete number of
receptive field is 1 in human control and 2 in human dystonia (+1, increased receptive field), and
1 in emulation control and 2 in emulation dystonia (+1, increased receptive field).
Figure 19. Comparison of motor overflow between human and emulated data.
Signals only at diagonal windows represent healthy human control. In dystonia,
signals in the non-diagonal windows increases due to motor overflow. Similar
pattern is observed between emulation of control and dystonia (right).
4.4 Discussion
The purpose of this study was to use neurmorphic emulator to understand the origin and
development of focal hand dystonia (FHD) because developmental mechanisms of motor
overflow in FHD have yet to be clarified. We started from physiological evidences of sensory
deficits found in FHD and speculated that the correlated sensory activity could a direct cause for
79
the development of motor overflow in FHD. In other words, this is to emulate prolonged,
repetitive use of the hand as a significant contributing factor in the development of the sensory
abnormality in FHD. Considering the fact that somatosensory dysfunction has a direct link to
motor disorder (Konczak and Abbruzzese, 2013), and the somatosensory cortex has an important
influence on the motor system (Hoon et al., 2009; Sanger and Merzenich, 2000) we assumed
that emulating the development of sensory disorganization would directly imply the
disorganization in motor system, leading to motor overflow. In order to emulate the dynamic
growth of strength in the neural structure, we implemented spike-timing-dependent plasticity
(STDP) in customizable hardware to build a biologically realistic high-speed emulation platform
to test the temporal correlation hypothesis of cortical dynamic processing in FHD—namely we
postulate that correlated movements two adjacent fingers is sufficient to cause confusion in the
sensory topography and eventually produce irreversible motor overflow between the two fingers.
The hypothesis as an extension from the rate-based computational model of the development of
focal had dystonia that was previously developed (Sanger and Merzenich, 2000). The innovation
of this study is the use of spike-based simulation and high-speed emulation with a more
physiologically accurate model of synaptic plasticity. Our model demonstrates (1) growth of
crosstalk synapses and (2) the enlargement of the cortical finger representation and (3) increase
in receptive field of the digits under correlated sensory inputs. Another emerging phenomenon of
using STDP was the perpetuation of motor overflow after the synchronized input activity has
ended.
4.4.1 Relevance of the simulated neural structure to the biology
The two-layer neural structure is used to represent simplistic sensorimotor system in the cortex.
Four synaptic connections from input neuron substrate to output neuron substrate represent the
normal and crosstalk connection between two sensory digit representations. Although
admittedly the neural structure is an oversimplification of the complex system encompassing
from sensory neurons to the sensory representational map in the cortex, our model allows us to
explain how synchronized sensory inputs could lead to the development and perpetuation of
overflow of motor commands due to plasticity.
80
4.4.2 Sensitivity of the result to other synaptic learning models
The question could be asked of how sensitive the result will be to the various synapse learning
models. Although it is not of our interest in this study to test the different effects of various
implementations of STDP update mechanisms, it can be said that the result is generalizable if 1)
update mechanisms dictate that correlated pre- and postsynaptic spikes causes potentiation and
uncorrelated pre- and postsynaptic spikes causes depression of the synaptic weight, 2)
postsynaptic firing rates are not in the range that results in net depression. Our implementation
has asymmetric STDP curve, and is all-to-all (Pfister and Gerstner, 2006) and standard additive.
Variation in curve shape, nearest-neighbor implementation instead of all-to-all, multiplicative in
stead of additive update will still generate the same results in this study if the above conditions
are met.
4.4.3 Implication of the emulated result in clinical treatment of FHD
The emulation of the FHD suggests how to minimize the risk of development or possibly prevent
the development of FHD. Since the synchronous sensory activity is the direct cause that makes
the crosstalk synapse to grow, the first caution should be taken not to get heavily involved in
effortful exercises that generate correlated sensory movements—forced grips or any
simultaneous contraction of hand muscles, e.g. in labored writing. It is equally important that the
potential patients must limit their duration of exercising any synchronous activity, if they must
do some. For musician who are at a risk of developing musician’s cramp, limiting the play time
will help to prevent the crosstalk to grow beyond a point that will make the effective treatment
harder due to the perpetuating nature of the overflow. Among the list of currently available
treatment options in FHD, Botulinum toxin is proven to have a benefit but many patients
discontinue due to dissatisfactions with the result. Sanger and Merzenich’s model (Sanger and
Merzenich, 2000) implies that some of the beneficial effect of injecting the toxin may be due to
reducing the overall gain of the sensorimotor loop by reducing the sensory afferents. Present
study implies that once the FHD is fully developed there may not be an easy way to return to the
healthy initial state by eliminating the crosstalk connections. One of the possible rehabilitation
methods might involve completely inactivating the affected hand for long enough time to hope
for the cortical representational map of the hand to shrink in size (or un-mapping process) such
that the remapping can be done easily.
81
The fact that we observe a rapid change in cortical organization upon stimulation also suggests
higher-than-normal plasticity in patients with focal hand dystonia. It would be interesting future
work to emulate the effect of learning rate in developing focal hand dystonia.
4.5 Conclusion
The purpose of this study was to understand the pathology of developing focal hand dystonia by
using biologically realistic neural structure to test our temporal correlation hypothesis. As
predicted by Sanger and Merzenich (Sanger and Merzenich, 2000), and demonstrated by
Merzenich and Byl (Byl et al., 1996), correlated sensory inputs lead to sensory disorganization
and have the potential to produce motor dysfunction. Here we extended this work using a more
realistic plasticity model with spike-timing dependent plasticity.
4.6 Author Contributions
Designed the experiment: WS CN. Performed the experiments: WS. Developed the hardware
and visual software: WS. Analyzed the data: WS. Wrote the paper: WS. Presented by WS at
the Annual Meeting of the Neural Control of movement (NCM) in April 2014 and April 2015.
Reviewed the manuscript: WS CN TS.
82
5 Study 3: How does the constraint-induced
therapy work?
83
Emergent phenomenon from a synaptic competition:
Constraint-induced intervention as a way to escape from the
suboptimal stable solution in biological systems
Won J. Sohn
1†
, Terence D. Sanger
1,2,3*
1
Department of Biomedical Engineering,
2
Biokinesiology, and
3
Neurology,
University of Southern California, 1042 Downey Way, Los Angeles, California,
90089
* E-mail: tsanger@usc.edu
Abstract. The principle of constraint-induced therapy is widely practiced in
rehabilitation. In the animal model of hemiplegic cerebral palsy (CP), the
impaired contralateral corticospinal projection due to postnatal unilateral injury
was repaired after imposing temporary constraint in one of the less affected
hemisphere. Despite some differences, such impairment in motor control by early
brain injury bears a resemblance to amblyopia in that it involves inter-hemispheric
activity-dependent synaptic plasticity. Previously, the mechanism for amblyopia,
equivalent to hemiplegic CP in visual system, has been explained within the
framework of BCM theory, a rate-based synaptic modification theory, but here we
attempt to provide a fundamental explanation for the general biological
phenomena that involve inter-hemispheric synaptic competition in spike-based
theory. In this study, we choose to emulate the restoration of the postnatal
hemiplegic CP in terms of the competition between ipsilateral and contralateral
CST. The importance of the study comes from the fact that there is a considerable
gap between clinical practice and the understanding of the neural mechanism
underlying the therapeutic method due to the limited methodological availability
in electrophysiology, despite the success in the many applications of constraint-
induced therapy. We strive to overcome the limitation by leveraging our
neuromorphic high-speed hardware emulation platform created in previous work
to study the neural underpinning of constraint-induced therapy. We hypothesize
that the mechanism of constraint-induced therapy can be demonstrated when a
simplified neural descending tract with 2 layers of spiking neurons that represent
cortical and spinal neural substrate are simulated with spike-timing-dependent
84
plasticity (STDP). In the 19-days of continuous emulation that includes periods of
the development of hemiplegia and the recovery from it due to imposing an
alternate constraint on the uninjured hemisphere, we observed the activity-
dependent synaptic competition as a key mechanism that accounts for the
formation of persistent deficits which is suboptimal due to transient
developmental injury and learned the fact that a forced intervention is essential to
transition into an improved state. Although our model is a highly simplified and
limited representation of descending corticospinal system, it offers an explanation
of how constraint as an intervention can help the system to escape from the
suboptimal solution. This is a display of an emergent phenomenon from the
synaptic competition.
Keywords: Spike-timing-dependent plasticity (STDP), Constraint-induced
therapy, Constraint-induced movement therapy (CIMT), corticospinal
development, suboptimal system.
85
5.1 Introduction
Recent discoveries regarding how the central nervous system responds to injuries have prompted
development of rehabilitative training for the patient to reacquire lost function. For instance, new
families of techniques called constraint-induced movement therapy (CIMT) have been developed
and proven to be effective in producing large improvement in limb use of patients after
cerebrovascular accidents (CVA) (Taub et al., 1999), and sensory and motor CNS injuries
(stroke, etc.). It is constraining the use of less-impaired upper extremities while intensive and
repetitive training of motor activities are conducted for up to 6 hours per day, for 2-4 weeks with
rewarding to the successive approximation to the target task (Taub et al., 1999; Taub et al., 2007;
Brady and Garcia, 2009), and it became a ‘paradigm shift’ in rehabilitation of CNS injury.
