Primate Suborders Figure 10.1: Summary of traditional primate classification.
Neurophysiological Characterization of a Non-Human Primate ...
Transcript of Neurophysiological Characterization of a Non-Human Primate ...
I
Neurophysiological Characterization of a Non-Human Primate Model
of Spinal Cord Injury Utilizing Intramuscular Electromyography
by
Farah Masood
A Thesis
presented to
The University of Guelph
In partial fulfilment of the requirements for the degree of
Doctor of Philosophy
in
Engineering
Guelph, Ontario, Canada
© Farah, January 2020
III
ABSTRACT
NEUROPHYSIOLOGICAL CHARACTERIZATION OF A NON-HUMAN PRIMATE MODEL OF SPINAL CORD INJURY UTILIZING INTRAMUSCULAR
ELECTROMYOGRAPHY
The lack of a valid non-human primate (NHP) model of spinal cord injury (SCI) and
a robust assessment tool are the key reasons behind the low rate of successful clinical
trials for the development of SCI medication. Accordingly, developing such an NHP model
would fill the gap between preclinical and clinical studies. A histopathological analysis
would help to validate a developed NHP model, while neurophysiological
characterizations could offer a reliable assessment tool for the effect of the injury. The
goal of this work is to submit a neurophysiological analysis using intramuscular
electromyography (EMG) signals collected from the agonist-antagonist pair of tail
muscles of Macaca fasicularis during pre- and post-lesion, and for a treatment and control
groups. The main goal was achieved by setting three sub-goals to be addressed in this
work. First, the general volitional muscle activity was analyzed utilizing some basic
amplitude features. Second, an analysis of the motor unit discharge properties was
submitted using wavelet transforms (WT) and the relative power (RP) of the multi-
resolution EMG signals. Third, the development of an SCI classification system was
studied and analyzed by employing various classification approaches, including regular
machine learning (ML) and deep learning (DL) classification techniques
Farah Masood
University of Guelph, 2019
Dr. Hussein A. Abdullah
III
The findings were consistent with that of the literature. Evidence suggested that
the newly proposed general volitional muscle metric (Q-metric) was related directly to the
gross recruitment of motor units required for volitional muscle control pre- and post-lesion.
As well, the results of the multi-resolution analysis demonstrated the capability of utilizing
the WT and the RP as a decomposition method to characterize the discharge properties
of the dynamic muscle activity. The findings of the SCI classification analysis implied that
the proposed Q-metric and WT-RPs features could be used to build an SCI classification
system using K-nearest neighbors (KNN) technique. Finally, the deep learning
classification, which consisted exclusively of convolutional neural network (CNN) using
raw EMG segments as inputs, showed high potential and promising results for use as an
SCI classification system with an F-measure of 84.0% and 86.9%. for the left and the right
side.
III
DEDICATION
To the soul of my great father with love …
Mohammed Ridha Masood
To my wonderful mother and my first teacher …
Sabah Yousif
To my best friends and my sisters …
Noor and Shahad
To my wonderful husband who supports me and has always believed in me …
Dr. Wisam Al-Wajidi
To my lovely daughter and son who support me with their continuous love and patience…
Mina & Yousif
Thank you all for being my strong, inspiring, and devoted family…
You have made me what I am today
IV
ACKNOWLEDGEMENTS
First and foremost, I would like to thank and praise Allah, the Almighty, the most merciful
and compassionate, for His support, help, and generosity, and I pray to Him for guidance
in the future.
I would like to express my deepest appreciation to all those who have contributed directly
and indirectly to the completion of this dissertation.
I am very grateful to my academic supervisor, Dr. Hussein A. Abdullah, for his valuable
guidance, and patient encouragement throughout the work for this dissertation. Great
appreciation goes to Dr. Shanker Nesathurai for his mentoring and kind support through
the long journey of this research and for giving me the great opportunity to be part of his
distinguished team. I would like to extend my thanks to the members of my doctoral
advisory committee, Dr. Bob Dony, Dr. Karen Gordon, Dr. Gerarda Darlington, and Dr.
David Chiu for their constructive suggestions and recommendations. My appreciation also
to Dr. Graham Tylor, and my colleagues Dr. Nitin Seth, Dr. Noor Al-Qazzaz, and Dr. Abeer
Al-Hyari, for their time, accessibility, and advice throughout the research period. Personal
thanks are given to the Iraqi Cultural Counselor in Ottawa, Dr. Asaad Toma Omran, for
his help and support. Thank you to my colleagues in the Robotics Institutes Lab,
especially Cole Tarry and Patrick Wspanialy for their support.
I gratefully acknowledge the Iraqi Ministry of Higher Education and Scientific Research
(IMHESR) for providing me a scholarship to finish my Ph.D. degree in Canada.
V
Table of Contents
Contents ABSTRACT............................................................................................................................................... III
DEDICATION ............................................................................................................................................ III
ACKNOWLEDGEMENTS ...................................................................................................................... IV
Table of Contents .................................................................................................................................... V
List of Figures ...................................................................................................................................... VIII
List of Tables ......................................................................................................................................... XII
List of Abbreviations .......................................................................................................................... XIII
Chapter 1 .................................................................................................................................................... 1
Introduction ........................................................................................................................................... 1
1.1. Spinal Cord Injury .................................................................................................................... 1
1.2. Problem Formulation .............................................................................................................. 1
1.3. Research Objectives ............................................................................................................... 2
1.3. Research Challenges .............................................................................................................. 3
1.4. Non-human Primate Model .................................................................................................... 3
1.6. Research Approaches ............................................................................................................ 4
1.7. Impact of the Work ................................................................................................................... 5
1.8. The Structure of the Thesis ................................................................................................... 6
Chapter 2 .................................................................................................................................................... 7
Background and Related Work ........................................................................................................ 7
2.1. Introduction ............................................................................................................................... 7
2.2. Spinal Cord Injury .................................................................................................................... 7
2.3. Motor Neuron Disorder ........................................................................................................... 9
2.4. Nonhuman Primate Model ................................................................................................... 12
2.5. Electromyography Signal .................................................................................................... 13
2.6 EMG Signal Acquisition........................................................................................................ 16
2.7. SCI Assessment Tools.......................................................................................................... 17
2.8. EMG Signal Conditioning ..................................................................................................... 20
2.9. EMG Feature Extraction ....................................................................................................... 23
2.10. EMG Decomposition .......................................................................................................... 25
2.11. Wavelet Transforms .......................................................................................................... 28
VI
2.12. SCI Classification System ................................................................................................ 31
2.12.1 Machine Learning Classification Technique ........................................................... 33
2.12.2 Deep Learning Classification Technique ................................................................. 35
Chapter 3 .................................................................................................................................................. 39
Materials and Methodology ............................................................................................................. 39
3.1. Introduction ................................................................................................................................ 39
3.2. Materials and Experimental Procedures .............................................................................. 39
3.2.1. Subjects enrolled ................................................................................................................ 39
3.2.2. The initial surgical procedure to facilitate the collection of EMG data ................ 39
3.2.3. The subsequent surgical procedure to create an experimental spinal cord
injury .................................................................................................................................................. 42
3.2.4. Monitoring of subjects ....................................................................................................... 45
3.2.5. Euthanasia and post-mortem examination .................................................................. 45
3.3. EMG Analysis Methods ............................................................................................................. 47
3.3.1. EMG signal preprocessing ............................................................................................... 50
3.3.2. General volitional muscle activity metric ..................................................................... 51
3.3.3. Motor units discharge properties ................................................................................... 59
3..4 SCI Classification System ........................................................................................................ 71
3.4.1 SCI classification system using classical ML classification technique ................ 71
3.4.2 SCI classification using DL technique compared to ML technique ........................ 75
Chapter 4 .................................................................................................................................................. 81
Results and Discussions ................................................................................................................. 81
4.1. Introduction ................................................................................................................................. 81
4.2. General Volitional Muscle Activity Metric ............................................................................ 82
4.2.1. Conventional EMG metrics ............................................................................................... 82
4.2.2. The proposed peak number method (Q-metric) ......................................................... 85
4.3. An Analysis of the EMG Discharge Properties .................................................................. 88
4.3.1. Conventional EMG decomposition................................................................................. 88
4.3.2. The proposed EMG discharge properties analysis utilizing wavelet transform
decomposition and relative power ............................................................................................ 92
4.4. SCI Classification System ...................................................................................................... 101
4.4.1 Classification performance evaluation ........................................................................ 102
4.4.2. SCI classification system using classical ML classification technique ............. 105
VII
4.4.3. SCI classification using DL technique compared to ML technique..................... 115
4.5. Summary ................................................................................................................................ 121
Chapter 5 ................................................................................................................................................ 124
Conclusions and Future Work ...................................................................................................... 124
5.1. Conclusions ............................................................................................................................... 124
5.2. Future Work ............................................................................................................................... 128
References ............................................................................................................................................. 130
List of Publications .............................................................................................................................. 142
VIII
List of Figures
Figure 2.1 The main causes of SCI [4] ____________________________________________________ 8
Figure 2.2 The different types of paralysis resulted from SCI [19] _______________________________ 9
Figure 2.3 The complete course of motor neurons starting from the motor cortex (brain) to the spinal
cord[26] ___________________________________________________________________________ 11
Figure 2.4 The two different pathways of the upper motor neuron [26] __________________________ 11
Figure 2.5 Animal models of SCI [38] ____________________________________________________ 13
Figure 2.6 Motor unit [41] _____________________________________________________________ 15
Figure 2.7 Schematic representation of the anatomical and the physiological model of EMG signal (adopted
from [46]) __________________________________________________________________________ 15
Figure 2.8 Different assessment method of animal model of spinal cord injury [38] ________________ 19
Figure 2.9 EMG raw recording with ECG spikes [65] _______________________________________ 22
Figure 2.10 EMG raw recording contaminated by power noise [65] ____________________________ 23
Figure 2.11 Illustration of EMG signal decomposition [41], [63], [64] ____________________________ 27
Figure 2.12 Classification process: training and testing steps _________________________________ 32
Figure 3.1The main components of the utilized experimental setting [5] _________________________ 41
Figure 3.2 Graphical representation of the collected EMG signal [160] __________________________ 42
Figure 3.3 The CT image of the inflated balloon in the epidural space of the thoracic spine. The balloon is
inflated with air, which appears black in the spinal column (red arrow). Note that 60% of the column is
occupied by the balloon, and the spinal cord is displaced. This image is from a cadaveric subject [6] __ 44
Figure 3.4 Standard radiograph of the catheter inserted into the epidural space via laminotomy in a
cadaveric subject. The balloon is not inflated in this image [6] _________________________________ 44
Figure 3.5 A general block diagram for the EMG analysis - Part 1 _____________________________ 48
Figure 3.6 A general block diagram for the EMG analysis- Part 2 ______________________________ 49
Figure 3.7 Flowchart of the peak number algorithm ________________________________________ 57
Figure 3.8 An illustration of the peak number method _______________________________________ 58
Figure 3.9 Mother wavelet function (db4) _________________________________________________ 66
IX
Figure 3.10 Wavelet analysis; decomposition and reconstruction steps [160] _____________________ 67
Figure 3.11 Collected Data Structure: The data in this work was collected from the left (L) and the right (R)
side of the tail for five subjects throughout the days of the experiment (d1, d2, …, dn) for the pre- and post-
lesion period. The data for each day was decomposed into seven frequency sub-bands (D1, D2, D3, …,
A6). Three of the subjects received a treatment combining TRH, selenium, and vitamin E post-lesion
(treatment group); the remaining two subjects did not receive any treatment post-lesion (control group)
[160] _____________________________________________________________________________ 70
Figure 3.12 General schematic illustrating the SVM classification technique _____________________ 74
Figure 3.13 A graphical illustration of the K-nearest neighbor classification technique ______________ 75
Figure 3.14 Convolutional neural network structures utilized to build the EMG-based SCI classification
system ____________________________________________________________________________ 80
Figure 4.1 The aggregate fitted Q-values, representing tail movement, are plotted for the left (top) and the
right (bottom) side. Values are normalized to a range between 0 and 1, with higher values indicating greater
muscle activity. Of note, on both the left and the right side, the Q values decreased immediately after the
creation of the lesion. On the left side, the effect of the lesion in the treatment group (orange) was
attenuated, compared with that of the nontreatment group (grey). On both the left and the right side, the Q-
metric improved over time [6] __________________________________________________________ 86
Figure 4.2 Example of valid and invalid MUPs, their MUP templates, and firing patterns [43] ________ 89
Figure 4.3 From (1) to (6), different samples of the detected and clustered MUPs using the K-means
clustering method and fifteen various EMG features ________________________________________ 92
Figure 4.4 The estimated mean of the RP of seven reconstructed EMG sub-bands prior to the creation of
SCI for the left and the right side of the tail. Of note for each band, there is a difference in the RP value