Conceptually similar interventional strategy that promotes activity of the injured system or
demotes the activity of uninjured system in the animal model of hemiplegic cerebral palsy (CP)
has been shown to contribute to significant corticospinal tract (CST) repair and motor recovery.
Martin et al. showed that temporarily imposed constraint in one of the less affected hemisphere
and active stimulation in the affected hemisphere harness activity-dependent plasticity to repair
the diminished connectional strength (Martin, 2005; Friel and Martin, 2007; Martin et al., 2011).
When activity in one motor cortex is blocked pharmacologically during an early sensitive period,
CST axons withdraw their projections (Friel and Martin, 2007) and constraining the use of one
limb during a similar period produced similar effect on the development of contralateral CST
projections (Martin et al., 2004).
Amblyopia is a unilateral reduction of visual acuity that bears a resemblance to the
impairment in the corticospinal system by early brain injury, as it is referred as hemiplegic CP in
the visual system. Amblyopia is the most common cause of lifelong monocular blindness
(Attebo et al., 1998), caused by abnormal visual experience during postnatal development. The
commonly practiced intervention involves patching the dominant eye while forcing the child to
use the weaker eye to revive the visual impairment of the weaker eye. In animal studies, the
effect of monocular deprivation is reversible only if the treatment is applied during the critical
period because the treatment causes changes in visual cortex for recovery (Movshon, 1976;
Movshon & Van Sluyters, 1981). The classic idea is that the patching gives a competitive
advantage to the amblyopic eye to overcome the dominance established by the other eye.
86
Although the underlying mechanisms behind the effect of the patching in the visual system may
not identical to the effect of imposing a constraint in the corticospinal system, it is still evident
that the principle of constraint-induced therapy can be highly effective across the modalities in
the biological system as long as inter-hemispheric competition is present.
Previously, the treatment mechanism for amblyopia has been explained within the framework
of BCM theory, a rate-based synaptic modification theory. Here, we attempt to provide a simple
yet fundamental mechanism that accounts for the general biological phenomena that involve
inter-hemispheric synaptic-competition using spike-based plasticity theory. Specifically in this
study, we choose to emulate the development of and restoration from the abnormal corticospinal
projections in postnatal hemiplegic CP. The importance of the current study comes from the fact
that there is a considerable gap between clinical practice and the understanding of the neural
mechanism underlying the therapeutic method due to the limited methodological availability in
electrophysiology, despite the success in the many applications of constraint-induced therapy.
We strive to overcome the limitation by leveraging our neuromorphic high-speed hardware
emulation platform created in previous work to study the neural underpinning of constraint-
induced therapy. We hypothesize that the mechanism of constraint-induced therapy can be
demonstrated when simplified neural descending tracts with two layers of spiking neurons
representing cortical and spinal neural substrate are simulated with spike-timing-dependent
plasticity (STDP). With 19-days of continuous simulation of the effect of alternate constraints
on the cortical neural substrate, we observe the formation of persistent deficits which is
suboptimal due to transient problems and learn the fact that a forced intervention is essential to
transition into an improved state.
5.1.1 Postnatal brain damage and its effect in corticospinal projection
If there is damage to the corticospinal tract (CST) in the postnatal period, there could be a
substantial and long-term consequence for motor function. Unilateral brain injury of this kind
can result in hemiplegic cerebral palsy (CP) (Farmer et al., 1991; Carr et al., 1993; Eyre et al.,
2001; Eyre, 2007). If the injury causes unilateral inactivation in the developing sensorimotor
cortex, the reduced activity in the affected hemisphere could cause increased withdrawal of its
existing contralateral corticospinal projection thus results in a weakened contralateral
corticospinal projection from the affected cortex. In contrast, the uninjured cortex could develop
87
increased ipsilateral corticospinal projection, subsequently displacing the contralateral
projection, developing suboptimal bilateral projections from the less affected side to the spinal
cord, contributing to the persistence of the impairment in motor control due to the insufficiency
of the ipsilateral projection from the less affected side to replace the contralateral projection from
the affected side (Hendricks et al., 2003; Vandermeeren et al., 2003; Eyre, 2007).
In feline model of CP, Martin found that with hemiplegic CP, the impaired contralateral CST
is at disadvantage early in development, compared with the ipsilateral CST that is undamaged, in
forming a strong synaptic connection in the spinal cord (Martin et al., 2011). It was proposed
that activity-dependent synaptic plasticity is a key mechanism both in the development of
hemiplegia and restoring a normal spinal termination pattern and restoring skilled motor
function. It was shown that electrical simulation of the impaired CST and inactivation of
unimpaired side (reverse inactivation) restored the normal CST termination and the normal
motor control was recovered. The feline experiment was significant in understanding the
development and treatment for postnatal hemiplegic CP because human CST development
parallels closely that of the cat (Martin et al., 2011).
5.1.2 Spike-based synaptic learning rule
Spike-timing-dependent plasticity (STDP) is a synaptic learning rule, based on biological
evidences, that uses correlated spiking activity between pre- and postsynaptic neurons to
implement synaptic plasticity effects and it is considered a good approximation to learning and
information storage in the brain and also for the development and refinement of neuronal circuit
in the brain. Depending on the temporal correlation between the spikes, either the synapse gets
strengthened or weakened (Bi and Poo, 2001). If the presynaptic spike arrives a few milliseconds
prior to the postsynaptic spike, then the pair will contribute to the long-term potentiation (LTP)
of the synapses and vice versa for the long-term depression (LTD) of the synapses. The change
of the synaptic weight according to the relative timing of pre- and postsynaptic spikes are shown
in the STDP function or learning window (see Fig. 21). The curve is based on biological
responses from previous studies (Froemke and Dan, 2002). The update of the synaptic weight
follows additive, all-to-all scheme (Pfister and Gerstner, 2006).
A rate-based BCM synaptic modification theory (Bienenstock et al., 1982) has shown
success in providing an explanation for learning in the visual cortex, developed in 1982. It
88
proposes first order mathematical prediction of how synaptic plasticity stabilizes postsynaptic
activity by a sliding threshold. Although useful to some extent, the explanation is superficial due
to its poor biorealism and, therefore, its limitation is defined. The use of STDP over the rate-
based models as a synaptic plasticity rule is an advance in terms of biorealistic property.
5.1.3 Benefit of simulating a biological system
We used programmable Very-Large-Scale-Integrated-circuit (VLSI) system, which allows us to
create emulations of neurons that communicate using spikes, with the potential to increase the
number of emulated neurons without sacrificing speed. A spike-based model is valuable because
we want to understand the high-level process in the physiological system that arises from
synaptic plasticity based on the neural spiking activity. In this way, we could use spike-based
plasticity theory such as spike-timing-dependent plasticity (STDP) to account for the key
experimental results. The system allows us to emulate the long-term effect of an intervention, a
change in input characteristics to the neural structure, to the change in the synaptic weights in the
network. The system further allows us to monitor the emergent phenomena of the STDP in the
context of synaptic competition in the simplistic neural structure representing the descending
corticospinal system. In this study, we emulate 19-days of continuous simulation of the effect of
alternate constraints in the developing brain modeling the experiment on the animal model by
Martin et al (Friel and Martin, 2007).
We observed the formation of persistent deficits that is suboptimal due to transient
developmental injury and learned the fact that a forced intervention is essential to transition into
an improved state. Specifically, our model demonstrates (1) how unilateral inactivation could
create hemiplegic bilateral projection, which is suboptimal stable state, (2) the inability of the
system to spontaneously restore the diminished contralateral projection after the unilateral
inactivation stage has ended, and (3) how alternate inactivation enables the diminished
contralateral projection to be restored back by using a principle of constraint-induced therapy.
Although our model is a highly simplified and limited representation of descending corticospinal
system, it allows us to understand how the principle of constraint-induced therapy can be
explained in terms of synaptic competition by activity-dependent plasticity.
89
5.2 Materials and methods
We hypothesize that the mechanism of constraint-induced therapy can be demonstrated when a
simplified neural descending tract with two layers of spiking neurons that represent cortical and
spinal neurons are simulated with spike-timing-dependent plasticity (STDP). We customized this
two-layer neuronal network from our recently developed neuromorphic hardware. We focus on
using spike-based emulation to determine the functional role of constraint-induced intervention,
especially their sufficiency for causing restoration of the diminished contralateral projection. The
hardware emulation of the spiking neurons and the associated neural structure are constructed
using field programmable gate arrays (FPGA, Xilinx Spartan-6), a programmable version of
VLSI electronic chips. The emulation includes four spiking neurons using Izhikevich’s
approximation to the classic Hodgkin-Huxley neurons (Izhikevich, 2003a). Although the system
is designed to easily increase the number of emulated neurons to hundredfold without sacrificing
speed, current study limits the number of neurons to minimal such that we could focus on
understanding the fundamental effect of the intervention in the context of synaptic competition in
the simplistic setting.