when compared to the left and the right side of the tail. The D2 sub-band of the left and the right side has
the maximum RP, and the significant difference between the two sides is at the D1, D2, D4, and D5 sub-
bands. The star indicates a significant difference [160] ______________________________________ 93
Figure 4.5 The estimated mean of the RP of seven reconstructed EMG sub-bands prior and post to the
creation of the SCI for the left side of the tail. Of note, the RP values for the frequency sub-bands (D4, D6,
X
and A6) are significantly higher in the post-lesion period. The star indicates a significant difference [160]
_________________________________________________________________________________ 96
Figure 4.6 The estimated mean of the RP of seven reconstructed EMG sub-bands prior to and post to the
creation of the SCI for the right side of the tail. Of note, the RP values for the lower frequency sub-bands
(D4, D6, and A6) are significantly higher in the post-lesion period, while the higher frequency sub-bands
(D1 and D2) are significantly lower in the post-lesion period. The star indicates a significant difference [160]
_________________________________________________________________________________ 97
Figure 4.7 The estimated mean of the RP of seven reconstructed EMG sub-bands post-lesion for the left
side of the treatment (Tr) and the control (Ctrl) groups; of note, the RP values for the frequency sub-bands
(D1, D2, D3, and D6) are significantly different. Subjectively, the distribution of the RP in the treatment
group is similar to the RP distribution in the pre-lesion period for all the subjects. The star indicates a
significant difference [160] ____________________________________________________________ 99
Figure 4.8 The estimated mean of the RP of seven reconstructed EMG sub-bands post-lesion for the right
side of the treatment (Tr) and the control (Ctrl) group; of note, the RP values for frequency sub-bands (D1,
D3, D4, D5, D6, and A6) are significantly different. Subjectively, the distribution of the RP in the treatment
group is similar to the RP distribution in the pre-lesion period for all the subjects. The star indicates a
significant difference [160] ___________________________________________________________ 100
Figure 4.9 The basic components of a confusion matrix [205], [206] ___________________________ 103
Figure 4.10 An example of the confusion matrix results _____________________________________ 104
Figure 4.11 Confusion matrix of the SVM classifier using EMG amplitude features / left side ________ 106
Figure 4.12 Confusion matrix of the KNN classifier using EMG amplitude features / left side ________ 107
Figure 4.13 Evaluation metrics of the KNN and the SVM classifier using EMG amplitude features/ left side
________________________________________________________________________________ 108
Figure 4.14 Confusion matrix of the SVM classifier using EMG amplitude features / right side ______ 109
Figure 4.15 Confusion matrix of the KNN classifier using EMG amplitude features / right side _______ 110
Figure 4.16 Evaluation metrics of KNN and SVM classifier using EMG amplitude features/ right side _ 110
Figure 4.17 Confusion matrix of the KNN classifier using the Q-metric and the RP features of the left side
EMG data ________________________________________________________________________ 112
XI
Figure 4.18 Confusion matrix of the SVM classifier using the Q-metric and the RP features of the left side
EMG data ________________________________________________________________________ 112
Figure 4.19 Confusion matrix of the KNN classifier using the Q-metric and the RP features of the right side
EMG data ________________________________________________________________________ 113
Figure 4.20 Confusion matrix of the SVM classifier using the Q-metric and the RP features of the right side
EMG data ________________________________________________________________________ 113
Figure 4.21 Evaluation metrics of the KNN and the SVM classifier using the RPs and the peak number
features/ left side ___________________________________________________________________ 114
Figure 4.22 Evaluation metrics of the KNN and the SVM classifier using the RPs and the peak number
features/ right side __________________________________________________________________ 115
Figure 4.23 The accuracy curve for the EMG data classification of the left side using the proposed CNN
architecture _______________________________________________________________________ 117
Figure 4.24 The loss curve for the EMG data classification of the left side using the proposed CNN
architecture _______________________________________________________________________ 118
Figure 4.25 The accuracy curve for the EMG data classification of the right side using the proposed CNN
architecture _______________________________________________________________________ 118
Figure 4.26 The loss curve for the EMG data classification of the right side using the proposed CNN
architecture _______________________________________________________________________ 119
Figure 4.27 Evaluation metrics of the CNN and the KNN classifier / left side ____________________ 120
Figure 4.28 Evaluation metrics of the CNN and the KNN classifier / right side ___________________ 120
XII
List of Tables
Table 2.1 A simple comparison between surface and indwelling electrodes [47] _____________________ 17
Table 2.2 Summary of the most common EMG filtering methods __________________________________ 21
Table 2.3 Most common features of EMG signals [48], [68], [69], [71], [84]–[88] _____________________ 25
Table 2.4 The most common applications of WT analysis in EMG signal field _______________________ 30
Table 2.5 Previous works related to EMG signal and deep learning techniques ______________________ 37
Table 3.1 The frequency ranges of the seven EMG reconstructed sub-bands _______________________ 66
Table 3.2 Parameters used in the equation of the mixed model ____________________________________ 69
Table 3.3 Optimized CNN architecture __________________________________________________________ 78
Table 4.1 The results of the statistical model (lesion effect) of the conventional EMG features for the left
side ________________________________________________________________________________________ 84
Table 4.2 The results of the statistical model (treatment effect) of the conventional EMG features for the
left side _____________________________________________________________________________________ 84
Table 4.3 The results of the statistical model (lesion effect) of the conventional EMG features for the right
side ________________________________________________________________________________________ 84
Table 4.4 The results of the statistical model (treatment effect) of the conventional EMG features for the
right side ____________________________________________________________________________________ 84
Table 4.5 Standardized Q-metric_______________________________________________________________ 87
Table 4.6 The utilized hyperparameters________________________________________________________ 116
XIII
List of Abbreviations
Abbreviation Definition
A6 Reconstructed EMG of Approximate Coefficient/Level 6
ALS Amyotrophic Lateral Sclerosis
AR Auto-regressive
BC Binary Cross-Entropy
Bior Biorthogonal
BPF Band Pass Filter
cA6 Wavelet Approximate Coefficient / Level 6
cD1 Wavelet Detail Coefficient / Level 1
cD2 Wavelet Detail Coefficient / Level 2
cD3 Wavelet Detail Coefficient / Level 3
cD4 Wavelet Detail Coefficient / Level 4
cD5 Wavelet Detail Coefficient / Level 5
cD6 Wavelet Detail Coefficient / Level 6
CNN Convolutional Neural Network
Coif Coiflets
Colab Collaboratory
CWT Continuous Wavelet Transform
D1 Reconstructed EMG of Detail Coefficient/Level 1
D2 Reconstructed EMG of Detail Coefficient/Level 2
D3 Reconstructed EMG of Detail Coefficient/Level 3
D4 Reconstructed EMG of Detail Coefficient/Level 4
D5 Reconstructed EMG of Detail Coefficient/Level 5
D6 Reconstructed EMG of Detail Coefficient/Level 6
Db Daubeches
DL Deep Learning
DWT Discrete Wavelet Transform
ECG Electrocardiography
EMG Electromyography
FC Fully Connected
FD Frequency Domain
XIV
FN False Negative
FP False Positive
FT Fourier Transform
GAP Global Average Pooling
HF High Frequency
IP Interference Pattern
KNN K-Nearest Neighbors
LDA Linear Discriminant Analysis
LF Low Frequency
LMN Lower Motor Neuron
LPF Low Pass Filter
MAV Maximum Absolute Value
MFP Muscle Fiber Potential
ML Machine Learning
MLP Multilayer Perceptron
MN Motor Neuron
Morl Morlet
MSPCA Multiscale Principal Component Analysis
MSR Muscle Stretch Reflex
MU Motor Unit
MUD Motor Unit Disease
MUNE Motor Unit Number Estimation
MUP Motor Unit Potential
MUPT Motor Unit Potential Train
MUSIC Multiple Single Classifications
NE Needle Electrode
NHP Non-Human Primate
P2P Peak to Peak
PUK Pearson VII Function-Based Universal Kernel
Q The Suggested Metric of Volitional Limb Movement
RBF Radial Basis Function
ReLUs Rectified Linear Units
RMS Root Mean Square
XV
RNN Recurrent Neural Network
RP Relative Power
SBP Sub-Band Power
SCI Spinal Cord Injury
SD Standard Deviation
SMOTE Synthetic Minority Oversampling Technique
SVM Support Vector Machine
Sym Symlets
TD Time Domain
TFD Time Frequency Domain
TN True Negative
TP Total Power
TP True Positive
TRH Thyrotropin Releasing Hormone
UMN Upper Motor Neuron
WT Wavelet Transforms
ZC Zero Crossing
1
Chapter 1
Introduction
1.1. Spinal Cord Injury
Spinal Cord Injury (SCI) is a serious neurological condition. Worldwide, there are
approximately 930,000 new cases each year [1]–[3]. The most common causes of SCI
are motor vehicle collisions, falls, sports-related activities, and interpersonal violence [4],
[5]. SCI typically damages the motor, sensory and autonomic nerve fiber tracts. Patients
experience a spectrum of clinical abnormalities such as limb paralysis and dysesthesia,
as well as bowel and bladder dysfunction [5]–[7].
SCI has a devastating effect on the physical, psychosocial, and emotional aspects of
patients and caregivers. The burden of this condition is shared by the person affected,
family members, the community, as well as the healthcare system. Long-term survival
and quality of life have improved due to enhanced rehabilitation and medical
interventions; however, it has not been convincingly demonstrated that current
pharmacological treatments have substantially improved spinal cord function. One of the
main challenges for developing a suitable pharmacologic intervention is the absence of a
valid non-human model for SCI [7], [8].
1.2. Problem Formulation
Despite the vast budget allocated for the development of SCI medication, the rate of
successful clinical trials is still low [8], [9]. The main reason behind this low rate is the lack
of a valid non-human primate (NHP) model of SCI and a reliable assessment tool [8], [9].
2
Developing such a model would therefore have a significant impact on filling the
translational gap between preclinical and clinical research. As well, it would be an
essential step in the broader plan for the drug development industry, as NHP data from a
proper and accurate design gives a clear idea about possible clinical outcomes. Due to
the importance of the NHP models in translational medicine, investors and the
government health institutions offer great financial support for developing and optimizing
this significant stage in the development of healthcare treatment [9].
1.3. Research Objectives
Conceptually, the overriding goal was to build an NHP model of SCI by characterizing
the injury neurophysiologically using intramuscular electromyography (EMG) signals.
This contribution would help in building a model to test new therapies for SCI and thereby
assist in improving the functionality and quality of life for human beings. In order to
achieve this general objective, three specific objectives were set to be addressed in this
study. These objectives included the following:
1. To investigate a method (metric) by which general volitional muscle activity pre-
and post-SCI could be measured utilizing the amplitude characteristics of
intramuscular EMG signals.
2. To characterize the discharge properties of muscle motor units by analyzing
the frequency content of the intramuscular EMG signal during pre- and post-
SCI for a treatment and a control group, and for an agonist-antagonist pair of
muscles.
3
3. To design a classification system for SCI using machine learning and deep
learning techniques.
1.3. Research Challenges
Challenges associated with developing a valid NHP model are:
• Minimizing the number of NHP involved in biomedical experiments due to strict
ethical standards and considerations, especially if the experiment requires any
physiological changes to the animal, such as causing an impairment [9].
• Balancing the limited animal numbers and the statistical power necessary for
rejecting the null hypothesis.
• Collecting and processing the large amount of data needed in order to capture the
full picture of the disease effects, which is costly and time-consuming.
1.4. Non-human Primate Model
Conceptually, the understanding of SCI impairments and recovery can be advanced
through animal models [6], [7], [10]–[13]. NHP models are particularly valuable due to the
anatomical and physiological similarities between NHP and human beings [10], [11].
Millions of human lives have been saved and many medical conditions have been cured
by the critical health advances that have resulted from research with NHPs [14]. NHP
models have been linked to many health advances, such as the development of the polio
vaccine, the development of the first animal model of Parkinson’s disease, advanced
treatment for prostate cancer, and the treatment of many other diseases [14].
4
1.5. Electromyography Signal
EMG is a widely available, cost-effective technique concerned with the recording and
the analysis of the myoelectric signal. An EMG signal is the electrical potential that is
formed when the muscle is electrically or neurologically activated. This technique has
been used in numerous clinical practices, e.g. to study the health condition of a muscle
and its innervated neurons (motor neurons). As well, it has been used as a diagnostic tool
for neuromuscular diseases, and as an assessment tool (biofeedback) for muscular
abnormalities and rehabilitation [15], [16]. EMG techniques are mainly utilized in medical
research (e.g. functional neurology, gait and posture analysis, motor control applications),
rehabilitation (e.g. neurological rehabilitation and post-surgery/accident), and sport
science (biomechanics and movement analysis) [15], [16].
The EMG signal provides information about the activity (electrical signal) of the
corresponding muscle, in which the time of the signal represents the duration of the
muscle activity. Moreover, the EMG amplitude is the level of the detected electrical signal,
which reflects the intensity of the muscle activation. As well, the frequency analysis of the
EMG signal is used to identify some neuromuscular disorders and symptoms, such as
muscle fatigue and dystonia [17]–[19].
1.6. Research Approaches
The collection and analysis of EMG data for this study are unique. Specifically, this
dataset contains EMG data from the tail musculature of NHPs. In addition, the data were
collected longitudinally for two phases: pre- and post-creation of an experimental SCI.
5
Each recording date served as a unique data point. This feature facilitated inferences
related to the trajectory of recovery from SCI. Furthermore, the muscles under
consideration were an agonist-antagonist pair. The statistical power in these experiments
was enhanced by the collection of the pre- and post-lesion EMG data. This inclusion
permitted the normalization of the post-lesion data for each subject, and theoretically,
decreased statistical variances. Consequently, this experimental strategy reduced the
number of required subjects. In this context, no customary or established analysis
paradigms were available.
The methodology of this study consisted of four main sections. The first section was
the collecting and preprocessing steps for the EMG data (recording, filtering, removing
outliers, segmentation). The second section was the characterization of the EMG signal
amplitude using the most common metrics (area, root mean square, turns, zero-crossing).
It also used a newly proposed metric, named the Q-metric, for free volitional muscle
activity. Next, an analysis of the discharge properties (frequency content) for the acquired
EMG signals was performed utilizing wavelet transforms and relative power for each EMG
sub-band. Finally, the SCI classification method was investigated using various EMG
extracted features and by employing different classification methods, including machine
learning (ML) and deep learning (DL) techniques.
1.7. Impact of the Work
Developing a new NHP model of SCI during pre- and post-lesion will have the
following impacts:
6
1. It provides a better understanding and characterization for the SCI effect on muscle
volitional activity by using an affordable and applicable neurophysiological
technique (EMG data).
2. It provides a new preliminary model to test the efficacy of suggested treatments
for SCI.
3. It may help in developing a new assessment tool that can help medical
professionals to determine the severity and progression of SCI.
4. It introduces a new application of deep learning for SCI classification using an
intramuscular EMG signal.
In addition, creating such a model and a metric will help to improve the approaches
used by the healthcare system to characterize spinal cord injury. This will occur in
different aspects such as developing new specialized rehabilitation techniques and
providing an easy assessment approach for SCI patients.
1.8. The Structure of the Thesis
The structure of the thesis is as follows: Chapter 2 provides a detailed literature
review of the main components of this work. Chapter 3 describes the proposed
methodology that was applied to the collection of the EMG signals. Chapter 4 presents
and discusses the results of the research work, and finally, the conclusions and
suggestions for future work are provided in Chapter 5.
7
Chapter 2
Background and Related Work
2.1. Introduction
This chapter provides a detailed background of the main research topics presented
in this thesis. SCI is also defined; its levels, the leading causes, and the consequences
are presented. The next section explains the NHP model and discusses its importance.
The definition and the physiological background of the EMG signal are also presented in
this chapter. Later in the chapter, an overview is provided of the EMG signal properties,
and the common artifacts contaminating this type of signal. The most common EMG
signal conditioning techniques are presented in this chapter. The SCI assessment tools
in the literature are reviewed in this chapter. The current understanding of the EMG signal
abnormalities associated with SCI in animals and human beings are presented. Chapter
2 also provides specifics regarding state-of-the-art EMG feature-extraction methods. The
applied wavelet analysis is explained as well as its relevant applications. This is presented
in the form of a table within this chapter. Finally, this chapter provides a background for
the employment of machine learning and deep learning techniques, as well as their main
applications in the EMG field.