90
Figure 20. Model neural structure. A) Simplified schematic for descending CST. CSTs
initiate from the motor cortex and terminate on the cervical gray matter in the spinal cord. The
bold descending lines represent contralateral projection (e.g. right hemisphere to left spinal
gray matter) and the dotted line represents ipsilateral projection. The projection strength of the
descending tract is represented by the relative thickness of the lines. In normal development,
contralateral projection dominates the ipsilateral projection. Distribution of CST projection
within the gray matter is not considered for simplification. B) In the simulated neural
structure, two layers of spiking neurons representing cortical neurons (input neurons) and
output neurons (spinal neurons) are connected via synapses (triangles). A strength of the
synapse is represented both by the relative size of the triangle in the structure and by the
opacity of the colors in the color matrix. Red corresponds to the cortical projection from the
right hemisphere and black for the left hemisphere.
5.2.1 Neural structure
We modeled the descending corticospinal system in two layers of neurons (Fig.20B). The two
input neurons (R/L) in the input layer, representing cortical neurons in the right and left motor
cortices, project to two output neurons in the output layer, representing neurons in the spinal gray
matter (Fig.20A). Among the four synaptic connections present in the structure, the strength of
the connection is represented by the size of the weight in the connecting synapses. Izhikevich
neurons are used because they permit use of biologically realistic variables including
transmembrane currents, yet they can be implemented much more efficiently in hardware than
the more complex Hodgkin-Huxley equations that they approximate (Izhikevich, 2003a). In the
network, all signals are encoded by neurons as spikes and the spikes pass through synapses. The
timing of the presynaptic and postsynaptic spikes determines the strengthening and weakening of
the connection according to the spike timing dependent plasticity (STDP) rule (Fig.21A). The
standard additive STDP model with all-to-all algorithm where all pairwise combination of
presynaptic and postsynaptic spikes contributes to plasticity is used (Fig.21C). In this
implementation, maximum time difference between of spike pairs is 64ms due to hardware
restriction. The number of points in the curve is adjusted to preserve the ratio of area under LTP
91
and LTD. In case of presynaptic spikes causing postsynaptic spikes to fire, causally correlated
pre-post firing activity facilitates long-term potentiation (LTP) in the synapse under certain
presynaptic frequency range (<20Hz) in the simulation. Stochastic current input to the input
neuron is used as an activity generator. Input current to the neuron (I_input =[0.6*I_th,
1.7*I_th ], I_th: neuron firing threshold) has a random variation centered around a mean value
that causes presynaptic firing rate to be about 15Hz. The feeding of the current to the input
neurons represents constant sensory activity over time.
The equations that govern the update of the postsynaptic current and the synaptic weight are as
follows:
𝜏
!"#
𝑑𝐼(𝑡)
𝑑𝑡
= 𝐼 𝑡 +𝛿 𝑡− 𝑡
!
𝑔(𝑡)
𝜏
!"#
𝑑𝑔(𝑡)
𝑑𝑡
=𝑔 𝑡 +𝑓 𝑡
!"#$
− 𝑡
!"#
First equation is a rule for updating the postsynaptic current. The postsynaptic current has an
exponential decay (𝜏
!"#
≅ 15𝑚𝑠) and incoming presynaptic spike adds to a current the size of
the strength of the synapse. Synaptic decay has a natural decay (𝜏
!"#
) with a time constant of
~x ms. Synaptic strength (𝑔 𝑡 ) is updated according to the STDP kernel (𝑓) additively (Fig.
21A, C). The area under the STDP curve is larger in LTD curve than LTP curve by
approximately 1.4 fold. In our implementation, discretized curve has a time resolution of 1ms.
Fig. 1C is an illustration of the all-to-all algorithm in STDP(Kempter et al., 1999; Song et al.,
2000). All pairwise combination of presynaptic and postsynaptic spikes contributes to plasticity.
In this implementation, a maximum time difference between of spike pairs is 64ms due to
hardware restriction. The number of points in the curve is adjusted to preserve the ratio of area
under LTP and LTD.
92
Figure 21. STDP model: the model includes standard all-to-all, additive STDP with synaptic
decay and stochastic current input to the neuron as an activity generator. A) Spike timing
dependent plasticity curve implemented on FPGA. Synapse potentiates when postsynaptic
spike arrives a few milliseconds after presynaptic spike arrives and depresses if the order is
reversed. The parameters A+= 103%, A- = -51%, tau+ = 0.014 sec, tau- = 0.034 sec are taken
from Froemke and Dan (2002). The area under the curve is larger in LTD curve than LTP
curve by approximately 1.4 fold. In our implementation, discretized curve has a time
resolution of 1ms. B) Illustration of the spikes traveling from presynaptic neurons to
postsynaptic neurons via synapses. The postsynaptic current generates postsynaptic spikes. C)
Illustration of the all-to-all algorithm in STDP. All pairwise combination of presynaptic and
postsynaptic spikes contributes to plasticity. In this implementation, a maximum time
difference between of spike pairs is 64ms due to hardware restriction. The number of points in
the curve is adjusted to preserve the ratio of area under LTP and LTD.
5.2.2 Experimental procedure
Experimental procedure of the emulating the constraint-induced therapy closely follows the
original experimental procedure conducted by Martin in kittens. The initial stages represents (1)
normal contralateral predominant structure (“initial state”) and followed by (2) unilateral
inactivation of one of the hemisphere by blocking the activity in one of the input neurons
(“unilateral inactivation”), (3) normal bilateral activation by removing the blockade (“bilateral
93
activation”), (4) reverse inactivation of the other hemisphere by blocking the activity in the input
neuron that wasn’t blocked before (“reverse inactivation”), and returning to (5) normal bilateral
activation which is identical to the third stage (“bilateral activation”). The initial unilateral
inactivation is to create a hemiplegic state from the normal contralateral-dominant state. In this
stage bilateral projection is established from the activated side by the growth of initially weak
ipsilateral connection. Subsequent bilateral activation is to prove that once the bilateral
projection, which is suboptimal in terms of biological performance, is established, the state will
not change spontaneously without an intervention. The reverse inactivation from the source of
bilateral projection is applying a constraint as an intervention to escape from the otherwise
consolidated (“stuck”) state. After keeping the constraint for a while, the constraint is removed to
return to the normal bilateral activation for the hope that the intervention was effective and has a
lasting effect. The synaptic decay rate of the synaptic weight of the contralateral connection is
set to be lower than that of the ipsilateral connection (see discussion for more detail). We keep
track of the change of synaptic weights from the four synapses during the dynamic changes in
the input current profile to input neurons.
94
5.2.3 Constraint-induced therapy uses the property of synaptic competition.
95
Figure 22. Simulating activity-dependent constraint-induced therapy by STDP in 4
neurons, 4 synapses neuronal structure. Schematic of five stages and corresponding
simulation windows. The five stages parallels both the development of and the recovery
from the hemiplegic CP in Martin’s experiment (Friel and Martin, 2007). The sizes of
synaptic weights (w1-w4) represent snapshots at the end of each phase. A) a. Initial
stable state has normal predominant contralateral projection and weaker ipsilateral
projection. b. Unilateral inactivation (blocking R) causes the inactive side to develop a
diminished projection (w2), and the active side to develop bilateral connections (w1, w3).
The ipsilateral projection in the active side (w3) is better able to compete with less active
contralateral projection (w2) for synaptic connections with the output neurons. c. Bilateral
activation after the removal of unilateral inactivation shows that the diminished
contralateral projection (w2) could not be spontaneously restored. In other words, the
state is stuck (fig. 3A) at bilateral projection. d. Reverse inactivation (blocking L)
enables the diminished contralateral projection (w2) to be restored by inactivating the
strong ipsilateral projection (w3), which has been preventing the recovery of w2. The
duration of this stage is an experimental design choice. To ensure the effect of constraint
therapy, this stage needs to be prolonged until w2 becomes comparable in size respect to
w3. e. Bilateral activation after the reverse inactivation resumes the competition between
contralateral and ipsilateral projection competing for the synaptic connection to the output
neurons. If the reverse inactivation stage ended at contralateral connection being stronger
than the ipsilateral connection, the competition leads to the contralateral dominant state
with a high likelihood. B) Simulated windows for the five stages that correspond A. Note
that the direction of the signals in this window is from left to right. The miniature plot in
the bottom keeps track of the real-time change of the synaptic weights. The magnified
version is in fig. 5.
96
5.3 Results
5.3.1 Illustration of activity-dependent synaptic competition
Fig. 23 shows illustrations of three contingencies for a synaptic competition according to
STPD when two input neurons project to an output neuron. The outcome of the competition is
dependent on the initial conditions of the synapses. The STDP rule dictates that the relationship
between pre- and postsynaptic spikes will induce long-term potentiation (LTP) of the synapse if
input and output are correlated, and long-term depression (LTD) if input and output are not
correlated. In all cases, the input neurons (R/L) receive stochastic input current that are
statistically independent to each other, and the synaptic weight sizes are capped (maximum
weight is four times bigger than the maximum weight). When the synaptic strength of one has
predominant initial condition, the predominant connection does not allow a chance for the
weaker connection to grow, thus the states are kept stuck as they initial were (Fig. 23A). Even if
the different in strength between two competing connection are small, the relatively stronger
connection tends win and leads to the predominant connection with high likelihood (Fig. 23B).
In order to escape from the stuck state, transient constraint of input activity in the dominant input
neuron is useful. The transient constraint induces a switch in the predominant connection (Fig.
23C). In other words, the transient blockade of the activity to the predominant neuron provided a
chance to the disadvantaged connection to overcome the dominance once established. The
effective transient constraint will last until the relative weight of the two synaptic weights
become comparable to each to other.