2.2. Spinal Cord Injury
SCI is damage occurring to any part of the spinal cord. It often results in permanent
perturbations in sensation, strength and the function of the body below the level of the
injury. Typically, SCI is the result of motor vehicle accidents, falls, and violence [4], [5], as
shown in Figure (2.1).
8
Figure 2.1 The main causes of SCI [4]
People with SCI have a range of clinical impairments, which can be attributed to
impaired fiber tracts. Based on the location of the SCI along the spinal cord and the
severity of the injury, the ability of limb control and motion is determined. The neurological
level of the SCI is below the level of the injury and the severity depends on whether the
injury is complete or incomplete, as follows:
• Complete SCI: when all sensory and all motor functions are lost below the
neurological level of the injury.
• Incomplete SCI: when there is some residual sensory or motor function below the
neurological level of the injury. There are different degrees of incompleteness.
The paralysis resulting from an SCI may be referred to as [20]:
9
• Tetraplegia (Quadriplegia): The paralysis of arms, hands, trunk, legs and pelvic
organs.
• Paraplegia: The paralysis of all or part of the trunk, legs and pelvic organs. Figure
(2.2) illustrates the different types of paralysis resulting from SCI.
Figure 2.2 The different types of paralysis resulted from SCI [19]
2.3. Motor Neuron Disorder
Motor neurons (MN) are a specialized type of neuron located within the spinal cord
and the brain. There are two main subtypes, namely the upper motor neuron (UMN) and
the lower motor neuron (LMN), as shown in Figure (2.3). The UMNs originate from the
10
primary motor cortex in the brain and travel downward. UMNs synapse on LMNs in the
brain stem and form a corticobulbar tract or they connect to the LMNs in the spinal cord
to form a corticospinal tract [21]. Figure (2.4) illustrates the different pathways of UMN
[22], [23]. LMNs originate in the brainstem and the spinal cord and they innervate skeletal
muscles directly [24].
Motor neuron disorders (MNDs) are sporadic, fatal neurologic diseases characterized
by progressive degeneration of the motor neurons. These disorders can affect either or
both types of motor neurons (UMNs, LMNs). The MNDs result from different types of
diseases and lesions such as SCI, stroke, cerebral palsy, and amyotrophic lateral
sclerosis (ALS). The main clinical features of an upper motor neuron disorder (the case
in this study) include:
1. Spasticity: Stiffness or tightness of a muscle, which affects its normal movement
[23].
2. Hyperreflexia: High muscle stretch reflex (MSR), which results from losing the
periodicity of the UMN stimulation on the LMN [22].
3. Clonus: Rhythmic contraction of the antagonist muscles, such as any rapid
movement (pushing or pulling) to any one of antagonist muscles (upper or lower
limb). This leads to an involuntary movement of the limb because each muscle
stimulates the other one [22].
4. Hypertonia: Increased muscle tension and difficulty in muscle stretch ability [22].
5. Extensor Plantar Response/Babinski Sign: “occurs after the sole of the foot has
been firmly stroked, the big toe then moves upward or toward the top surface of
the foot while other toes fan out.” [25].
11
Figure 2.3 The complete course of motor neurons starting from the motor cortex (brain) to the spinal cord[26]
Figure 2.4 The two different pathways of the upper motor neuron [26]
12
2.4. Nonhuman Primate Model
Animal models are essential in biomedical research. In the field of SCI, various
species have been used, including mice, rats, cats, dogs, rabbits, and NHP, as shown in
Figure (2.5). In these models, the experimental spinal cord injury has been created
through diverse methods, some of which require surgical exposure of the spinal cord and
lesion induction by using a sharp instrument [12], [27], [28] or a device that provides static
or dynamic compression [29]–[31]. In other models, the dura remains intact, and
compression is applied by using a circumferential clip [32]. These approaches are compli-
mentary. Some pathophysiologic and clinical features of SCI differ substantially between
humans and other species; for example, despite complete anatomic transaction, dogs
and cats—unlike humans and NHP—may be able to walk [33]–[36]. The tail in NHPs is
analogous to a human limb. The tail is integral to performing functional tasks such as
standing on the hindlimbs as well as reciprocal movements that aid in balance during
ambulation [37]. As described later, the tail exhibits other features analogous to human
beings such as limb dominance or preference. In this context, NHP models have an
essential role in advancing the understanding of SCI as well as identifying candidate
treatments.
13
Figure 2.5 Animal models of SCI [38]
2.5. Electromyography Signal
An EMG represents the electrical activity of skeletal muscles during the activation
state (contraction), which is generated as a result of physiological changes in the muscle
fibers while they are moving [39]. Scientists have known about the existence of
myoelectrical signals since the 1600s, but the lack of equipment required for detecting
and measuring the electrical signals has limited the research in this field [19], [40];
however, in the early 1900s, the development of electrical components enabled scientists
to invent suitable tools for detecting the EMG signal [41], [42]. During that time, physicians
were analyzing the EMG signals qualitatively, again as a result of the lack of appropriate
electronic equipment. However, around the 1960s, this problem was solved, and the EMG
amplitudes started to be quantified [41], [42]. Step by step, EMG devices and signal
processing techniques have been developed, and this progress has enabled physicians
to make a more effortless quantitative analysis of the EMG signal. Now the EMG signal
14
is considered an important tool in understanding the neuromuscular system, in addition
to its role as a biological signal in discovering underlying disease mechanisms [41], [42].
The EMG signal is a series of voltages that are generated from muscle fibers
during muscle contractions [39]. Skeletal muscle is composed of densely packed muscle
fibers, and these fibers are innervated in groups by motor neurons [41]. The combination
of a group of muscle fibers and the single motor neuron that innervates them is known as
a motor unit (MU), as is seen in Figure (2.6) [41]. Each muscle consists of multiple MUs
which coactivate together to generate a coordinated muscle contraction. Each active
single muscle fiber generates a wave, which results from the depolarization and
repolarization of the muscle fiber membrane. This wave is called the muscle fiber potential
(MFP). The summation of all the MFPs of a single MU is called the motor unit potential
(MUP). To increase or maintain the generated force of a muscle, the MUs fire repeatedly
and generate a motor unit potential train (MUPT). A detected EMG signal is comprised of
the superposition of the generated MUPTs of all fired MUs and any accompanying
interference signal, such as the background noise [Figure (2.7)] [43]–[45].
The EMG signal is an important biological signal, which is used in different fields
such as medical research, rehabilitation, and sport science. EMG signal analysis enables
researchers to look directly into the muscle to reflect on the internal structure and activity.
As well, muscle performance could be evaluated by utilizing the EMG signal. In addition
to the previous benefits, the EMG plays an important role in choosing the desired muscle
training program and a suitable treatment [16].
15
Figure 2.6 Motor unit [41]
Figure 2.7 Schematic representation of the anatomical and the physiological model of EMG signal (adopted from [46])
16
2.6 EMG Signal Acquisition
In order to get an accurate EMG analysis, it is vital to consider the parameters that
affect the quality of an acquired signal. Electrode type, position, level of muscle
contraction, and many other parameters play an important role in the outcomes of the
EMG analysis. As well, the different kinds of noise combining in the EMG signal could be
reduced by using a proper electrode type and a suitable setting [41].
There are a variety of methods used to obtain EMG data from humans and
animals. As a general construct, there are two types of electrodes: a surface, and an
indwelling electrode (needle or wire) [47]. The neurophysiological characteristics and the
interpretation of the data derived are not synonymous. Surface electrodes are placed over
the skin; they can be manufactured from a variety of materials and formed into various
sizes and shapes. Surface recording over the skin is subject to a distortion in the EMG
signal due to the interposition of the fascia and the subcutaneous tissues. With limb
movement, the skin, the subcutaneous tissue and the muscles move asymmetrically. This
contributes further to movement artifacts. These tissues are effectively a low pass filter.
In comparison, needle electrodes are made of metal; subtypes include monopolar,
concentric, single fiber and macroelectrodes. Wire electrodes, in contrast, can vary in
diameter and length. Wire electrodes allow for long term implantation in subjects and are
generally well tolerated. When compared to surface electrodes, there is less cross-talk
[16], [41]. As a conceptual argument, data from wire electrodes aggregates the EMG
signal from muscle fibers in multiple motor units. The signal is potentially affected by
multiple factors, including the proximity of the signal generator to the wire, the phase
17
cancellation, the size of the fibers, and the type of fibers (i.e., fast-twitch or slow-twitch).
As well, there is some factors related to the neural network that controls the agonist-
antagonist pair [41]. Table (2.1) explains the main differences between the two kinds. As
well, the difference in the quality of the acquired EMG signal using these two types can
be noticed in the table, as the signal of the indwelling type is more precise and cleaner
(small area, and less cross-talk) [47].
Table 2.1 A simple comparison between surface and indwelling electrodes [47]
Surface Electrode Indwelling Electrode
Electrode Type Flat disk Needle, fine wire
Position Placed over selected muscle skin Inserted into selected muscle
Pick Up Area Large Small
Cross-talk High rate Low rate
Type of Muscle Superficial muscle Deep muscle
2.7. SCI Assessment Tools
SCI leads to variable functional consequences; therefore, an early and accurate
assessment is vital in determining the neurological deficits, defining the required therapy,
and predicting the prognosis of the injury. An essential step of the muscle functional
neurophysiological assessment entails the acquisition, as well as a qualitative and a
quantitative analysis, of the EMG signals.
The qualitative EMG analysis includes behavioral, visual, and auditory
assessments. In contrast, the quantitative analysis, which is a more accurate and
objective tool, involves neurological measures and biomarkers such as the EMG signal
[48], [49]. Anderson et al. [50] submitted a comparison study of various assessment
18
methods for measuring the effects of cervical hemisection lesions in rats. The results
revealed that with an incomplete cervical hemisection, there was a recovery stage,
whereas a complete hemisection did not demonstrate a recovery stage. The assessment
techniques used to determine recovery in this study included behavioral testing, grip
strength meter (GSM), food pellet reaching tasks, horizontal rope walking, and a dynamic
plantar aesthesiometer. GSM was the most convenient method for assessing the strength
of the muscle. In [12], the researchers utilized rhesus monkeys to build a non-human
model that described natural recovery after a spinal cord injury. This study also
highlighted the vital role of such primates in translational medical research. Then in [51],
Nout et al. explored the use of a behavioral test as an assessment approach for motor
function recovery after impairment. This study examined the sensitivity of the test to the
changes in the motor functionality level. It was revealed that behavioral measurements
represent a reliable method to assess motor function after spinal cord hemisection in
primates. Moreover, such assessment methods play a vital role in testing the efficacy of
suggested therapeutics and help in understanding their positive and negative effects on
the primate body. The assessment methods in these studies mainly depended on
behavioral tests and reaching tasks.
As mentioned earlier, the impairments of the SCI are usually measured by
observing behavior as well as histopathological assessment, as seen in Figure (2.8) [38].
However, the EMG signal confers certain advantages over these other methods. It is
particularly useful for assessing SCI for EMG data collected on a serial basis as this
approach facilitates the characterization of MU activation and recruitment [12]. EMG
signals can be recorded through multiple channels, permitting the simultaneous
19
assessment of multiple muscle groups. This is particularly helpful in evaluating agonist-
antagonist muscle pairs, in which the evaluation of muscle co-contractions is associated
with central nervous system disorders [6], [13]. EMG data can be obtained through
surface recordings or intramuscular electrodes placed into the muscles. There is limited
literature regarding the nature of intramuscular EMG signals after SCI in animals and
humans. The majority of human studies have been based on surface EMG electrodes.
As a general construct, the SCI results in a perturbation in the EMG signal. The next
section reviews the current understanding of EMG abnormalities associated with SCI in
animals and human beings.
Figure 2.8 Different assessment method of animal model of spinal cord injury [38]
Calancie et al. [52] studied the recovery of volitional activity after acute SCI in
human beings utilizing pairs of surface EMG electrodes. This study demonstrated
perturbations in some EMG characteristics including abnormalities in recruitment. Lewko
[53] utilized surface EMG electrodes and noted disturbed behavior in spinal cord
20
conductivity with quiet standing. Nout and Rosenzweig, among others [12], [13]
developed an NHP model of SCI, and the results demonstrated a significant difference in
EMG amplitude and temporal patterns between the healthy and the SCI subjects. As well,
they noted uncoordinated muscle activity during the post-lesion condition. Wiegner et al.
and Shahani et al. studied the recruitment pattern of MUs post-SCI [54] and cortical
pathology [55] using needle EMG signals. The results indicated that MUs fire irregularly
with a low discharge rate post-SCI. As well, the inter-discharge intervals (IDIs) had a
positive serial correlation which resulted from the decreased variability in the length of the
adjacent intervals. Capogrosso et al. [27] developed a brain-spine interface to modulate
the consequences of SCI in NHP. EMG signals during continuous locomotion were
recorded and averaged to calculate the spatiotemporal maps of the motoneuron
activation in monkeys. The results suggested a practical translational pathway for
conceptual analysis studies and investigational applications in humans with SCI.
2.8. EMG Signal Conditioning
Signal processing is a wide field that deals with different scientific aspects, and it
is an essential tool for analyzing any bio-signal and exploring the underlying information.
In the literature, different signal processing techniques were applied to the raw EMG
signal to enhance its quality. In this section, the most common filtering methods are
explored. Table (2.2) summarizes the common methods in the literature for EMG signal
filtering.
21
Table 2.2 Summary of the most common EMG filtering methods
Study Filter Type Frequency Limits
Hz
Sampling Frequency
Hz
[56] Analog/ Band Pass filter (BPF) 5-300 1024
[57] Analog/ BPF 10-500 2000
[58] Digital/ BPF/4th Butterworth 20-500 2000
[59] Analog/ BPF 20-1000 2000
[60] Digital & Analog / BPF /4th Butterworth 100-3000 20000
[61] Analog / BPF & Digital/ Low Pass Filter (LPF) 5hz-10khz & 8khz 20000
[62] Analog/ BPF 20 – 460 Hz 1000
[63] Digital/ BPF / Butterworth 20-500 Hz 1000
[64] Analog/ LPF 450 Hz 1000
An EMG filtering step has been implemented to attenuate the noise signals that
combine in the raw EMG signal on its way from the muscle to the electrode. There are
various types of noise that can affect the EMG shape and characteristics. These types
can be grouped as below:
1. Motion artifact (baseline shifts): This kind of artifact results from changes in the
geometry between the electrode location and the muscle or as a result of a repeated
cable shake. This problem can be solved adequately through the use of proper
electrode fixation and good skin preparation [41], [65]–[67].