97
98
Figure 23. Contingencies for a synaptic competition according to STPD when two input
neurons project to an output neuron. The change in synaptic weights representing the
connectional strengths between input and output neurons are plotted with different initial
conditions. The results demonstrate how STDP affects potentiation and depression of the
synapse over time when constant current activity is in the input neurons. The STDP rule
dictates that the relationship between pre- and postsynaptic spikes will induce long-term
potentiation (LTP) of the synapse if input and output are correlated, and long-term depression
(LTD) if input and output are not correlated. In all cases, the input neurons (R/L) receive
stochastic input current that are statistically independent to each other, and the synaptic
weight sizes are capped (maximum weight is four times bigger than the maximum weight).
A) When it starts with left predominant initial condition, the predominant connection does not
allow a chance for the weaker connection to revive, thus the states are kept stuck as they
initial were. B) When left is relatively stronger than the right, the synaptic competition leads
to left predominant state with high likelihood. C) Transient constraint of input activity in the
dominant input neuron induces a switch in the predominant connection. In other words, the
transient blockade of the activity to the left predominant neuron (L) provided a chance to the
disadvantaged connection (R) to overcome the dominance established by the other input
neuron. The effective transient constraint will last until the relative weight of the two
synaptic weights become comparable to each to other.
5.3.2 Simulating activity-dependent constraint-induced therapy by STDP
Change of synaptic weights of four synapses according to the five stages of input profiles are of
key interests because the connectional strength of corticospinal tract is represented by their
relative weights. The weight matrix helps to visualize the relative strength of the four synaptic
connections (fig. 24A). Note that w0 and w1 are in competition and likewise w2 and w3 are in
competition to gain synaptic connection to the output neurons. The normal state starts with
predominant contralateral projections. Unilateral inactivation causes the inactive side to
withdraw its projection (w2), and reciprocally causes active side to develop ipsilateral
connections (w3) because the ipsilateral projection in the active side (w3) is advantageous to
99
compete with less active contralateral projection (w2) for synaptic connections with the output
neurons. As a result, bilateral projection is developed from the active side (w1, w3). Bilateral
activation after the removal of unilateral inactivation shows that the diminished contralateral
projection (w2) could not be spontaneously restored. In other words, the state is stuck at
pathological bilateral projection from the active side. Reverse inactivation enables the
diminished contralateral projection (w2) to be restored by inactivating the strong ipsilateral
projection (w3), which has been preventing the recovery of w2. The duration of this stage is an
experimental design choice. To ensure the effect of the applying constraint, this stage is
prolonged until w2 becomes comparable in size respect to w3 such that w2 is not overpowered
by w3. If this condition were met, resumed competition between contralateral and ipsilateral
projection in the following bilateral activation stage would provide setting favorable to the
strengthening of the contralateral projection and reciprocally weakening of the ipsilateral
projection.
5.3.3 What is the condition that the states will be consolidated?
Present study demonstrates the strengths of the two projections that are competing for the
synaptic space in the output neuron are consolidated after a certain period of time as a result of
competition by STDP. The question can be asked about the conditions of initial weights of the
competing synapses that will decide the winner. Due to the fact that the presynaptic signals
includes random noise, the stochastic nature of the result prohibits us to clearly mark the
threshold of weights where one side will surely win the other side. However it is safe to state that
when the weight of one side is significantly greater that the other side, the one with a greater
weight will eventually displace the other and establish a predominant connection due to the
nature of STDP. There ought to be a region of middle ground in which the development of
predominance can go either way, which is highly unpredictable due to the stochasticity of the
input stream. In this simulation, it is our design choice that the decay rate of the synaptic weight
in the contralateral projection is set to be 50% higher than the decay rate of the synaptic weight
of the ipsilateral projection to coarsely model the comparative advantage of the contralateral
projection in forming a synaptic connection to the spinal neuron. This is because an axonal path
with smaller decay rate will provide a survival advantage when competing with a path with
higher decay rate. Although in real biology the winner between contralateral and ipsilateral
100
would form a denser projection that can be quantified from the relative number of synapses or
the volume of axons projecting to the spinal neurons, we attempt to quantify this to an increase in
synaptic weight size as a high-level rendering of the phenomenon. The level of advantage set
here is arbitrary yet biologically plausible assumption and the ratio of decay rates would not
affect the mechanism of constraint-induced intervention but it may affect how long it takes to
restore the normal contralateral predominant projection.
101
102
Figure 24. Simulating activity-dependent constraint-induced therapy by STDP: a
change of synaptic weights of 4 synapses according to the five stages of input
profiles. A/B. The weight matrix helps to visualize the relative strength of the four
synaptic connections. Note that w0 and w1 are in competition and likewise w2 and
w3 are in competition. a. The normal state with predominant contralateral
projections. b. Unilateral inactivation (blocking R) causes the inactive side to
develop a diminished projection (w2), and the active side to develop bilateral
connections (w1, w3). The ipsilateral projection in the active side (w3) is better able
to compete with less active contralateral projection (w2) for synaptic connections
with the output neurons. c. Bilateral activation after the removal of unilateral
inactivation shows that the diminished contralateral projection (w2) could not be
spontaneously restored. In other words, the state is stuck (fig. 3A) at bilateral
projection. d. Reverse inactivation (blocking L) enables the diminished contralateral
projection (w2) to be restored by inactivating the strong ipsilateral projection (w3),
which has been preventing the recovery of w2. The duration of this stage is an
experimental design choice. To ensure the effect of the constraint therapy, this stage
needs to be prolonged until w2 becomes comparable in size respect to w3. If this
stage goes on longer, bilateral projection from R will develop reversely, which is why
it is important to control the duration of this stage. e. Bilateral activation after the
reverse inactivation resumes the competition between contralateral and ipsilateral
projection competing for the synaptic connection to the output neurons. If the reverse
inactivation stage ended at contralateral connection being stronger than the ipsilateral
connection, the competition leads to the contralateral predominant state with a high
likelihood (fig. 3B).
*The inherent decay rate of the synaptic weight of the contralateral connection is set
to be lower than that of the ipsilateral connection.
103
5.4 Discussion
The purpose of this study was to use biologically realistic simulation tool to understand how
biological system responds to injuries and also to find a fundamental and generalizable
mechanism of constraint-induced therapy as a rehabilitative treatment after injuries. Although
not identical in mechanism, constraint-induced movement therapy (CIMT) after central nervous
system injury due to stroke or cerebrovascular accidents, constraint-induced intervention in
hemiplegic cerebral palsy (CP), and restraining the use of unaffected eye in amblyopia all share a
conceptually similar rehabilitative methodology, aiming to restore lost function in the affected
side of the body. In this study, we hypothesize that the mechanism of constraint-induced therapy
in the injured corticospinal system in hemiplegic CP can be demonstrated when simplified neural
descending tracts with two layers of spiking neurons representing cortical and spinal neural
substrate are simulated with spike-timing-dependent plasticity (STDP). Although we could have
picked other injury, we chose an example from a study of kitten with hemiplegic CP (Martin,
2005; Friel and Martin, 2007; Martin et al., 2011) to simulate in order to demonstrate the entire
process—from the development of hemiplegic CP to restoration to the normal CST connectivity
by constraint-induced therapy as a treatment.
5.4.1 Relevance of the simulated neural structure to the biology
The two-layer neural structure is used to represent a simplistic descending corticospinal system
(Fig. 20). Four synaptic connections from input neuron substrate to output neuron substrate
represent the contralateral and the ipsilateral corticospinal projection from left and right
hemispheres to the spinal termination in gray matter. This structure forms competition sites at
each output neurons, competitions between the contralateral projection from one hemisphere and
the ipsilateral projection from the other hemisphere, which represents activity-dependent
competition for spinal synaptic space based on physiological reports in primates including
human (Nathan et al., 1990; Lacroix et al., 2004; Eyre et al., 2007). The stages of development
of and treatment from hemiplegic CP in simulation in present study (fig. 24B) are relevant to
hemiplegic CP in human. Electrophysiological studies confirm that persistent unilateral
abnormality by perinatal stroke is highly predictive of the development of hemiplegia (Kato et
al., 2004). Significant hypertrophy of the CST observed arising from the nonimpaired
hemisphere (Scales and Collins, 1972), and a reciprocal relation between the diameter of
104
ipsilateral axons from the nonimpaired hemisphere and that of contralateral axons from the
impaired hemisphere (Eyre et al., 2007) supports the activity-dependent competition hypothesis
at spinal cord termination (Martin and Lee, 1999).
5.4.2 Similarity to the treatment of amblyopia
Amblyopia is a unilateral reduction of visual acuity that bears a resemblance to the
impairment in the corticospinal system by early brain injury, as it is referred as hemiplegic CP in
the visual system. The neural architecture of binocular vision includes the binocular zone of
visual cortex that responds to projection from both eyes via lateral geniculate nucleus
(Casagrande and Boyd, 1996). It is known that monocular deprivation in the developmental
period leads to reduction of neurons in the visual cortex driven by binocular eyes and also the
reduction of neuron driven by amblyopic eye (Wiesel and Hubel, 1963, 1965; Shatz and Stryker,
1978). Therefore it can be said that the development of amblyopia due to monocular deprivation
resembles the development of hemiplegic CP due to unilateral injury in the developing brain in
that in both cases activity-dependent synaptic competition is taking place in the activity-
receiving neuron: neuron in the spinal gray matter in hemiplegic CP and neuron in the visual
cortex in case of amblyopia. The treatment for amblyopia is remarkably in sync with the concept
of constraint-induced therapy. The patching of the dominant eye while forcing the child to use
the weaker eye is essentially applying a transient constraint to the stronger side in order to revive
the visual impairment of the weaker eye. The similar principle being effective demonstrates that
the principle of constraint-induced therapy can be highly effective across the modalities in the
biological system as long as inter-hemispheric competition is present.
Previously, the treatment mechanism for amblyopia has been explained within the framework of
BCM theory, a rate-based synaptic modification theory (Cooper and Bear, 2012). Here, we
attempt to provide a simple yet fundamental mechanism that accounts for the general biological
phenomena, which includes the treatment of amblyopia, that involve inter-hemispheric synaptic-
competition using spike-based plasticity theory.
5.4.3 Escaping from the suboptimal stable state
In the third stage of the experiment (“bilateral activation”), we observe a persistence of
suboptimal stable solution due to hysteresis effect. The bilateral projection from one hemisphere
105
was due to a history of transient blockade of activity in the affected hemisphere. It is suboptimal
because right intervention will move it to a better state. It is interesting phenomenon that
traditional learning theories do not explain. Indeed, STDP could be the synaptic mechanism that
allows the persistent suboptimal solution at a synapse level. Given that this competition
mechanism proposed based on STDP is true, the same mechanism could be the thing that is
operative everywhere when there is a persistent suboptimal solution in a biological system
because spike-based plasticity proposed here is a generalizable mechanism.
5.4.4 Sensitivity of the result to other synaptic learning models
The question could be asked of how sensitive the result will be to the various synapse learning
models. Although it is not of our interest in this study to test the different effects of various
implementations of STDP update mechanisms, it can be said that the result is generalizable if 1)
update mechanisms dictate that correlated pre- and postsynaptic spikes causes potentiation and
uncorrelated pre- and postsynaptic spikes causes depression of the synaptic weight, 2) synaptic
weight has upper limit to prevent infinite growth, 3) postsynaptic firing rates are not in the range
that results in net depression and 4) synaptic decay is present to implement decrease in synaptic
weight on inactivity. Our implementation has asymmetric STDP curve, and is all-to-all (Pfister
and Gerstner, 2006) and standard additive. Variation in curve shape, nearest-neighbor
implementation instead of all-to-all, multiplicative in stead of additive update will still generate
the emergent phenomena of constraint-induced intervention in this study if the above conditions
are met.
5.4.5 Caveats in the simulation of biological systems
Is the criticism “simulations doomed to succeed” valid in this study? If the simulation is
attempting to explain the phenomena of a biological system, the statement is not true. In fact,
simulation do fail if we do not account for the sufficient complexity of the neural circuit to
explain a biological phenomenon, and studying the conditions that make it doom to fail is of
scientific importance by itself. The statement simply reminds us the fact that there must be a
verification of the simulation conditions and results when evaluating a simulation. The questions
to ask in the verification process is whether the parameters used are biologically realistic,
whether the theories used are well-established and sound, otherwise the simulation could be
106
deceptively manipulated to produce whatever result a researcher wants. In this study, we used
well-established biologically realistic models of neurons (Izhikevich, 2003a), synapse with
spike-based learning rule (Bi and Poo, 2001; Froemke and Dan, 2002), and realistic conditions,
e.g. synaptic weight with lower and upper limits, the existence of synaptic decay, etc., as well as
biologically realistic range of parameters that generate realistic output range using the models.
The simulation failed to produce a valid synaptic competition due to STDP if the synaptic weight
limit are not set, which will lead to unrealistic ever-growing synapse. Although our two-layer
neuronal structure of the descending corticospinal system is clearly an over-simplification that
does not compute the effects of other descending tracts, e.g. rubrospinal tract, it still provides a
valuable insight of how constraint-induced therapy might work in minimal complexity using
spike-based learning model. Provided that simulation is properly verified, it is useful in making
prediction, explanation, retrodiction and in exploring emergent explanation (Grim et al., 2013)
that are otherwise impossible in real physiology due to practical and ethical reasons.
5.5 Conclusion
In this study we presented a simplified yet fundamental mechanism that provides explanation for
some of the key phenomena in constraint-induced therapy by spike-based plasticity in synaptic
level. By simulating the activity-dependent synaptic competition as a key mechanism that might
harness synaptic plasticity to repair damaged coritocispinal system, this study suggests general
principle of how biological system could escape from a suboptimal stable state by applying a
forced transient constraint to a more competitive side of a system in order to transition into an
improve state which otherwise could result in the suboptimal pathological deficit to become
fixated.
5.6 Author Contributions
Designed the experiment: WS TS. Performed the experiment: WS. Developed the hardware and
visual software: WS. Analyzed the data: WJ. Wrote the paper: WS. Presented by WS at the
Annual Meeting of the Computational and Systems Neuroscience (Cosyne) in March 2015.
Reviewed manuscript: WS TS.
107
6 Conclusions and future work
6.1 Conclusion
Movement disorders are neurological conditions that affect speed, fluency, quality, and ease of
movement in a negative direction. In that regard, investigating the neurological underpinning of
the cause of the movement disorder is desirable. The emulation study presented here is one of the
alternative responses to overcome the limitation posed by human studies due to practical and
ethical reasons. My study is centered on understanding the dystonia and the study is extended to
understanding the neural mechanism of the constraint-induced rehabilitative intervention which
is highly effective across multiple modalities in biological systems.
The first study proposes two plausible neurological mechanisms that lead to behavioral
characteristics of dystonia and the outcomes from the emulation are compared with available
data from subjects with dystonia (chap 3). The study has shown through neuromorphic
emulation that there are at least two possible mechanisms of cortical abnormality that could
cause increased long-latency stretch reflexes and result in the clinical phenomenon of dystonia.
While an emulation of this type does not prove that this is the mechanism of dystonia, it does
provide evidence that a mechanism of this type is sufficient to cause the clinical features of
dystonia, and is, therefore, worthy of further study. We hope that these results provide both
insight and guidance for future clinical studies to test whether reducing the long-latency stretch
reflex may alleviate symptoms of dystonia.
The second study investigates the origin and development of motor overflow in focal hand
dystonia in the context of spike-based plasticity mechanism (chap 4). The purpose of this study
was to understand the pathology of developing focal hand dystonia by using biologically realistic
neural structure to test our temporal correlation hypothesis. As predicted by Sanger and
Merzenich (Sanger and Merzenich, 2000), and demonstrated by Merzenich and Byl (Byl et al.,
1996), correlated sensory inputs lead to sensory disorganization and have the potential to
108
produce motor dysfunction. Here we extended this work using a more realistic plasticity model
with spike-timing-dependent plasticity.
The third study investigates the mechanism of constraint-induced therapy, a popular
rehabilitative method in impaired biological systems with spike-based plasticity mechanism
(chap 5). This study presents a simplified yet fundamental mechanism that provides an
explanation for some of the key phenomena in constraint-induced therapy by spike-based
plasticity in synaptic level. By simulating the activity-dependent synaptic competition as a key
mechanism that might harness synaptic plasticity to repair damaged coritospinal system, this
study suggests general principle of how biological system could escape from a suboptimal stable
state by applying a forced transient constraint to a more competitive side of a system in order to
transition into an improved state which otherwise could result in a suboptimal pathological
deficit to become consolidated.
The second and the third study use the same neural structure that is assembled in a different
way to emulate two different biological phenomena. The neurons in the input and output
substrates represent sensory afferents to sensory representation in the cortex in the second study
whereas they represent neurons on the motor cortex and spinal neurons in the corticospinal
projection in the third study. The major difference between the studies is that the input current
profile to the input substrate being different. The second study investigates the effect of
correlated sensory inputs to the development of crosstalk and also shows transient correlated
input could lead to perpetuated structural motor overflow. The third study investigates the effect
of differential current input—on or off at times—in developing bilateral projection and
producing a therapeutic effect from a constraint. The model in this study includes synaptic decay
as a mean to diminish the strength of the synaptic connection when the input is inactivated.
Synaptic decay is an essential assumption that leads to the result obtained by this study.
The studies required extremely fast and customizable hardware, which provides a unique
benefit of accelerated emulation of the development of neurological system under relevant
circumstances, at an affordable cost without having to spend on a system with extremely high
computing power such as that of supercomputer. It required multi-scale emulation for the
biological system both in time domain and space domain. Multi-scale in time domain accounts
for the drastically different time scales that neural processes in developmental diseases usually
operate on, e.g. spinal reflex in milliseconds versus learning in years, and multi-scale in space in
109
biological system accounts for the details, ranging from cellular level spiking activity to behavior
level limb biomechanics. The spike level cellular activity was the highest level of abstraction
essential to achieve sufficient biorealism in the neurological system that are still affordable to
implement in the available hardware space. This daunting task was achieved both by highly
customizing clock-level computation and efficient use of memory in commercially available
field-programmable gate arrays (FPGAs), and also by carefully designing scientific studies using
the built tool. The task was realized by our brilliant team members who are listed as authors in
the three studies.