2. Muscle cross-talk (physiological cross-talk): The neighboring muscles can generate
some additive EMG signals that are added to the desired raw EMG signal. This type
of noise can be solved by narrowing the measured area or changing the electrode
type; for example, using an intramuscular electrode instead of a surface electrode. An
electrocardiography (ECG) artifact, which is the electrical signal of the heart, is an
22
avoidable biological cross-talk noise that contaminates the EMG signal with ECG
bursts, as shown in Figure (2.9) [65]. This problem occurs when the muscle being
studied is located at the left upper trunk / near the heart. Many signal processing
procedures can be applied to solve this problem without destroying the original EMG
signal [41].
Figure 2.9 EMG raw recording with ECG spikes [65]
3. Power-line noise (interference power line): Any devices or equipment with poor
grounding that are close to the recording area can generate an interference power line
artifact, which is usually between 50 and 60 Hz. To overcome this type of noise, it is
important to ground all devices, especially those equipped with an EMG recording
system, such as training devices, treadmills, etc. It is also important to use a minimum
number of power plug connectors. Figure (2.10) [65] represents an EMG signal
contaminated with a power-line artifact [41], [65], [66].
23
Figure 2.10 EMG raw recording contaminated by power noise [65]
4. Baseline offset: This type of noise has a constant value and it results in shifting the
zero baseline of the original signal to a specific value. This kind of artifact may occur
if any change has been done after setting the calibration. Fortunately, it is easy to
remove the baseline offset noise by utilizing one of the designed offset correction
methods, such as “detrend” [41], [65], [66].
2.9. EMG Feature Extraction
Feature extraction refers to the transformation of raw data into a relevant and more
representative structure, which is known as a feature set. In order to find an informative
feature set, noise or irrelevant data needs to be discarded while highlighting the useful
information [68]. An EMG signal is one of the informative electrophysiological signals;
however, useful and important information is hidden within the complex structure of the
signal. This complexity makes EMG signal usage in medical and engineering applications
24
challenging [69]. During the last three decades, various EMG feature extraction methods
have been developed.
There are different types of EMG features. Generally, these features can be
organized in four basic types: time-domain (TD), frequency domain (FD), time-frequency
domain (TFD), and morphological features. These features have been considered in
many EMG applications and analyses, such as pattern recognition studies, and
neuromuscular disease diagnosis studies [70], [48], [68], [71], [72], [68], [64], [69], [73]–
[77].
TD features represent the most common and efficient computational set for EMG
features and have been adopted in most of the EMG studies. FD and TFD features are
used to understand the frequency contents of an EMG signal. They have been utilized in
different types of EMG analysis but mostly in fatigue of muscle analysis and recruitment
studies of MUs [71]. Morphological features have been used widely in the EMG
quantitative and semi-quantitative analysis for disease diagnosis as they characterize the
extracted MUPs of the active MUs in the studied area [43], [48], [78], [79]. These types of
features generally describe the morphological aspects of MUPs, such as shape, size,
global complexity, and local complexity [48], [80], [81]. This type of analysis characterizes:
1) the pathological effects on the morphology aspects of MUPs, and 2) the stability of
MUPTs (firing rate and recruitment style) of the active MUs in the studied area [48], [82],
[83]. Table (2.3) illustrates some of the features that are utilized in the literature with their
names, definitions, and some references.
25
Table 2.3 Most common features of EMG signals [48], [68], [69], [71], [84]–[88]
Feature
Type Feature Name Feature Definition
Tim
e D
om
ain
Featu
res
Area Integration of the absolute value of each sample of the EMG
signal over its duration.
Turn number (signal
slope changes)
The number of changes in the slope of a line from negative to
positive and vice-versa.
Peak to peak (P2P) The difference between the maximum and minimum peak
values for the EMG signal.
Root mean square
(RMS)
A statistical measure representing the square root of the mean
of the squares of all EMG signal samples.
Maximum absolute
value (MAV) The average of the absolute value of the EMG signal
Zero-crossing (ZC) The number of EMG signals crossing the zero line
Auto-regressive
coefficients (AR) The coefficients of the AR model
Mo
rph
olo
gic
al
Featu
res
Thickness The ratio of area to P2P
Turn-amplitude The ratio of amplitude to turns.
Turn-length The ratio of length to turns
Turn-area The ratio of area to turns
Turn-width The ratio of width to turns
Length “The summation of the absolute amplitude differences for every
two consecutive samples” [47]
Shape-width The ratio of area to length
Phase number The part of EMG that falls above and below the crossing line
Phase-area The ratio of area to phase
Fre
qu
en
cy
Do
main
Peak frequency Maximum frequency.
Mean power “Average power of the EMG power spectrum” [71]
Total power “An aggregate of the EMG power spectrum” [71]
Mean frequency The mean frequency of a power spectrum
Median frequency The median frequency of a power spectrum
Power density drop
ratio The maximum-to-minimum drop in power density ratio
2.10. EMG Decomposition
There are four main approaches to analyze the EMG signal quantitively, including
MUP analysis (decomposition), interference pattern (IP) method, motor unit number
estimation (MUNE), and muscle fiber jitter [89]. The MUP analysis approach, which is
26
tested in this study, analyzes the firing behavior of the active MUs and provides useful
information regarding the MU size and density in the studied area [89].
Each muscle in the skeletal system consists of a large number of muscle fibers,
and these fibers are innervated by α- nerves, which represent the LMN of the motor nerve
system [39]. A motor unit represents the functional unit of the motor neuron system, in
which each α- neuron axon innervating a group of muscle fibers represents one MU [90].
The summation of the action potentials from all the muscle fibers within one MU is called
the MUP, and during muscle contraction, the MU fires a train of these MUPs in order to
keep the stability of the muscle force [65]. The recorded EMG signal contains a
combination of electrical activity from several MUs; in other words, the EMG signal is a
combination of a series of MUPs from the active MUs, which is known as the MUPT [89],
[91]. The EMG signal is a superposition of multiple motor unit activations. Therefore, to
understand the hidden mechanism of muscle fibers and motor neurons, it is important to
decompose the signal and find the activity of constituent functional MUs [41], [92].
Different approaches regarding EMG decomposition have been suggested, such as
manual, semi-automatic, and automatic methods [89]. Figure (2.11) explains the general
idea of EMG decomposition.
27
Figure 2.11 Illustration of EMG signal decomposition [41], [63], [64]
The MUP analysis consists of seven main steps: 1) EMG signal acquisition, 2)
EMG signal pre-processing, 3) detection and segmentation of the MUPs, 4) feature
extraction from the MUPs, 5) clustering and supervised classification of the candidates,
6) resolving superimposed MUPs, and 7) finding the templates of the MUPs and the firing
pattern information of the motor units [43]. In the previous sections, the first two points
have been explained. The main EMG extracted features have also been reviewed earlier.
This type of analysis focuses more on the morphological perturbations of MUPs through
different aspects, such as shape, size, and complexity, which reflect how MUPs are
varying across an MUPT [89], [91].
The segmentation step aims to extract the EMG segments that could have an MUP
generated from an active MU in the measured muscle area. This step is applied using a
threshold crossing approach [93]–[102], and it results in a set of MUP candidates that will
28
be used in the clustering step. The aim of applying the clustering step before the
supervised classification is to get some prior information about the segmented MUPs and
to group them into initial sets of MUPTs that represent the activity of most active MUs.
There are various types of clustering algorithms that have been employed in the literature
regarding MUP analysis, such as single linkage [91], [100], [103]–[105], K-means [93],
[95], [106], [107], and many other approaches. The main challenges that face applying
clustering techniques are as follow: 1) there is no previous information about the data
labels, 2) the challenge of finding the optimal number of clusters, and 3) the different
results that could be obtained from a data set using different similarity measurements and
clustering techniques [43].
2.11. Wavelet Transforms
Wavelet transform (WT) is a well-known mathematical multiresolution
decomposition method, and it maps a function onto time-scale space. Wavelet transform
has time and frequency localizations, unlike the Fourier transform (FT). WT has been
developed to solve the non-stationarity problem of some signals, such as bioelectrical
signals [108]. WT decomposes a signal into a set of basis functions consisting of dilations,
contractions, and translations of a fixed function known as a mother wavelet. The wavelet
function is produced by the shifted mother wavelet through a shifting parameter, b, and
dilated/contracted by a scaling parameter, a, as in Equation (2.1) [109]:
𝜓𝑎,𝑏(𝑡) = 1
√𝑎 ψ (
𝑡 − 𝑏
𝑎) (2.1)
29
where 𝑎 ∈ 𝑅+ and 𝑏 ∈ 𝑅 , are the scaling and the translating parameters. When
the original signal f(t) is projected locally on a scaled and shifted wavelet function 𝜓𝑎,𝑏(𝑡),
the wavelet transforms of function f(t) are produced, as in Equation (2.2) [109]:
𝑊𝑎,𝑏 = ∫ 𝜓𝑎,𝑏(𝑡) 𝑓(𝑡) 𝑑𝑡
∞
−∞
(2.2)
Discrete wavelet transform (DWT) can be obtained by discretizing the value of the
translating parameter (𝑏 = 𝑛 𝑏0𝑎0𝑚) and the scaling parameter (𝑎 = 𝑎0
𝑚), as in Equation
(2.3) [109]:
𝐷𝑊𝑇𝑚,𝑛(𝑓) = 𝑎0
−𝑚2⁄
∫ 𝑓(𝑡) 𝜓(𝑎0−𝑚𝑡 − 𝑛𝑏0)𝑑𝑡 (2.3)
The WT is an efficient mathematical analysis method for temporally nonstationary
and spatially nonhomogeneous bio-signals [108]. As well, this approach has proven its
ability to represent precise measurements and to extract useful information from non-
stationary biomedical signals [110]–[113]. Decomposition, denoising, and pattern
classification are the most common applications of the WT in the EMG field [57], [69],
[99], [110], [114]–[119]. Phinyomark et al. [57] investigated the usefulness of extracting
EMG features from the multi-resolution wavelet decomposition process. The results
showed that the reconstructed EMG of the first and second level (detail coefficients) had
improved the class separability. Yamada et al. [120] introduced a new EMG
decomposition algorithm by adopting the principal component analysis of wavelet
30
coefficients. This method showed a higher decomposition accuracy when compared to
conventional wavelet methods. Fang et al. [114] decomposed EMG signals into their
constituent single motor units using wavelet spectrum matching, and the results were
satisfying. Table (2.4) summarizes the most common applications of WT analysis for
EMG signals.
Table 2.4 The most common applications of WT analysis in EMG signal field
# Study WT Application WT Type, Mother Wavelet
1 [121] Denoising DWT, Sym8,
2 [122] Denoising DWT, Db6
3 [123] Denoising Db4
4 [124] Visualization of the neuropathy and myopathy activity Db2
5 [118] Discrimination of healthy and post-partum CWT
6 [125] Assessment of muscle load and fatigue CWT, Db5
7 [126] A measure of muscle activation CWT
8 [127] The usefulness of the EMG features DWT, Db2
9 [128] Recruitment pattern of slow and fast fibers CWT, Db 4
10 [114] Multi-unit EMG signal decomposition Db
11 [117] Pattern classification DWT, Sym5, Bior6.8
12 [129] Pattern classification Sym4 and Db4
13 [110] Motor control DWT, Coif 5
14 [130] Pattern classification CWT, Morl
15 [131] Pattern classification DWT, Sym 5
16 [132] Muscle fatigue analysis DWT, Sym4 and Sym5
17 [133] EMG control system DWT, Sym4 and Sym5,
18 [69] Pattern classification DWT, Db7, Db8
19 [134] EMG classification Db6
20 [57] Pattern classification DWT, Db7
21 [135] EMG classification CWT, Coif 4
22 [136] EMG M-wave pattern recognition Db4
31
where:
• CWT: Continuous Wavelet Transform
• DWT: Discrete Wavelet Transform
• Db: Daubeches; Sym: Symlets; Morl: Morlet; Coif: Coiflets; Bior: Biorthogonal.
2.12. SCI Classification System
Classification is a supervised learning process, which tries to establish the
relationship between a set of features and a specific target of interest. It is mostly used in
machine learning and statistics. The classification problem could be a binary or a multi-
class problem. Figure (2.12) illustrates the two main steps of the learning process, which
include:
1)Training Step: In this step, using training instances, a model is constructed.
2)Testing Step: In this step, a label is assigned to the unlabeled testing instances using
the model constructed in the training step.
32
Figure 2.12 Classification process: training and testing steps
An EMG signal represents a useful biomarker to assist in understanding many
neuromuscular diseases, such as myopathy and neuropathy disorders [61], [129], [137]–
[142]. An EMG classification analysis is mainly performed for applications regarding
prosthetic motor control and neuromuscular disease assessment [61], [129], [139], [140],
[142]. Developing a reliable EMG classification system would assist in the diagnosis of
these conditions and would facilitate the periodic assessment of a patient. As well, it would
provide an easy system for testing the efficacy of new pharmaceutical therapies.
33
In this study, different EMG features were extracted using numerous methods
including TD, morphological, and WT analysis. The ability of these features to differentiate
between the pre- and post-lesion EMG data were tested using different classifications of
machine learning techniques. As well, a new application of deep learning was
investigated and compared to the previous approaches regarding the SCI classification
system.
2.12.1 Machine Learning Classification Technique
Neuromuscular diseases were identified based on EMG signals using different
machine learning techniques, such as support vector machine (SVM), multilayer
perceptron (MLP), and K-nearest neighbors (KNN). Guler and Koçer [143] proposed
classifying the EMG signals of myopathic, neuropathic, and healthy subjects using SVM
and MLP techniques. A comparison of the accuracy for these two applied classifiers
showed that the SVM had the better accuracy at 92.3% while the accuracy of the MLP
was 91.6%. Kaur et al. [144] classified the EMG data from healthy and diseased
(myopathic and neuropathic) subjects using two SVM classifiers. The first SVM classifier
was used to differentiate between the healthy and the diseased data sets. The diseased
data sets were then classified further with a second SVM into myopathy and neuropathy
signals.
Elamvazuthi et al. [145] used the MLP method to perform a classification task to
differentiate between three classes: myopathy, neuropathy, and a control group. The
results of this work verified the MLP as a useful tool to classify the three different
conditions based on some extracted EMG features. In the study by Oskoei et al. [64], an
34
SVM classifier was used to perform classification tasks on upper limb movements utilizing
EMG signals. The authors suggested a feature set, a data segmentation technique, and
a parameter adjustment approach for improving the SVM classifier. Combining all these
approaches offered a robust performance, exceptional accuracy, and a low computational
load for the SVM compared to that of other classifiers, like linear discriminant analysis
(LDA) and MLP. Artug et al. [88] used a scanning EMG method to build a new dataset;
four new EMG features were then extracted. MLP, SVM, KNN, and radial basis function
(RBF) classifiers were used to investigate the effect of these features. The best classifiers
were the SVM and the KNN algorithm as they obtained the highest classification accuracy
at 97.78%, while the MLP was the second-best classifier with an accuracy of 97.33%. In
Subasi et al. [61], the EMG signals of normal, myopathic, and neurogenic subjects were
classified by an SVM classifier. The results confirmed the SVM as a promising
classification technique. It was important to carefully set the kernel function and its
parameter values while building the SVM model because improper parameter settings
could result in misclassifications.
Gokgoz et al. [146] tested the effect of a multiscale principal component analysis
(MSPCA), a denoising method, and a multiple single classifications (MUSIC) feature
extraction method on classification performance. Using ANN, KNN, and SVM classifiers,
combined with a MUSIC decomposition method with MSPCA denoising, three EMG
signal patterns [normal, myopathic and ALS] were classified. The results showed that the
accuracy of the SVM was improved by 35.55%, the ANN by 39.87% and the KNN by
31.66%, in comparison to results in which the MSPCA denoising and the MUSIC
35
decomposition methods were not used. Bose et al. [147] used the cross-correlation based
feature extraction technique for the classification of neuromuscular diseases employing
the EMG signal to consider myopathy, neuropathy and healthy subjects. The authors
used the KNN and SVM classifiers for this work. The best results (100% accuracy) were
found when they used K=4 for the KNN and a Gaussian radial basis function for the SVM
classifier.
2.12.2 Deep Learning Classification Technique
Deep learning and convolutional neural network (CNN) have recently received increased
attention in several machine learning fields, such as computer vision and speech
recognition [148]. The CNN consists of multiple various building layers, and the three
main types of layers are a convolutional layer, fully connected layer, and output layer
[149], [150].
I. The convolution layer: It is an essential component of CNN, it includes linear and non-
linear operations (convolutional layer and activation function). The convolutional layer
comprises a specific number of filters (kernels) that are applied over the input data.
These filters work as detectors to extract different features of the input data. The filter
convolves with the input data elements (image: pixels, signal: data points) by shifting
it horizontally and vertically in a certain number of steps that is known as a stride. All
the outputs from the coevolution step will be combined in one volume that is named a
feature map. Next, a nonlinear function which is usually rectified linear unit (ReLU) will
be applied to the created feature maps. ReLU function has shown to be useful in
solving the vanishing gradient problem during the backpropagation process [151]
36
which helps in reducing the training time. Then, to reduce the size of the extracted
feature maps, pooling operations are applied using different types of operations (Max,
Sum, Average). Finally, the output of the last convolutional layer pass through a flatten
operation to transform it into one vector to be used later as input to the fully-connected
layer. The CNN is trained in a supervised manner, and the filters (kernels) are learned
automatically during the training process of the CNN. The hyperparameters related to
the kernels such as their size, their numbers, padding and stride need to be designed
before the CNN training process. The difference between the ML classification and
DL classification is the type of the utilized features. The ML classification uses hand-
crafted features that are extracted manually from the input data to perform the
classification task. While the DL classification automatically learns unique
representations of the input data (learned feature maps) that can classify the data into
their related classes.
II. The fully connected layer: It is a fully connected feed-forward neural network that
consists of neurons which are fully connected to all the neurons in the following layer
by learnable weights. It receives the output of the flattening step, which is the last
pooling step of the last convolutional layer.
III. The output layer: It includes an activation function that is selected according to the
type of the classification task (binary classification: Sigmoid function, multiple class
classification: Softmax function, regression-continues values: Linear function)
Despite the growing attention that has been given to deep learning techniques,
including CNNs, autoencoder (AE), and recurrent neural networks (RNNs) in the EMG
37
motor control area, to the best of this author’s knowledge, the EMG-based neuromuscular
disease classification system is still a new application, and has not been investigated
previously. Therefore, it seems reasonable to start investigating this type of learning style
and compare it to the conventional machine learning techniques that have been studied
extensively in the literature. CNN has been successful in various EMG motor control
studies, and it achieved a competitive result compared to other machine learning
classification techniques. Table (2.5) summarizes the most recent applications of deep
learning in the EMG motor control field.
Table 2.5 Previous works related to EMG signal and deep learning techniques
# Month / Year
Study DL Model EMG Feature Application Conclusion
1 June / 2019
[152]
Improved dual parallel channels convolutional neural network (IDPC-CNN).
Spectral features of sEMG
Fall detection using sEMG
IDPC-CNN is significantly better than other applied methods
2 May / 2019
[153] CNN Spectrogram images
Pattern recognition
CNN classification accuracy of 88.04 %
3 April / 2019
[154] Regression CNN Raw EMG signal
Simultaneous EMG control
CNN-based system outperformed SVM-based system
4 Apr / 2019
[154] Compact CNN Raw EMG signal
Gesture classification
CNN outperformed SVM
5 Jan / 2019
[155] CNN + transfer learning
Raw EMG, spectrograms and continuous wavelet transform (CWT)
Gesture classification
Transfer learning enhanced the performance for the CWT-based CNN
6 Oct / 2017
[156] CNN sEMG Gesture recognition
CNN achieved an average accuracy of 97.81%
7 Jul / 2017
[157] CNN sEMG Pattern recognition
CNN showed a better performance compared to SVM
8 Feb / 2017
[158] CNN sEMG images Hand gesture recognition
CNN is better than LDA, SVM, KNN, and RF (state-of-the-art methods)
9 Nov/ 2016
[159] CNN sEMG images Gesture recognition
CNN is better than LDA, SVM, KNN, MLP and RF
38
10 Oct / 2016
[160] CNN sEMG spectrograms
Robotic arm guidance
CNN achieved state-of- the-art results
11 Sep / 2016
[148] CNN Raw sEMG Pattern recognition
CNN produced accurate results with a simple architecture compared to the classical methods
39
Chapter 3
Materials and Methodology
3.1. Introduction
This chapter describes the work methodology, including all the details regarding the
EMG acquisition process. This includes the equipment, the experimental setting, and the
protocol. All of these details are described in the Materials and Experimental Procedures
section of this chapter. A brief description is made of the subjects which were enrolled. A
detailed description is then made of the surgical procedures that were required to implant
the intramuscular EMG electrodes and to create the designed SCI. The monitoring of the
subjects, euthanasia, and postmortem examinations are also explained at the end of this
section. The EMG Analysis Methods section discusses all the applied EMG analyzing
steps and methods that were implemented.
3.2. Materials and Experimental Procedures
3.2.1. Subjects enrolled
The subjects were healthy, research-naïve, adult male cynomolgus macaques
(Macaca fascicularis; n = 6; weight, 6 to 13 kg). The first four subjects received a
combination of pharmacologic treatment; the remaining two subjects did not receive
treatment.
3.2.2. The initial surgical procedure to facilitate the collection of EMG data
On the lower back of each macaque, a midline incision was placed superior to the
proximal tail, and a small pocket was dissected between the underlying muscle and the
subcutaneous fat. Additional small incisions were made on the left and the right side of
40
the tail to expose the flexor cauda longus and brevis muscles; these muscles are an
agonist-antagonist pair. A small telemetry device (PhysioTel D70, Data Sciences
International, Minneapolis, MN) was inserted into this pocket in the lower back. The
telemetry device was attached through a set of intramuscular electrodes, which were
implanted into the left and the right flexor cauda longus and brevis. The EMG signals
were measured using intramuscular electrodes with a 10 mm inter-electrode distance, a
10 mm exposed wire length, a 1000 Hz sampling frequency, and a common ground.
Figure (3.1) illustrates the experimental setup. The tail of Macaca fascicularis can
be considered a fifth limb, which is involved in the performance of functional tasks and
balance [37], [161], [162]. The tail has well-developed sensory and motor areas. EMG
data were collected by a radio frequency link to a computer for 30 days to determine
baseline tail movements. Data collection occurred Monday through Friday (excluding
holidays) for one hour daily. The attached intramuscular electrodes were implanted into
the left and the right flexor cauda longus and brevis muscles (tail muscles) and used to
measure EMG signals. In a NHP, the tail is integral to functional tasks; this includes
reciprocal movements that aid in balance during ambulation [37]. As such, EMG data from
the tail is analogous to EMG data obtained from the limb of a human being with SCI.
Baseline EMG data were collected during voluntary movement of the NHPs within their
home enclosures for 30 days. Intramuscular electrodes allow for long-term implantation
in subjects and are generally well tolerated. When compared to surface electrodes, there
is less cross-talk. As a conceptual argument, data from intramuscular electrodes
aggregates the EMG signal from muscle fibers in multiple motor units.
41
EMG data were collected from all six subjects. However, for one subject that
received treatment, the information from one side was not recorded. As such, its raw data
was not included in the work analysis. The resulting analyses represent the data of only
three of the subjects that received the combination therapy (Subjects 1, 2, and 3), and
two subjects who did not receive treatment (Subjects 4 and 5). The EMG data were
collected during the free movement of normal daily activities of the subjects (non-
stimulated manner). Figure (3.2) illustrates a sample of the recorded raw EMG data.
Figure 3.1The main components of the utilized experimental setting [5]
42
Figure 3.2 Graphical representation of the collected EMG signal [163]
3.2.3. The subsequent surgical procedure to create an experimental spinal cord
injury
At 30 days after the implantation of the transmitter, the subjects underwent a
second surgery. The anesthesia and the postsurgical pain management plan was the
same as described earlier. A small laminotomy was performed at the level of the fifth
lumbar vertebra. An epidural balloon catheter was inserted and advanced approximately
10 cm cranial, to the level of the lower thoracic spinal cord. The balloon was inflated
rapidly and remained inflated for 1 min. Conceptually, this procedure corresponds to
human SCI, in that there is a rapid transfer of energy (that is, initial balloon inflation),
followed by residual displacement of tissues such as disk material, boney fragments, or
hematoma (that is, continued balloon inflation for 60 s). The balloon was then deflated.
The catheter was removed, and the surgical incision was closed. The focus of this model
was to create a lesion that mimics human SCI. The SCI was designed to cause only
43
pathological and neurophysiological impairments without any clinical impairments that
could affect the function of the subject’s internal organs.
The radiographic image in Figure (3.3) highlights the location of the experimental
lesion in the lower thoracic vertebra marked by the red arrow. The CT image in Figure
(3.4) demonstrates how the lesion is created and the displacement of the thoracic spinal
cord by the epidural catheter. After the lesion was created, the subject remained
anesthetized for 1 h. This time period is a typical duration between injury and the
availability of emergency medical treatment in humans.
After one hour, four macaques (treatment group) received an intravenous bolus of
thyrotropin releasing hormone (TRH) (dose, 0.2 mg/kg) followed by a continuous
intravenous infusion for one hour of selenium (60 mg, 0.2 mg/kg/h). In addition, vitamin E
(dose, 80 IU) was administered orally once daily starting one day after surgery and
continuing for 90 d. In addition, two macaques (untreated control group) received an
infusion of normal saline for 1 h. These control subjects did not receive any selenium or
vitamin E. All the subjects experienced a spinal cord lesion.
44
Figure 3.3 The CT image of the inflated balloon in the epidural space of the thoracic spine. The balloon is inflated with air, which appears black in the spinal column (red arrow). Note that 60% of the column is occupied by the balloon, and the spinal cord is displaced. This image is from a cadaveric subject [6]
Figure 3.4 Standard radiograph of the catheter inserted into the epidural space via laminotomy in a cadaveric subject. The balloon is not inflated in this image [6]
45
3.2.4. Monitoring of subjects
The investigators and the veterinary staff closely monitored all subjects. The
frequency of assessment was at least twice daily and was more frequent during the
immediate post-surgical periods. The macaques were assessed and monitored for clinical
indicators of illness including limb weakness, vomiting, diarrhea, jaundice, bleeding, and
anorexia. The animals were weighed at the time of the surgical procedure, and prior to
euthanasia. The subjects were monitored for wasting and emaciation, as clinical
indicators of weight loss.
3.2.5. Euthanasia and post-mortem examination
On day 90, the subjects were euthanized by sedation with ketamine (10 to 20 mg/kg
IM) followed by sodium pentobarbital (>50 mg/kg IV). The postmortem examination was
performed immediately after euthanasia. The gross and microscopic examination
included assessment of the brain, spinal cord, heart, liver, spleen, kidneys, and bladder.
Venous blood was obtained immediately prior to euthanasia; cerebrospinal fluid was also
obtained. The experimental SCI resulted in both a physiological and a histopathological
perturbation [6] as shown in Figure (3.5).
46
Figure 3.5 (A) Photomicrograph of the spinal cord at the epicenter of the lesion. The gray and white parenchyma are disrupted by dilated myelin sheaths (spongiosis; arrows), which are widespread. These findings are noted caudal and cephalad to the lesion, in both treated and untreated subjects. Hematoxylin and eosin stain; magnification, 40×. (B) Photomicrograph of the spinal cord (white matter funiculi) at epicenter of experimental injury. There are many dilated myelin sheaths (spongiosis; black arrows), some of which contain degenerating or swollen axons (red arrows). These findings are noted caudal and cephalad to the lesion, in both treated and untreated subjects. Haematoxylin and eosin stain; magnification, 400× [6].
The research involving the first two subjects was completed at the New England
Primate Research Center, where the animals were cared for in accordance with the
National Research Council’s Guide for the Care and Use of Laboratory Animals (8th
47
edition) and the standards of the Harvard Medical School Standing Committee on
Animals. Research on the remaining four subjects was completed at the Wisconsin
National Primate Research Center and was approved by the IACUC at the University of
Wisconsin at Madison. The macaques were under close supervision by the veterinary
staff and were monitored for any adverse effects.
3.3. EMG Analysis Methods
This work represents the only analysis that has been applied using this dataset. The
recorded EMG signal was analyzed using a methodology that consisted of four stages:
EMG preprocessing, general muscle activity metric analysis, motor unit discharge
analysis, and development of an SCI classification system. The first stage included the
filtering, the removal of outliers, and the segmentation steps of EMG signals. The second
stage was performed using the most common EMG features (area, root mean square,
turns, and zero-crossing) as well as a newly proposed metric named Q-metric. The third
stage was performed using a conventional decomposition method, and a proposed
wavelet and relative power method. The fourth stage included developing SCI
classification systems using the different EMG features and classical ML techniques
(KNN and SVM). It also included a new deep learning application for an SCI classification
system using CNN and raw EMG signals. Figure (3.6) illustrates the first three stages with
the main steps of each analysis, while Figure (3.7) represents the fourth stage of the
methodology.