The potential importance of this project was to gain knowledge in complex interplay between
development, plasticity, behavior and neurological injury because understanding the plastic
mechanism of how early brain injury leads to developmental disorder could be used to guide
early intervention such as in constraint-induced therapy. From the study of motor overflow and
constraint-induced therapy, we observed plasticity can work in both way: adaptive plasticity
taking place by transient inactivity (constraint) that facilitates the recovery of the injured
function, and also maladaptive plasticity that consolidates the suboptimal pathological state
which frustrates the recovery of the injured function unless effective intervention is applied.
The first study is published in Journal of Neural Engineering. The second and third studies
will be submitted to relevant journals (undecided at this point). The engineering technique and
general methodology behind the use of programmable hardware (chap 2) is published in neural
information processing systems (NIPS).
6.2 Future studies
Future studies should further focus on the unique advantage of using emulation as a way to
study and answer questions about finding the sufficient mechanism responsible for the movement
disorders that have a neurological origin. Since we have built and modularized many
neurological components that are proven to work, we can utilize the built library to test specific
disease hypotheses by building a neural circuit around a structure that is relevant. We can
conduct experiment that is generally prohibited in human studies due to practical and ethical
reasons, especially by maximizing the high-speed acceleration that the technology provides to
110
study plasticity and learning effects that span several years for development. We have used the
system around a sensorimotor reflex pathway to study the immediate effects of the change in
physiological parameters but did not incorporate the “learning” synapse in the study of dystonia.
The future study could strive to close the loop in the dystonia study to emulate the plasticity
effect of many pharmacological treatments, deep brain stimulation, cell death, and behavior
training. The general emulation platform is capable of opening a new line of projects that is non-
dystonia studies. We could expand the currently available single joint system to multi-joint
system to emulate the various postural and behavioral characteristics of tic, Parkinsonism, tremor,
and ataxia, etc. Because we believe the only way to test and characterize the high-level behavior
of a brain model is to actually build the closed loop system between the artificial nervous system
and the body (plant) acting in an environment, by elaborating the brain model we can interrogate
the system through a well-designed experiment to acquire a unique and practical knowledge
about the mechanism of biological phenomena and become a step closer to treating the disorders
by studying how the system breaks.
111
7 Bibliography
Air E L, Ostrem J L, Sanger T D and Starr P A 2011 Deep brain stimulation in children: experience and
technical pearls Journal of neurosurgery. Pediatrics 8 566-74
Albanese A, Bhatia K, Bressman S B, Delong M R, Fahn S, Fung V S, Hallett M, Jankovic J, Jinnah H A,
Klein C, Lang A E, Mink J W and Teller J K 2013 Phenomenology and classification of dystonia:
a consensus update Movement disorders : official journal of the Movement Disorder Society 28
863-73
Attebo K, Mitchell P, Cumming R, Smith W, Jolly N and Sparkes R 1998 Prevalence and causes of
amblyopia in an adult population Ophthalmology 105 154-9
Bara-Jimenez W, Catalan M J, Hallett M and Gerloff C 1998 Abnormal somatosensory homunculus in
dystonia of the hand Annals of neurology 44 828-31
Bara-Jimenez W, Shelton P and Hallett M 2000a Spatial discrimination is abnormal in focal hand
dystonia Neurology 55 1869-73
Bara-Jimenez W, Shelton P, Sanger T D and Hallett M 2000b Sensory discrimination capabilities in
patients with focal hand dystonia Annals of neurology 47 377-80
Beck S, Richardson S P, Shamim E A, Dang N, Schubert M and Hallett M 2008 Short intracortical and
surround inhibition are selectively reduced during movement initiation in focal hand dystonia The
Journal of neuroscience : the official journal of the Society for Neuroscience 28 10363-9
Bi G and Poo M 2001 Synaptic modification by correlated activity: Hebb's postulate revisited Annual
review of neuroscience 24 139-66
Bienenstock E L, Cooper L N and Munro P W 1982 Theory for the development of neuron selectivity:
orientation specificity and binocular interaction in visual cortex The Journal of neuroscience : the
official journal of the Society for Neuroscience 2 32-48
Blake D T and Merzenich M M 2002 Changes of AI receptive fields with sound density Journal of
neurophysiology 88 3409-20
Brady K and Garcia T 2009 Constraint-induced movement therapy (CIMT): pediatric applications
Developmental disabilities research reviews 15 102-11
Brashear A and Elovic E 2010 Spasticity: Diagnosis and Management: Demos Medical)
Breakefield X O, Blood A J, Li Y, Hallett M, Hanson P I and Standaert D G 2008 The pathophysiological
basis of dystonias Nature reviews. Neuroscience 9 222-34
Byl N N, Merzenich M M and Jenkins W M 1996 A primate genesis model of focal dystonia and
repetitive strain injury: I. Learning-induced dedifferentiation of the representation of the hand in
the primary somatosensory cortex in adult monkeys Neurology 47 508-20
Capaday C, Forget R, Fraser R and Lamarre Y 1991 Evidence for a contribution of the motor cortex to
the long-latency stretch reflex of the human thumb The Journal of physiology 440 243-55
112
Carpaneto J, Micera S, Galardi G, Micheli A, Carboncini M C, Rossi B and Dario P 2004 A protocol for
the assessment of 3D movements of the head in persons with cervical dystonia Clinical
biomechanics 19 659-63
Carr L J, Harrison L M, Evans A L and Stephens J A 1993 Patterns of central motor reorganization in
hemiplegic cerebral palsy Brain : a journal of neurology 116 ( Pt 5) 1223-47
Casagrande V A and Boyd J D 1996 The neural architecture of binocular vision Eye 10 ( Pt 2) 153-60
Ceballos-Baumann A O 1994 Overactivity of primary and associated motor areas in secondary hemi-
dystonia (SHD) due to thalamic or basal ganglia lesions: a PET study. Movement disorders :
official journal of the Movement Disorder Society 9 238
Chen R, Cohen L G and Hallett M 2002 Nervous system reorganization following injury Neuroscience
111 761-73
Colton C A, Brown C M, Czapiga M and Vitek M P 2002 Apolipoprotein-E allele-specific regulation of
nitric oxide production Annals of the New York Academy of Sciences 962 212-25
Cooper L N and Bear M F 2012 The BCM theory of synapse modification at 30: interaction of theory
with experiment Nature reviews. Neuroscience 13 798-810
De Luca C J and Hostage E C 2010 Relationship between firing rate and recruitment threshold of
motoneurons in voluntary isometric contractions. Journal of neurophysiology 104 1034-46
DeLong M R and Wichmann T 2007 Circuits and circuit disorders of the basal ganglia Archives of
neurology 64 20-4
Edwards M J, Huang Y Z, Mir P, Rothwell J C and Bhatia K P 2006 Abnormalities in motor cortical
plasticity differentiate manifesting and nonmanifesting DYT1 carriers Movement disorders :
official journal of the Movement Disorder Society 21 2181-6
Edwards M J, Huang Y Z, Wood N W, Rothwell J C and Bhatia K P 2003 Different patterns of
electrophysiological deficits in manifesting and non-manifesting carriers of the DYT1 gene
mutation Brain : a journal of neurology 126 2074-80
Elbert T, Candia V, Altenmuller E, Rau H, Sterr A, Rockstroh B, Pantev C and Taub E 1998 Alteration of
digital representations in somatosensory cortex in focal hand dystonia Neuroreport 9 3571-5
Emonet-Denand F, Laporte Y, Matthews P B and Petit J 1977 On the subdivision of static and dynamic
fusimotor actions on the primary ending of the cat muscle spindle The Journal of physiology 268
827-61
Evarts E V and Tanji J 1976 Reflex and intended responses in motor cortex pyramidal tract neurons of
monkey Journal of neurophysiology 39 1069-80
Eyre J A 2007 Corticospinal tract development and its plasticity after perinatal injury Neuroscience and
biobehavioral reviews 31 1136-49
Eyre J A, Smith M, Dabydeen L, Clowry G J, Petacchi E, Battini R, Guzzetta A and Cioni G 2007 Is
hemiplegic cerebral palsy equivalent to amblyopia of the corticospinal system? Annals of
neurology 62 493-503
113
Eyre J A, Taylor J P, Villagra F, Smith M and Miller S 2001 Evidence of activity-dependent withdrawal
of corticospinal projections during human development Neurology 57 1543-54
Fahn S 1988 Concept and classification of dystonia Advances in neurology 50 1-8
Farmer S F, Harrison L M, Ingram D A and Stephens J A 1991 Plasticity of central motor pathways in
children with hemiplegic cerebral palsy Neurology 41 1505-10
Friel K M and Martin J H 2007 Bilateral activity-dependent interactions in the developing corticospinal
system The Journal of neuroscience : the official journal of the Society for Neuroscience 27
11083-90
Froemke R C and Dan Y 2002 Spike-timing-dependent synaptic modification induced by natural spike
trains Nature 416 433-8
Fuglevand A J, Winter D A and Patla A E 1993a Models of recruitment and rate coding organization in
motor-unit pools Journal of neurophysiology 70 2470-88
Fuglevand A J, Winter D A and Patla A E 1993b Models of recruitment and rate coding organization in
motor-unit pools. Journal of neurophysiology 70 2470-88
Gelfan S, Kao G and Ruchkin D S 1970 The dendritic tree of spinal neurons. The Journal of comparative
neurology 139 385-411
Ghez C, Gordon J and Hening W 1988 Trajectory control in dystonia Advances in neurology 50 141-55
Glowatzki E and Fuchs P A 2002 Transmitter release at the hair cell ribbon synapse Nature neuroscience