48
Figure 3.6 A general block diagram for the EMG analysis - Part 1
49
Figure 3.7 A general block diagram for the EMG analysis- Part 2
50
3.3.1. EMG signal preprocessing
The EMG signals were processed using three main steps which includes:
a) Removing outliers
EMG data were collected from all six subjects. However, in one subject, the
information from one side was not recorded due to a technical error. As such, the raw
data in this work represents the experience of three subjects that received the
combination therapy, and two subjects who did not receive treatment. As well, due to
some technical issues during the EMG recording process, EMG data from a few days
during the experiment were discarded.
b) Segmentation
The recording session from each day collected between 2 to 4 million EMG data
points, of which 1.5 million points were selected representing approximately 25 minutes
of data. These subsets were chosen to avoid the outliers that often occurred at the
beginning or the end of some of the recording sessions while still maintaining an equal
sample selected from each session. According to the stochastic nature of the EMG signal,
the instantaneous EMG data sample held relatively little information regarding the whole
muscle activity; therefore, another segmentation step was applied using a window of size
(1000 ms) on the EMG data of each experimental day.
c) Filtering
The physiological changes in the muscle fibers produced a myoelectric signal which
is known as the EMG signal [85]. When the EMG signal is obtained from dynamic
activities, it is usually contaminated with unwanted artifacts that might affect the
51
developed signal analysis and its interpretation. It must thus be carefully filtered before
proceeding to any forward processing step. The filtering signal process usually tries to
keep as much of the original signal information as possible, and then reduce or attenuate
the accompanying noise signals [164]. The quality of the EMG signal depends mainly on
the applied processing steps. Filtering represents an important step that eliminates the
combined noises from the raw EMG signal. The raw EMG data, obtained from daily
recordings, was filtered using:
1. Zero offset filter: eliminates the baseline offset noise and is implemented by using one
of the offset correction methods (“detrend”) in MATLAB.
2. Power-line noise filter: a notch filter with 60 Hz was also applied to eliminate the power
line noise.
3. Zero-phase bandpass filter: extracts the most effective band frequency region from
the raw signal by applying a bandpass filter (4th order Butterworth filter with a lower
and an upper cut off frequency of 10 and 450 Hz respectively). A phase shift problem
was solved by processing the input signal in both the forward and the backward
directions.
The EMG conditioning steps have been implemented using MATLAB software
(MathWorks, Natick, MA, USA).
3.3.2. General volitional muscle activity metric
3.3.2.1. Conventional EMG metrics
The goal of this section is to evaluate the ability to use some of the conventional
EMG amplitude metrics to evaluate volitional muscle activity pre- and post-SCI, as well
as while performing dynamic and multilevel muscle contractions. During the muscle
52
contraction, motor units recruit sequentially according to their size starting from the
smallest to the largest ones. First, the earliest MUs that are recruited, which are small in
size, generate a weak tension. Then to increase the level of contraction, the larger MUs
fire progressively until they reach the desired level [165]. A strong muscle contraction
generates a complex EMG signal, which is known as the interference pattern EMG signal.
This type of signal is built up from the activity of multiple motor units, and it results when
the MUs recruit at the same time, their potentials get overlapped, and they interfere with
each other [165].
A conventional EMG signal analysis, using signals detected during constant force
and isometric contractions, is useful in many applications. However, experiments that
represent real-life conditions are important to extend the reliability of these methods
especially for clinical applications [166]. It remains unclear whether existing EMG
features, which are largely used based on signals collected during constant-force
isometric contractions, are still effective for use with this type of EMG signal (dynamic
activity and multilevel muscle contractions).
An EMG signal has a stochastic and complex nature [167]. Therefore, to describe
the overall muscle activity, the averaged/windowed signal is utilized because it is more
informative than instantaneous EMG samples. EMG features are commonly calculated
by using a sliding window over the EMG signal [86], [138], [168]. The window length differs
based on the characterization of the signal (sampling frequency, total length), and the
application. The segmentation step could be applied using either a joint (overlapped) or
a disjoint windowing process [169], [170]. It has been observed that the overlapped
windowing process has a better classification performance compared to the disjoint
53
windowing [169]. The overlapped method does, however, require high computational
costs and time for some of the classification methods.
In this analysis, the filtered EMG signal for each individual day was used, then
followed with a segmentation step using a window of size 1000 ms. Initially, a group of
standard and popular EMG amplitude features were measured including area, RMS, turns
number and zero-crossing number. These features were selected according to their easy
implementation, fast computation, and class discrimination factors [71], [85].
Subsequently, the calculated features were tested in a linear mixed model to check their
ability to differentiate between the two experimental phases (pre- and post-lesion), and to
test the effect of the utilized treatment. The EMG features were calculated as follows:
1) Area is the integration of the absolute value of each EMG sample over the selected
duration and it reflects the quantity of the electrical impulse. It can be expressed as follows
[69], [71]:
𝐴𝑟𝑒𝑎 = ∑|𝑥𝑖| 𝑖 = 1, 2, … , 𝑁
𝑁
𝑖=1
(3.1)
𝑁: 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑎𝑚𝑝𝑙𝑒𝑠 , 𝑥𝑖 = 𝑡ℎ𝑒 𝑖 − 𝑠𝑎𝑚𝑝𝑙𝑒
2) Root mean square (RMS) is a statistical measure that represents the square root
of the mean of the squares of all the EMG signal samples. The RMS reflects the average
power of the EMG signal, and it is given by the following formula [69], [71]:
54
𝑅𝑀𝑆 = √1
𝑁∑ 𝑥𝑖
2
𝑁
𝑖=1
(3.2)
𝑁: 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑠𝑎𝑚𝑝𝑙𝑒𝑠 , 𝑥𝑖 = 𝑡ℎ𝑒 𝑖 − 𝑠𝑎𝑚𝑝𝑙𝑒
3) Zero-crossing is a simple frequency measure in the time domain which
represents the number of times that the EMG waveform crosses the amplitude zero line.
The zero-crossing counter will increase if two consecutive EMG samples satisfy the
following condition [69], [71]:
{𝑋𝑛 > 0 𝑎𝑛𝑑 𝑋𝑛+1 < 0} 𝑜𝑟 {𝑋𝑛 < 0 𝑎𝑛𝑑 𝑋𝑛+1 > 0}
4) Turns number is another feature that may measure the frequency content of the
signal by counting the number of times that the slope changes its sign from negative to
positive and vice-versa. The turns counter is incremented if three consecutive EMG data
points satisfy the following condition [48]:
{𝑋𝑛 > 𝑋𝑛−1𝑎𝑛𝑑 𝑋𝑛 > 𝑋𝑛+1} 𝑜𝑟 {𝑋𝑛 < 𝑋𝑛−1 𝑎𝑛𝑑 𝑋𝑛 < 𝑋𝑛+1}
3.3.2.2. Proposed peak number method (Q-metric)
According to the results of previous conventional EMG features and the mentioned
experimental settings (longitudinal recording, dynamic activity, and different levels of
contraction), a new general muscle activity metric is needed. The goal in this section is to
55
design a new metric, called Q-metric, to evaluate the general volitional muscle activity. In
calculating Q, the EMG data were processed based on the following fundamental
assumptions:
1. The lesion perturbs the EMG signal that was collected from the tail musculature.
2. The intramuscular electrodes detect all electrical EMG activity from all the motor units
‘close’ to the electrodes. Because motor units contain motor fibers, the EMG activity
also represents the EMG activity of all the motor fibers as well.
3. Insertion of the intramuscular electrode into the muscle may initially disrupt the tissue,
causing associated inflammation that may initially affect the signal variability; this
variability decreases as the inflammation subsides. The wire–tissue interface
therefore ‘matures’ overtime after initial implantation and produces more reliable
signals. Consequently, the EMG signal later in the recording period is less likely to be
affected by artifacts related to surgical implantation.
4. The number of peaks in the aggregate EMG signal acts as a surrogate of MUs
activation. From a clinical perspective, this activation translates into greater tail
movement.
5. The number of peaks or spikes in the interference (aggregated) EMG signal is
considered an important measurement, in which it is the only MU characteristic that
preserves itself in the interference EMG signal [171]–[173].
6. The average number of peaks reflects and estimates the activity of the MUs in the
pickup volume of the used electrode [106].
56
Figures (3.8) and (3.9) summarize the main steps of the proposed algorithm for
calculating the Q-metric which include:
a. Rectification: After the filtering step, the signals are rectified to convert all the
negative values to positive values. This process resolves positive–negative signal
cancellation.
b. Smoothing: A moving average window is applied for a continuous segment
containing 10,000 data points. This information is used to mathematically ‘smooth’
the rectified EMG signals and reduce the effect of artifacts.
c. Thresholding: The peaks with low amplitudes are discarded. This step ensures
that data from the motor units close to the electrode are included in the calculation
of Q.
57
Figure 3.8 Flowchart of the peak number algorithm
58
Figure 3.9 An illustration of the peak number method
The total number of peaks per daily recording session were calculated utilizing the
proposed EMG processing strategy. This value was reduced by the average (arithmetic
mean) number of peaks for pre-lesion period; that is, after the insertion of the transmitter
but before the creation of the spinal cord lesion. Finally, this number was divided by the
standard deviation in the number of peaks in the pre-lesion period. This approach
‘standardized’ the raw data relative to the activity in the pre-lesion period.
As such, for every day of EMG recording, a value of Q was calculated. The values of
Q were calculated separately for the muscles on the left and the right sides of the tail. A
mixed model was used to determine whether the tail movement before the lesion differed
from that afterward, and to ascertain whether combination therapy had an effect,
The first effect tested by using the model relates to the EMG consequences of the
lesion. Stated more explicitly, did the lesion cause a change in Q? The second effect
59
tested relates to whether the combination treatment (TRH, selenium and vitamin E)
resulted in improved Q scores when compared with an absence of treatment. In other
words, was the combination treatment effective?
Due to the longitudinal nature of the study, the rates of each effect were tested. When
the effects were not significant (P > 0.05), they were subsequently removed from the
model. Data from the right and left tail muscles were tested separately. The analyses
were conducted by using SAS 9.1 (SAS Institute, Cary, NC) and MATLAB (Mathworks,
Natick, MA).
3.3.3. Motor units discharge properties
To analyze and understand the discharge properties of the EMG signal, the signal
needed to be decomposed into its main functional components, which are MUPTs. The
MUPTs represents the sequences of MUPs from the active MUs within the EMG
recording area. Many decomposition methods have been developed and applied to
different types of EMG signals in automatic, semiautomatic, and manual ways [43]. The
decomposition task is difficult because it depends on many factors, such as the type of
electrode used and its position, the muscle condition (the clinical state of the
neuromuscular system of the patient), and the type of muscle contraction. An EMG
decomposition technique therefore needs to be able to address all the different factors in
effect and provide a robust result. Most of the decomposition methods require slight to
moderate muscle activity (< 50 % MVC) because the stronger muscle contraction, the
harder it is to analyze [43]. An EMG decomposition analysis helps in the
characterization/semi-quantification of pathological changes in a muscle in terms of
60
morphological changes to the MUPs. It also helps with the recruitment pattern of the MUs
and the discharge rate. This can be achieved by using different EMG features. The
recognition and semi-quantification outcomes belong to the MUs that are within the
volume of the recording of the intramuscular electrode.
In this NHP model, the induced lesion produced an SCI limited to the upper motor
neuron tracts that supply the lower motor neurons of the tail muscles [11]. Lesions of such
a nature resulted in a perturbation in the discharge properties of the MU’s, such as
abnormalities in the inter-discharge interval, the firing rate, and the floating serial
correlation coefficient [54], [55], [174]. These abnormalities have been typically
characterized in non-physiological experimental conditions, such as low force levels and
isometric contraction frequency [54], [55], [95], [96], [175]. According to the goal of this
work, EMG decomposition is a helpful tool to uncover the hidden mechanisms and
changes that result from the created SCI.
3.3.3.1. Conventional EMG decomposition method (MUPs analysis)
During the last three decades, the EMG signal has been broadly used by clinicians
and researchers as a powerful tool for the diagnosis of a variety of neuromuscular
diseases. Generally, EMG data has been analyzed qualitatively; however, more efforts
have recently been spent on quantitative EMG analyses. These analyses have been
extracted from different EMG quantitative features and evaluated according to their
diagnostic abilities. This transition is a result of the fact that qualitative tools depend on
the experience of the clinician, so they are a subjective assessment tool. Moreover,
61
longitudinal studies that were interested in measuring the severity and progression of a
disease required quantitative tools to facilitate the work [176].
The quantitative information regarding MUPs provides important diagnostic
information about neuromuscular disorders [106]. Various EMG quantitative analyses
have been introduced in the literature, including MUP analysis, IP, MUNE, and muscle
fiber jitter method [89]. EMG decomposition is one of the applied MUP analysis methods
[106]. The decomposition process includes three main steps: the detection of MUPs, the
clustering of MUPs, and the classification of MUPs into MUPTs. EMG decomposition is
the procedure that resolves the EMG signal into its constituent MUPTs based on the
morphological changes of MUPs. Each MUPT represents the electrical activity of an
individual MU; therefore, analyzing the activity of MUPTs provides information regarding
the discharge properties (firing rate) of the MUs, and the status of the MUPs that are
generated within a MUPT. In this work, an attempt was made to decompose an EMG into
its constituent functional components/motor units using a method similar to that of
Stashuk [106]. This method was established using the following three steps:
1. Detection and segmentation of MUPs
• Detection of MUPs:
The goal of this step is to isolate the active MUPs that form the MUPTs. It is known
that the active motor unit fibers closest to the electrode generate the highest amplitude in
motor unit potential; therefore, the thresholding technique [43], [94], [99], [106], [177],
[178] (>threshold value) will eliminate the low amplitude MUPs from motor fibers that are
further away. In this step, all of the signal peaks are scanned, and only the peaks that
62
exceed the detection threshold are considered to be candidates for MUP positions. There
are different signal features that could be utilized for calculating the threshold value, such
as the RMS, the max absolute, and the average absolute. In this work (1.5 RMS) was
selected [43].
• Segmentation of MUPs
A window with a specific length was applied over the MUP locations that were
detected in the previous step to extract the data that falls within the window. The extracted
segments of the signal were saved as candidate MUPs [179], [180]. The length of the
applied window was chosen based on the quality of the signal; the window length needed
to be large enough to cover the whole candidate MUP waveform, which usually contained
3 to 6 spikes. After a few experiments on the recorded EMG signals, it was found that a
suitable length was 180 data points.
• MUP feature extraction:
After the detection and segmentation of the candidate MUPs, various feature
extraction methods, including time domain, frequency domain, and morphological
domain, were applied to cover the meaningful information from the MUP waveforms.