5 147-54
Godde B, Spengler F and Dinse H R 1996 Associative pairing of tactile stimulation induces
somatosensory cortical reorganization in rats and humans Neuroreport 8 281-5
Gordon L M, Keller J L, Stashinko E E, Hoon A H and Bastian A J 2006 Can spasticity and dystonia be
independently measured in cerebral palsy? Pediatric neurology 35 375-81
Grim P, Rosenberger R, Rosenfeld A, Anderson B and Eason R 2013 How simulations fail Synthese 190
2367–90
Hallett M 2006 Pathophysiology of writer's cramp Human movement science 25 454-63
Hallett M 2011 Neurophysiology of dystonia: The role of inhibition Neurobiology of disease 42 177-84
Hashimoto T, Naito K, Kitazawa K, Imai S and Goto T 2010 Pallidotomy for severe tardive jaw-opening
dystonia Stereotactic and functional neurosurgery 88 105-8
Hendricks H T, Pasman J W, van Limbeek J and Zwarts M J 2003 Motor evoked potentials in predicting
recovery from upper extremity paralysis after acute stroke Cerebrovascular diseases 16 265-71
Henneman E 1957 Relation between size of neurons and their susceptibility to discharge. Science (New
York, N.Y.) 126 1345-7
Henneman E, Somjen G and Carpenter D O 1965 Functional Significance of Cell Size in Spinal
Motoneurons Journal of neurophysiology 28 560-80
114
Hill A V 1938 The heat of shortening and dynamics constants of muscles. Proc. R. Soc. Lond. B 126 136-
95
Hoon A H, Jr., Stashinko E E, Nagae L M, Lin D D, Keller J, Bastian A, Campbell M L, Levey E, Mori S
and Johnston M V 2009 Sensory and motor deficits in children with cerebral palsy born preterm
correlate with diffusion tensor imaging abnormalities in thalamocortical pathways Developmental
medicine and child neurology 51 697-704
Imer M, Ozeren B, Karadereler S, Yapici Z, Omay B, Hanagasi H and Eraksoy M 2005 Destructive
stereotactic surgery for treatment of dystonia Surgical neurology 64 Suppl 2 S89-94; discussion
S-5
Izhikevich E M 2003a Simple model of spiking neurons IEEE transactions on neural networks / a
publication of the IEEE Neural Networks Council 14 1569-72
Izhikevich E M 2003b Simple model of spiking neurons. IEEE transactions on neural networks / a
publication of the IEEE Neural Networks Council 14 1569-72
Jankovic J 2001 Can peripheral trauma induce dystonia and other movement disorders? Yes! Movement
disorders : official journal of the Movement Disorder Society 16 7-12
Kaas J H and Florence S L 1997 Mechanisms of reorganization in sensory systems of primates after
peripheral nerve injury Advances in neurology 73 147-58
Kahn J, Anderton B H, Gibb W R, Lees A J, Wells F R and Marsden C D 1985 Neuronal filaments in
Alzheimer's, Pick's, and Parkinson's diseases The New England journal of medicine 313 520-1
Kato T, Okumura A, Hayakawa F, Kuno K and Watanabe K 2004 Electroencephalographic aspects of
periventricular hemorrhagic infarction in preterm infants Neuropediatrics 35 161-6
Kempter R, Gerstner W and von Hemmen J L 1999 Hebbian learning and spiking neurons Phys Rev E 59
4498-514
Kojovic M, Parees I, Kassavetis P, Palomar F J, Mir P, Teo J T, Cordivari C, Rothwell J C, Bhatia K P
and Edwards M J 2013 Secondary and primary dystonia: pathophysiological differences Brain : a
journal of neurology 136 2038-49
Konczak J and Abbruzzese G 2013 Focal dystonia in musicians: linking motor symptoms to
somatosensory dysfunction Frontiers in human neuroscience 7 297
Krauss J K 2010 Surgical treatment of dystonia European journal of neurology : the official journal of the
European Federation of Neurological Societies 17 Suppl 1 97-101
Krutki P, Ciechanowicz I, Celichowski J and Cywinska-Wasilewska G 2008 Comparative analysis of
motor unit action potentials of the medial gastrocnemius muscle in cats and rats Journal of
electromyography and kinesiology : official journal of the International Society of
Electrophysiological Kinesiology 18 732-40
Kukke S N and Sanger T D 2011 Contributors to excess antagonist activity during movement in children
with secondary dystonia due to cerebral palsy Journal of neurophysiology 105 2100-7
115
Lacroix S, Havton L A, McKay H, Yang H, Brant A, Roberts J and Tuszynski M H 2004 Bilateral
corticospinal projections arise from each motor cortex in the macaque monkey: a quantitative
study The Journal of comparative neurology 473 147-61
Lebiedowska M K, Gaebler-Spira D, Burns R S and Fisk J R 2004 Biomechanic characteristics of
patients with spastic and dystonic hypertonia in cerebral palsy Archives of physical medicine and
rehabilitation 85 875-80
Lee R G, Murphy J T and Tatton W G 1983 Long-latency myotatic reflexes in man: mechanisms,
functional significance, and changes in patients with Parkinson's disease or hemiplegia Advances
in neurology 39 489-508
Levin M F and Feldman A G 1994 The role of stretch reflex threshold regulation in normal and impaired
motor control. Brain research 657 23-30
Lomont C 2003 Fast inverse square root. In: Technical Report: Available at
http://www.lomont.org/Math/Papers/2003/InvSqrt.pdf)
Martin J H 2005 The corticospinal system: from development to motor control The Neuroscientist : a
review journal bringing neurobiology, neurology and psychiatry 11 161-73
Martin J H, Chakrabarty S and Friel K M 2011 Harnessing activity-dependent plasticity to repair the
damaged corticospinal tract in an animal model of cerebral palsy Developmental medicine and
child neurology 53 Suppl 4 9-13
Martin J H, Choy M, Pullman S and Meng Z 2004 Corticospinal system development depends on motor
experience The Journal of neuroscience : the official journal of the Society for Neuroscience 24
2122-32
Martin J H and Lee S J 1999 Activity-dependent competition between developing corticospinal
terminations Neuroreport 10 2277-82
Matthews P B 1991 The human stretch reflex and the motor cortex Trends in neurosciences 14 87-91
McIntyre C C, Savasta M, Kerkerian-Le Goff L and Vitek J L 2004 Uncovering the mechanism(s) of
action of deep brain stimulation: activation, inhibition, or both Clinical neurophysiology : official
journal of the International Federation of Clinical Neurophysiology 115 1239-48
Mileusnic M P, Brown I E, Lan N and Loeb G E 2006a Mathematical models of proprioceptors. I.
Control and transduction in the muscle spindle Journal of neurophysiology 96 1772-88
Mileusnic M P, Brown I E, Lan N and Loeb G E 2006b Mathematical models of proprioceptors. I.
Control and transduction in the muscle spindle. Journal of neurophysiology 96 1772-88
Morimoto C, Distaso J A, Cheney J J, Reinherz E L and Schlossman S F 1984 Cellular interaction
between subsets of T8 population for maximal suppression of antigen-specific antibody response
Cellular immunology 88 75-84
Murray W M, Delp S L and Buchanan T S 1995 Variation of muscle moment arms with elbow and
forearm position Journal of biomechanics 28 513-25
Nathan P W, Smith M C and Deacon P 1990 The corticospinal tracts in man. Course and location of
fibres at different segmental levels Brain : a journal of neurology 113 ( Pt 2) 303-24
116
Nelles G, Spiekermann G, Jueptner M, Leonhardt G, Muller S, Gerhard H and Diener H C 1999
Reorganization of sensory and motor systems in hemiplegic stroke patients. A positron emission
tomography study Stroke; a journal of cerebral circulation 30 1510-6
Niu C M, Nandyala S K and Sanger T D 2014 Emulated muscle spindle and spiking afferents validates
VLSI neuromorphic hardware as a testbed for sensorimotor function and disease Frontiers in
computational neuroscience 8 141
Nudo R J 2013 Recovery after brain injury: mechanisms and principles Frontiers in human neuroscience
7 887
Palmer E and Ashby P 1992 Evidence that a long latency stretch reflex in humans is transcortical The
Journal of physiology 449 429-40
Pfister J P and Gerstner W 2006 Triplets of spikes in a model of spike timing-dependent plasticity The
Journal of neuroscience : the official journal of the Society for Neuroscience 26 9673-82
Playford E D, Fletcher N A, Sawle G V, Marsden C D and Brooks D J 1993 Striatal [18F]dopa uptake in
familial idiopathic dystonia Brain : a journal of neurology 116 ( Pt 5) 1191-9
Prescott I A, Dostrovsky J O, Moro E, Hodaie M, Lozano A M and Hutchison W D 2013 Reduced paired
pulse depression in the basal ganglia of dystonia patients Neurobiology of disease 51 214-21
Pruszynski J A, Kurtzer I, Nashed J Y, Omrani M, Brouwer B and Scott S H 2011 Primary motor cortex
underlies multi-joint integration for fast feedback control Nature 478 387-90
Quartarone A, Bagnato S, Rizzo V, Siebner H R, Dattola V, Scalfari A, Morgante F, Battaglia F, Romano
M and Girlanda P 2003 Abnormal associative plasticity of the human motor cortex in writer's
cramp Brain : a journal of neurology 126 2586-96
Raphael G, Tsianos G A and Loeb G E 2010 Spinal-like regulator facilitates control of a two-degree-of-
freedom wrist. The Journal of neuroscience : the official journal of the Society for Neuroscience
30 9431-44
Ridding M C, Sheean G, Rothwell J C, Inzelberg R and Kujirai T 1995 Changes in the balance between
motor cortical excitation and inhibition in focal, task specific dystonia Journal of neurology,
neurosurgery, and psychiatry 59 493-8
Rodriguez I, Gila L, Malanda A, Gurtubay I G, Mallor F, Gomez S, Navallas J and Rodriguez J 2007
Motor unit action potential duration, I: variability of manual and automatic measurements
Journal of clinical neurophysiology : official publication of the American
Electroencephalographic Society 24 52-8
Sanger T D 2010 Neuro-mechanical control using differential stochastic operators. In: Engineering in
Medicine and Biology Society (EMBC), 2010 Annual International Conference of the IEEE, pp
4494-7
Sanger T D 2011a Distributed control of uncertain systems using superpositions of linear operators
Neural computation 23 1911-34
Sanger T D 2011b Distributed control of uncertain systems using superpositions of linear operators.