Many feature combinations were formed and tested in this work, and the results that were
reported represent an example of these combinations. They consist of fifteen features,
including area, RMS, ZC, turn number, peak to peak, thickness, phase, phase complexity,
phase area, shape width, turn amplitude, turn area, length, mean frequency, and median
frequency, which describe the morphological changes, shape, and size of the MUPs.
63
Methods for calculating the selected features are explained in Chapter 2, Table (2.3). It
is important to note the size of the feature vector to avoid the curse of dimensionality [43].
2. The MUP clustering step
Clustering is the task of dividing a group of unlabeled objects into logical clusters.
These clusters contain some similar characteristics which describe these objects. The
clustering process aims to group similar objects into one cluster. Each cluster represents
the objects that are as similar as possible, while objects with more differences are divided
into different clusters. Any discriminative criteria therefore need to be able to increase the
similarity within one cluster and decrease it between groups.
After the detection and segmentation step for the MUP is performed, the estimated
number and the template shape of the MUPTs of the significant contributing motor units
need to be determined. These requirements can be achieved by applying the clustering
step. The first objective of this step is to allocate the candidate MUPs into a set of MUPTs
(clusters), in which each MUPT represents the electrical activity of one MU. Furthermore,
the average of all the MUPs that belong to one cluster (MUPT) defines the template shape
of that MUPT. Finally, the clustering step can help in determining the firing pattern of each
MUPT that is extracted.
3. MUP classification step
Decomposition techniques are designed mostly with a clustering step that is
followed by an MUP classification step [46], [181], [182]; however, some of these
techniques depend only on the MUP clustering step [178], [183], [184]. For the first group
of techniques, the thresholds of each MUPT template are determined after acquiring
64
some initial information, such as the number of classes (MUPTs), the initial shape
templates of MUPTs, and the MU firing rate statistics. The MUPs are then classified into
MUPTs through a number of iterations. When the MUPTs are stable, the classification
process is terminated. Finally, using the extracted MUPTs, different features are
calculated and based on these features, the status of a muscle is determined [44], [185].
In this work, according to the clustering results, the classification step was not applicable
(see Chapter 4 for more details).
3.3.3.2. The proposed EMG discharge properties analysis utilizing wavelet
transform decomposition and relative power
In this work, the intramuscular EMG data were analyzed using the WT as an EMG
decomposition method, and relative power (RP) as a metric of the active MUs within each
WT sub-band. Specifically, this work addresses the following questions:
1. During the pre-lesion period, do the subjects demonstrate symmetrical volitional tail
activity as measured by EMG activity?
2. In the post-lesion period, is there a change in the EMG activity attributed to the
experimental spinal cord injury and how it could be characterized in terms of
frequency content?
3. What is the difference in the EMG activity between the control and the treatment group
in the post-lesion period (treatment effect)?
65
In the NHP model of this work, the induced lesion produced an SCI limited to the upper
motor neuron tracts that supply the lower motor neurons of the tail muscles [11]. Lesions
of such nature result in a perturbation in the discharge properties of an MU; i.e.,
abnormalities in the inter-discharge interval, firing rate, and floating serial correlation
coefficient [54], [174]. These abnormalities have been typically characterized in non-
physiological experimental conditions, such as low force levels and isometric contraction
frequency [54], [95], [96], [175]. In this work, the EMG data were collected when the
subjects engaged in physiological activities (i.e., locomotion in a cage). The proposed
EMG analysis method consisted of the following two steps:
A. Decomposition step
A decomposition process was applied using wavelet transforms. Each WT sub-band
was assumed to represent the firing rate of a group of MUs. As well, it was assumed that
the RP of each individual sub-band reflected the level of activity for these MUs. Increases
or decreases in the RP may thus characterize the recruitment pattern process of the MUs
throughout different conditions in the experiment. The filtered EMG signals were broken
down into seven frequency sub-bands using the WT. DWT were selected for this work
because they have non-redundant results, and require less computational time and costs
[186], [187]. A Daubechies mother wavelet of fourth-order, ‘db4’, was used due to its
similarity to the triphasic pattern of the motor unit action potential, as seen in Figure (3.10)
[188]. Consistent with the analysis of other bio-signals, DWT decomposition was
performed using six levels [87], [187], [189]–[191]. Figure (3.11) illustrates the wavelet
analysis which was performed in two steps including:
66
a) The EMG signals were decomposed into seven sub-bands, one approximate
coefficient (cA6), and six detailed coefficients (cD1, . . ., cD6).
b) The EMG signal was then reconstructed at each level using an inverse discrete
wavelet transform. Seven EMG reconstructed signals (A6, D1, . . ., D6) were
obtained from their coefficients (cA6, cD1, . . ., cD6). Table (3.1) shows the
frequency ranges of the seven EMG sub-bands.
Figure 3.10 Mother wavelet function (db4)
Table 3.1 The frequency ranges of the seven EMG reconstructed sub-bands
Wavelet
Level Frequency Range/ Hz
Reconstructed EMG Sub-
Bands
1 250 - 500 D1
2 125 - 250 D2
3 62.5 - 125 D3
4 31.25 – 62.5 D4
5 15.63 – 31.25 D5
67
6 7.81- 15.63 D6
6 0- 7.81 A6
Figure 3.11 Wavelet analysis; decomposition and reconstruction steps [163]
B. Feature extraction step
To evaluate the changes in EMG sub-bands during different phases of the
experiment, these changes were characterized using the RP. The probabilistic distribution
of the spectral power was quantified by calculating the relative power of each spectral
component [192]. To obtain the RP, the power spectral density was first determined for
68
each reconstructed EMG sub-band signal. The RP for each individual sub-band was then
calculated using the following formula [193]:
𝑅𝑃(%) =𝑆𝐵𝑃
𝑇𝑃, (3.3)
where:
𝑅𝑃: is the relative power of the desired sub-band.
𝑆𝐵𝑃: is the power of the sub-band, such as the power of A6, D1, … or D6.
𝑇𝑃: is the total power of all the sub-bands (A6 + D1, ..., + D6).
C. Statistical analysis
The goal was to address the research questions by analyzing specific subsections
of the data and testing specific effects. To test the lateralization effect (attributed to limb
dominance), only the pre-lesion data were considered for the two sides (left and right).
The effect of the lesion was then tested using the data from both the pre- and the post-
lesion periods, and two separate models were fitted, one for each side. To test the
treatment effect, two separate models (left and right) were fitted using only the post-lesion
data. The data in this work is considered a clustered longitudinal dataset with three levels;
days are nested in frequency sub-bands, which are in turn nested within subjects (day is
Level 1, sub-bands is Level 2, and subject is Level 3). As a result of such a hierarchical
structure, shown in Figure (3.12), a non-independence problem for the observations
occurred, which required an appropriate statistical analysis method. A mixed model is a
statistical method that was developed specifically to address the non-independence
69
(correlated or repeated measurements) issue by including the random effect term in its
model. The developed mixed models controlled for the within-subject correlation by
including a random effect for the subject and the nested sub-bands within each subject,
with a variance-covariance structure and a restricted maximum likelihood estimation. The
models were implemented in the statistical software R using the lme4 package. The
significance level was set to 𝛼 = 0.05. The RP for any given subject at day i for sub-band
j nested within subject k, denoted as RPijk, is represented in Equation (3.4) [see a
summary of the parameters in Table (3.2)]:
𝑅𝑃𝑖𝑗𝑘 = 𝛽0 + 𝛽1 𝐷𝑎𝑦𝑖𝑗𝑘 + 𝛽2 𝐸𝐹𝐹𝐸𝐶𝑇𝑗𝑘 + 𝛽3 𝐸𝐹𝐹𝐸𝐶𝑇𝑗𝑘 𝐷𝑎𝑦𝑖𝑗𝑘
+ 𝛽4𝐹𝑟𝑒𝑞𝑗𝑘 + 𝛽5 𝐹𝑟𝑒𝑞𝑗𝑘 𝐷𝑎𝑦𝑖𝑗𝑘 + 𝛽6 𝐹𝑟𝑒𝑞𝑗𝑘 𝐸𝐹𝐹𝐸𝐶𝑇𝑗𝑘
+ 𝛽7 𝐹𝑟𝑒𝑞𝑗𝑘 𝐸𝐹𝐹𝐸𝐶𝑇𝑗𝑘 𝐷𝑎𝑦𝑖𝑗𝑘 (𝑓𝑖𝑥𝑒𝑑) + 𝛼0𝑘
+ 𝛼1𝑘 𝐷𝑎𝑦𝑖𝑗𝑘 + 𝛼0𝑗𝑘 + 𝛼1𝑗𝑘 𝐷𝑎𝑦𝑖𝑗𝑘 + 𝜀𝑖𝑗𝑘 (𝑟𝑎𝑛𝑑𝑜𝑚)
(3.4)
Table 3.2 Parameters used in the equation of the mixed model
Parameter Definition
𝑹𝑷𝒊𝒋𝒌 response (relative power)
𝑫𝒂𝒚𝒊𝒋𝒌 experiment day
𝑭𝒓𝒆𝒒𝒋𝒌 frequency sub-band (D1, D2, D3, D4, D4, D6, and A6)
EFFECT the main effect in the models, so for:
Model 1: the side effect [categorical variable with two levels (left and right
side)]
Model 2: the lesion effect [categorical variable with two levels (pre and post-
lesion)]
Model 3: the treatment effect [categorical variable with 2 levels (control and
treatment group)]
𝜷𝟎 𝐭𝐡𝐫𝐨𝐮𝐠𝐡 𝐭𝐨 𝜷𝟕 the fixed effect associated with the following: intercept, Day, EFFECT,
EFFECT*Day, Freq, Freq*Day, Freq* EFFECT, and Freq* EFFECT*Day
70
𝜶𝟎𝒌, 𝜶𝟏𝒌 the subject random effect associated with the intercept and Day slope,
respectively
𝜶𝟎𝒋𝒌, 𝜶𝟏𝒋𝒌 the random effect of a frequency sub-band nested within a subject associated
with the intercept and Day slope respectively
𝜺𝒊𝒋𝒌 random error
Figure 3.12 Collected Data Structure: The data in this work was collected from the left (L) and the right (R) side of the tail for five subjects throughout the days of the experiment (d1, d2, …, dn) for the pre- and post-lesion period. The data for each day was decomposed into seven frequency sub-bands (D1, D2, D3, …, A6). Three of the subjects received a treatment combining TRH, selenium, and vitamin E post-lesion (treatment group); the remaining two subjects did not receive any treatment post-lesion (control group) [163]
71
3..4 SCI Classification System
3.4.1 SCI classification system using classical ML classification technique
In this analysis, an SCI classification system has been investigated using various EMG
features and ML classification techniques. This analysis has been performed using the
following steps:
1. EMG features extraction and pre-processing
All the features that were measured in the previous sections, including the four EMG
amplitude features (area, RMS, turn number, and ZC number), the Q-metric, and the RP
of the EMG frequency sub-bands (RPA6, RPD6, RPD5, RPD4, RPD3, RPD2, and RPD1),
have been used as input data for the generated SCI classification systems. Two systems
were built; the first system utilized the EMG amplitude features, while the second system
was based on the two newly proposed features.
According to the experimental setting, more post-lesion instances were available than
the pre-lesion class, which generates a class imbalanced dataset. The imbalanced
dataset, in turn, leads to a misclassification problem that reduces the accuracy of a
classifier. The degree of imbalance is usually small, but it appears extensively in certain
applications, such as medical diagnoses, and fraud detection applications. As an
example, the medical dataset often included a higher percentage of ‘normal’ class
samples as compared to the ‘abnormal’ or ‘interesting’ class samples. In this case, the
misclassification rate of considering an ‘abnormal’ instance as a ‘normal’ instance is
higher than vice versa, which results in a poor diagnostic system. To resolve this issue,
there are some pre-processing steps that could be applied to rebalance the classes within
72
a dataset. In the literature, the problem of an imbalanced dataset was solved by re-
sampling the data in two ways, either under-sampling the majority class or oversampling
the minority class, such as in the synthetic minority oversampling technique (SMOTE)
[194].
In this work, SMOTE was applied to address the imbalanced class issue. SMOTE is
one of the pre-processing techniques that over-sample the pre-lesion (minority class) and
generate a balanced dataset. The algorithm takes samples of the feature space for the
pre-lesion class and its nearest neighbors and merges them to generate new instances
[194], [195].
2. Classification techniques
In this study, two different machine learning classification algorithms were used.
These classifiers were chosen according to the requirements and the feasibility of the
EMG signal dataset. The classification results were analyzed and compared to find the
most reliable type. The SVM and KNN algorithms were the two classification techniques
that were employed. The SMOTE approach and the classification algorithms were
implemented using a special software (Weka) by the University of Waikato, New Zealand.
The two different machine learning classification algorithms are as follow:
a) Support vector machine
The SVM technique is one of the machine learning algorithms that has been
applied to different biomedical tasks, including classification and regression [196]. The
SVM is a binary classifier that aims to differentiate between two classes by creating an
optimal hyperplane that separates the classes and satisfies the following equation:
73
𝑤. x + b = 0 (3.5)
where 𝑤 is a weight vector, 𝑥 is an input vector, and 𝑏 is a bias term [197]. For the linearly
separable classification problem, the SVM algorithm tries to find the separating
hyperplane with a maximum (margin). Given that the two classes are -1 and +1, the
separability assumption refers to the existence of some values of w and b in which all the
input vectors satisfy the following constraints in Equations (3.6) and (3.7) [198]:
𝑤. x + b ≥ +1 for ∀ y = +1 (3.6)
𝑤. x + b ≥ +1 for ∀ y = −1 (3.7)
These two can be combined into Equation (3.8):
𝑦(𝑤. x + b) − 1 ≥ 0 for ∀ y (3.8)
In this study, the kernel function of the SVM was tested empirically and the optimal
result was obtained using the Pearson VII function-based universal kernel (PUK) [199],
[200]. Figure (3.13) illustrates a general schematic of the SVM classification.
74
Figure 3.13 General schematic illustrating the SVM classification technique
b) K-nearest neighbors
KNN is a simple classification technique; however, it is an effective method that has
been applied to various applications such as healthcare, handwriting detection, and
image recognition [201]. The KNN is a non-parametric and a lazy approach, which refers
to the fact that no assumptions are made, and the structure of the model is based entirely
on the data given to it. This algorithm is popular for its speed and simplicity, and it has
been utilized widely in EMG classification studies [146], [147], [202]. Like any other
machine learning algorithm, KNN has some drawbacks or limitations that include
choosing the appropriate value of K [146], [203]; however, by varying the value of K, the
increase in generalization and accuracy of the KNN classifier can be achieved. Figure
(3.14) illustrates the three main steps of applying the KNN technique. These steps are
listed below [204], so as to predict a label for point (?) in Figure (3.14):
75
1. Firstly, calculating the distance between point (?) and all the remaining points by using
distance measuring methods, such as Euclidean distance, Manhattan distance,
Minkowski distance, or Hamming distance.