Neural computation 23 1911-34
117
Sanger T D, Delgado M R, Gaebler-Spira D, Hallett M, Mink J W and Task Force on Childhood Motor D
2003 Classification and definition of disorders causing hypertonia in childhood Pediatrics 111
e89-97
Sanger T D and Merzenich M M 2000 Computational model of the role of sensory disorganization in
focal task-specific dystonia Journal of neurophysiology 84 2458-64
Sanger T D, Tarsy D and Pascual-Leone A 2001 Abnormalities of spatial and temporal sensory
discrimination in writer's cramp Movement disorders : official journal of the Movement Disorder
Society 16 94-9
Scales D A and Collins G H 1972 Cerebral degeneration with hypertrophy of the contralateral pyramid
Archives of neurology 26 186-90
Scheidt R A and Ghez C 2007 Separate adaptive mechanisms for controlling trajectory and final position
in reaching Journal of neurophysiology 98 3600-13
Scontrini A, Conte A, Defazio G, Fiorio M, Fabbrini G, Suppa A, Tinazzi M and Berardelli A 2009
Somatosensory temporal discrimination in patients with primary focal dystonia Journal of
neurology, neurosurgery, and psychiatry 80 1315-9
Shadmehr R and Wise S P 2005 (Cambridge, MA: MIT Press) pp 1-18
Shatz C J and Stryker M P 1978 Ocular dominance in layer IV of the cat's visual cortex and the effects of
monocular deprivation The Journal of physiology 281 267-83
Sitburana O and Jankovic J 2008 Focal hand dystonia, mirror dystonia and motor overflow Journal of the
neurological sciences 266 31-3
Song S, Miller K D and Abbott L F 2000 Competitive Hebbian learning through spike-timing-dependent
synaptic plasticity Nature neuroscience 3 919-26
Soska K C, Galeon M A and Adolph K E 2012 On the other hand: overflow movements of infants' hands
and legs during unimanual object exploration Developmental psychobiology 54 372-82
Suminski A J, Rao S M, Mosier K M and Scheidt R A 2007 Neural and electromyographic correlates of
wrist posture control Journal of neurophysiology 97 1527-45
Taub E, Griffin A, Nick J, Gammons K, Uswatte G and Law C R 2007 Pediatric CI therapy for stroke-
induced hemiparesis in young children Developmental neurorehabilitation 10 3-18
Taub E, Uswatte G and Pidikiti R 1999 Constraint-Induced Movement Therapy: a new family of
techniques with broad application to physical rehabilitation--a clinical review Journal of
rehabilitation research and development 36 237-51
Tinazzi M, Frasson E, Bertolasi L, Fiaschi A and Aglioti S 1999 Temporal discrimination of somesthetic
stimuli is impaired in dystonic patients Neuroreport 10 1547-50
Trompetto C, Avanzino L, Marinelli L, Mori L, Pelosin E, Roccatagliata L and Abbruzzese G 2012
Corticospinal excitability in patients with secondary dystonia due to focal lesions of the basal
ganglia and thalamus Clinical neurophysiology : official journal of the International Federation
of Clinical Neurophysiology 123 808-14
118
Valero-Cuevas F J, Yi J-W, Brown D, McNamara R V, Paul C and Lipson H 2007 The tendon network of
the fingers performs anatomical computation at a macroscopic scale. IEEE transactions on bio-
medical engineering 54 1161-6
van Doornik J, Kukke S and Sanger T D 2009 Hypertonia in childhood secondary dystonia due to
cerebral palsy is associated with reflex muscle activation Movement disorders : official journal of
the Movement Disorder Society 24 965-71
Vandermeeren Y, Bastings E, Fadiga L and Olivier E 2003 Long-latency motor evoked potentials in
congenital hemiplegia Clinical neurophysiology : official journal of the International Federation
of Clinical Neurophysiology 114 1808-18
Vidailhet M, Vercueil L, Houeto J L, Krystkowiak P, Benabid A L, Cornu P, Lagrange C, Tezenas du
Montcel S, Dormont D, Grand S, Blond S, Detante O, Pillon B, Ardouin C, Agid Y, Destee A,
Pollak P and French Stimulation du Pallidum Interne dans la Dystonie Study G 2005 Bilateral
deep-brain stimulation of the globus pallidus in primary generalized dystonia The New England
journal of medicine 352 459-67
Weise D, Schramm A, Stefan K, Wolters A, Reiners K, Naumann M and Classen J 2006 The two sides of
associative plasticity in writer's cramp Brain : a journal of neurology 129 2709-21
Wiesel T N and Hubel D H 1963 Single-Cell Responses in Striate Cortex of Kittens Deprived of Vision
in One Eye Journal of neurophysiology 26 1003-17
Wiesel T N and Hubel D H 1965 Comparison of the effects of unilateral and bilateral eye closure on
cortical unit responses in kittens Journal of neurophysiology 28 1029-40
Young S J, van Doornik J and Sanger T D 2011 Finger muscle control in children with dystonia
Movement disorders : official journal of the Movement Disorder Society 26 1290-6
Abstract (if available)
Abstract
There is currently no quantitative model of how the functions of neurons affect the specific abnormalities observed in movement disorders. Although clinical experience has provided insight in making qualitative prediction of how certain kinds of injury might lead to particular outcome, qualitative clinical insight itself is of little use if the goal is to make a specific prediction on the effect of particular impairment–whether they are neuronal or anatomic in nature, based on specific quantifiable physiological measures in individual patients. In order to study the causal relationship between a particular neuronal injury and the resultant immediate or long-term biomechanical effect on the movement of a patient, we designed a multi-purpose high-speed emulation platform in scalable hardware. In this project, the platform is designed to emulate a subset of human sensorimotor nervous system that is speculated to be responsible for many movement disorders when it is impaired. Technical preparation section (chap 2) is dedicated for the methodological considerations. The structures of fundamental building blocks of the monosynaptic spinal stretch reflex pathway, including spiking neurons, spindle, muscle, and synapse, etc., as well as how fast computation are achieved in customizable hardware are described. Although the thesis is centered on understanding the pathology of dystonia, the third study (chap 5) extends the use of the technology to understand the mechanism for the popular rehabilitative method called constraint-induced therapy.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
On the electrophysiology of multielectrode recordings of the basal ganglia and thalamus to improve DBS therapy for children with secondary dystonia
PDF
Facilitating myocontrol for children with cerebral palsy
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
Asset Metadata
Creator
Sohn, Won Joon Eric
(author)
Core Title
Understanding the pathology of dystonia by hardware emulation
School
Viterbi School of Engineering
Degree
Doctor of Philosophy
Degree Program
Biomedical Engineering
Publication Date
07/27/2015
Defense Date
06/10/2015
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
Biomedical Engineering,constraint-induced therapy,dystonia,emulation,FPGA,movement disorder,neuromorphic,OAI-PMH Harvest,simulation,synthetic neuromuscular system
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Sanger, Terence D. (
committee chair
), Parker, Alice C. (
committee member
), Prasanna, Viktor K. (
committee member
), Valero-Cuevas, Francisco J. (
committee member
)
Creator Email
wonjoons@usc.edu,wonjsohn@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c3-608740
Unique identifier
UC11299127
Identifier
etd-SohnWonJoo-3708.pdf (filename),usctheses-c3-608740 (legacy record id)
Legacy Identifier
etd-SohnWonJoo-3708.pdf
Dmrecord
608740
Document Type
Dissertation
Format
application/pdf (imt)
Rights
Sohn, Won Joon Eric
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 a...
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
Tags
constraint-induced therapy
dystonia
emulation
FPGA
movement disorder
neuromorphic
simulation
synthetic neuromuscular system