2. Secondly, finding the closest K neighbors to point (?).
3. Thirdly, classifying point (?) by the majority voting of the K neighbors, so if K=3, it
would be labelled as an orange star, while if K=7, it would be labelled as a green dot.
Figure 3.14 A graphical illustration of the K-nearest neighbor classification technique
3.4.2 SCI classification using DL technique compared to ML technique
A comparison study has been performed between two SCI classification systems
using DL classification (CNN) and ML (KNN), the main steps of the analysis as follows:
76
1. EMG data
In this section, the filtered EMG signal for each individual day of the experiment was
segmented into a series of disjoint windows of size 1000 ms. A different number of EMG
segments were next tested as inputs to the classification system. After multiple trials,
53,350 pre-lesion EMG segments and 135,300 post-lesion EMG segments combined for
all the subjects were chosen to be used as classifier inputs. This number was selected as
a compromise between the improvement in the classification accuracy and the time
consumed. According to the nature of the experiment, the collected data was imbalanced
since it was collected for more days during the post-lesion period compared to the pre-
lesion period. An imbalanced dataset leads to inaccurate classification results [195]. To
address this problem, the dataset was corrected by applying a random over-sampling
technique for the pre-lesion class. The number of samples in the pre-lesion class was
increased to be equal to the number of samples in the post-lesion class, so a total of
270,600 instances were used as inputs to the classification system.
2. Classification techniques
Two classification techniques have been used in this comparison. These techniques are
as follows:
i) KNN classification
Using the prepared EMG segments, four of the standard EMG amplitude features
were extracted, which included the area, the RMS, the turns number and the zero-
crossing number. All the features were then standardized and normalized. The created
feature vector was then utilized as an input to the KNN classifier. Different K values were
tested, and the best accuracy was obtained using k=9.
77
ii) CNN classification
In this work, the CNN architecture was chosen empirically by running multiple
experiments with different architectures by considering various numbers of layers, and
different numbers of filters with different sizes and numbers of strides. The final selected
network consisted of five blocks, illustrated in Figure (3.15) and Table (3.3). The CNN
was structured as follows:
Input: An EMG segment (1000 x 1)
Block 1 was composed of:
- A convolutional layer with 32 filters having a size 5 x 1 and a stride of 2
- An activation layer using rectified linear units (ReLUs) as activation functions
- A subsampling layer (max-pooling); also, a dropout of 0.2 was applied to reduce
the overfitting.
Block 2 was composed of:
- A convolutional layer with 32 filters having a size of 5 x 1 and a stride of 2
- An activation layer of ReLUs used as activation functions
- A subsampling layer (max-pooling), and a dropout of 0.2 was applied
Block 3 was composed of:
- A convolutional layer with 64 filters having a size of 3 x 1 and a stride of 1
- An activation layer of ReLUs used as activation functions
- A subsampling layer (max-pooling), and a dropout of 0.2 was applied
Block 4 was composed of:
78
- A convolutional layer with 128 filters having a size of 3 x 1 and a stride of 1
- An activation layer of ReLUs used as activation functions
- A subsampling layer (max-pooling), and a dropout of 0.2 was applied
These four blocks were followed by:
A global average pooling layer (GAP) which computed the average value of each
individual input feature map and yielded a single feature map obtained by concatenating
the computed average values.
A fully connected layer (FC) of size 100 x 1 with ReLUs used as activation functions
and a dropout of 0.3 was also applied.
An output layer: A sigmoid activation function was chosen for the output layer to satisfy
the binary classification requirement. This was done to classify the input data using the
output from the FC layer into either a pre-lesion class (< 0.5) or a post-lesion class (>
0.5).
Table 3.3 Optimized CNN architecture
No Architecture Parameters Output Shape No. of Parameters
1 CONV1: 32 x (5,1), stride: (2) 500 x 32 192
2 Maxpool: (2,1), stride: (2) 250 x 32 0
3 Dropout1: (0.2) 250 x 32 0
4 CONV2: 32 x (5,1), stride: (2) 125 x 32 5152
5 Maxpoo2: (2,1), stride: (2) 63 x 32 0
6 Dropout2: (0.2) 63 x 32 0
7 CONV3: 64 x (3,1), stride: (1) 63 x 64 6208
8 Maxpoo3: (2,1), stride: (2) 32 x 64 0
9 Dropout3: (0.2) 32 x 64 0
79
10 CONV4: 128 x (3,1), stride: (1) 32 x 128 24704
11 Maxpoo4: (2,1), stride: (2) 16 x 128 0
12 Dropout4: (0.2) 16 x 128 0
13 Global Average Pooling 128 0
14 FC: 100 (ReLU) 100 12900
15 Dropout1: (0.3) 100 0
16 Output: 1 (Sigmoid) 1 101
Total No of Parameters 49257
80
Figure 3.15 Convolutional neural network structures utilized to build the EMG-based SCI classification system
81
Chapter 4
Results and Discussions
4.1. Introduction
This chapter describes the detailed results obtained after the implementation of the
proposed methodology in Chapter 3. The current chapter is organized as follows:
• The general volitional muscle activity analysis is illustrated in two sections.
o The first section reports the results of the fitted statistical models for the four
conventional EMG amplitude features.
o The second section reports and illustrates the results of the proposed peak
number method (Q-metric).
• The analysis of motor unit discharge properties is described in another two sections.
o The third section presents and illustrates the results of using a conventional
EMG decomposition method.
o The fourth section reports and illustrates the results of the proposed EMG
discharge analysis approach using a wavelet decomposition technique and
relative power.
• The SCI classification system is described in four separate sections.
o The first section explains and illustrates the definition and the mathematical
formula for the classifier performance evaluation methods which were used.
82
o The second section includes the results of the SCI classification system using
the four EMG amplitude features.
o The third section describes the results of the SCI classification system using
the proposed Q-metric and WT-RPs features.
o The fourth section presents and compares the results of the new deep learning
(CNN) application as an SCI classification system to those of the classical KNN
classification technique.
4.2. General Volitional Muscle Activity Metric
4.2.1. Conventional EMG metrics
In this section, the popular features in the analysis of the EMG signal and the most
utilized features that describe the EMG signal amplitude have been calculated. These
features are area, RMS, turns, and ZC, which have been chosen based on easy
implementation, fast computation, and class discrimination factors [71], [85]. Each
individual feature was used as a response variable in a statistical mixed model to test the
effect of the created lesion and the combination of treatments that was utilized.
Tables (4.1) and (4.2) present the results of the statistical models of the left side EMG
signals that related to the effect of the created lesion and the treatment that was used
respectively. In Table (4.1), the results showed a significant difference (p-value <0.05) in
the turns and the ZC value between the pre- and the post-lesion period, with increments
in the post-lesion period of Turn/pre = 0.420, Turn/post = 0.49, ZC/pre = 0.507, and
ZC/post = 0.544. In comparison, the statistical models of the lesion effect using area and
83
RMS had nonsignificant differences (p-value>0.05), as seen in Table (4.1). The treatment
effect models showed nonsignificant results using all four of the calculated features (p-
value>0.05), as seen in Table (4.2).
Tables (4.3) and (4.4) present the results of the statistical models of the right side
EMG signals that are related to the effect of the lesion that was created and the treatment
that was used respectively. ZC showed a significant difference between the pre- and post-
lesion values (p-value<0.05), with a reduction in the post-lesion phase of ZC/pre = 0.801
and ZC/post = 0.636, while the area, the RMS, and the turns had a nonsignificant
difference between the two phases. The results related to the treatment effect models
showed nonsignificant results using all four of the calculated features (p-value>0.05), as
seen in Table (4.4).
As can be seen from Tables (4.1) to (4.4), there is no consistency in the results.
On the left side, the Turn and the ZC showed a significant difference with the increments
in their values in the post-lesion period. In comparison, on the right side only the ZC
showed a significant difference with a reduction during the post-lesion period. However,
the significant results in the ZC on both sides may imply a potential effect of the created
lesion on the frequency content of the EMG signals because ZC represents a simple
frequency measure in the time domain. This disturbance or change in the parameters of
the frequency fits the inferences of previous works that referred to an alteration in the
recruitment pattern and firing rate of motor units after the occurrence of a UMN lesion
[54], [55], [205].
84
Table 4.1 The results of the statistical model (lesion effect) of the conventional EMG features for the left side
Feature Mean (Pre) Mean (Post) p-value
Area 0.459 0.244 0.964
RMS 0.504 0.315 0.075
Turns 0.420 0.492 < 0.001
ZC 0.507 0.544 < 0.001
Table 4.2 The results of the statistical model (treatment effect) of the conventional EMG features for the left side
Feature Mean (Ctrl) Mean (Tr) p-value
Area 0.365 0.338 0.171
RMS 0.434 0.384 0.230
Turns 0.478 0.435 0.208
ZC 0.450 0.601 0.797
Table 4.3 The results of the statistical model (lesion effect) of the conventional EMG features for the right side
Feature Mean (Pre) Mean (Post) p-value
Area 0.260 0.321 0.576
RMS 0.079 0.328 0.106
Turns 0.765 0.565 0.894
ZC 0.801 0.636 < 0.01
Table 4.4 The results of the statistical model (treatment effect) of the conventional EMG features for the right side
Feature Mean (Ctrl) Mean (Tr) p-value
Area 0.282 0.299 0.589
RMS 0.206 0.202 0.802
Turns 0.718 0.613 0.083
ZC 0.689 0.749 0.438
85
4.2.2. The proposed peak number method (Q-metric)
According to the results obtained from the conventional EMG metrics and the unique
dataset (longitudinal data recordings, dynamic activity, and different levels of contraction),
a new muscle activity metric was needed. In the methodology section, a new metric,
called Q-metric, was proposed by applying multiple signal processing steps, which
included filtering, rectifying, smoothing, thresholding, and counting the number of peaks
that satisfied the calculated threshold. Next, a mixed model was fitted to test the effect of
the created lesion and the treatment that was used on the EMG signal represented in the
Q-metric.
Table (4.5) summarizes the Q values at key time points, and Figure (4.1) represents
a fitted linear regression model for the aggregate data of Q before and after the creation
of the lesion. Consistent with a maturation of the wire–muscle interface, the value of Q
for both the left and the right side of the tail increased during the pre-lesion period. Im-
mediately after the lesion, the value of Q decreased on both the left and the right side in
both the treatment and control groups. The Q scores before and after the creation of the
spinal cord lesion were significantly different (left side, P = 0.021; right side, P = 0.01)
with a reduction in the mean value of the Q-metric during the post-lesion period. This
finding suggests that the lesion led to decreased EMG activity, resulting in decreased tail
movement, and supports the construct that Q can be used as a measure of impairment
after experimental SCI. On the left side, the treatment group was associated with a trend
toward higher Q-values when compared with the untreated control group (P = 0.075); this
effect was not noted on the right side (P = 0.519). Post-lesion data was compared to the
86
pre-lesion baseline to identify consistent changes across the subjects. Fitted Q values
(Table 4.1) increased with time in both groups, although the values on both sides were
higher after treatment. Overall, the EMG data were insufficient to attribute the effect to
the treatment.
Figure 4.1 The aggregate fitted Q-values, representing tail movement, are plotted for the left (top) and the right (bottom) side. Values are normalized to a range between 0 and 1, with higher values indicating greater muscle activity. Of note, on both the left and the right side, the Q values decreased immediately after the creation of the lesion. On the left side, the effect of the lesion in the treatment group (orange) was attenuated, compared with that of the nontreatment group (grey). On both the left and the right side, the Q-metric improved over time [6]
87
Table 4.5 Standardized Q-metric
Subject Day of
implantation
Pre-lesion,
midpoint
Postlesion,
day 1
Post-lesion,
midpoint
Study end
Right side
1 0.548 0.693 0.016 0.322 0.645
2 0.553 0.696 0.016 0.320 0.639
3 0.552 0.696 0.016 0.320 0.641
4 0.604 0.731 0.014 0.283 0.566
5 0.585 0.719 0.015 0.296 0.591
Left side
1 0.855 0.902 0.005 0.103 0.207
2 0.695 0.793 0.011 0.217 0.435
3 0.798 0.863 0.007 0.144 0.288
4 0.836 0.888 0.006 0.117 0.235
5 0.948 0.965 0.002 0.037 0.073
The Q-metric for each subject increased after the time of insertion and reached a
plateau during the pre-lesion period. This pattern reflects the maturation of the wire–
muscle interface. The Q-metric decreased immediately after the creation of the lesion and
gradually increased (improved) over time. The values of Q relative to time are nonlinear.
Data from the pre- and post-lesion period suggest that a change is taking place and can
be measured in a quantifiable method. Given the EMG data from these experiments, Q
can be used to evaluate impairment after TSCI. Specifically, Q decreased immediately
after the experimental SCI and increased over time; however, the evidence was
insufficient to ascribe a treatment-associated effect. Regardless, the results from these
88
experiments have provided insights into the statistical variances of Q and will serve as a
guide or future experiments.
4.3. An Analysis of the EMG Discharge Properties
The discharge properties of the collected EMG signals were studied using two
different approaches. First, one of the conventional EMG decomposition techniques was
used. It was similar to that of Stashuk [106], which had been widely used in the literature.
This method was applied to test its efficacy with the unique nature of the EMG data
(dynamic, longitudinal, and multi-levels of contraction) from the current study. Second, a
combination of wavelet transforms, and relative power was used as a decomposition
method. The frequency content of the EMG signals was studied by decomposing the
signal into its constituent frequency sub-bands. The relative power of each sub-band was
then calculated and analyzed statistically using a mixed linear model.
4.3.1. Conventional EMG decomposition
In this section, the results after the implementation of the conventional EMG
decomposition method (isolating MUPs) are presented. This method includes three main
steps: MUP detection and segmentation, MUP clustering, and MUP classification into
MUPTs. The segmentation and MUP detection steps were applied using a thresholding
technique. This step requires choosing a certain threshold level and a specific length for
the segmentation window; in this study, it was 1.5 RMS [43] and 180 ms respectively.
According to this method, the EMG peaks that exceeded the selected threshold level were
then used as the center point for the segmentation window. The EMG data points that fell
within the selected window were finally segmented and saved as candidate MUPs.
89
Next, various types of morphological, time, and frequency domain features were
measured for each MUP segment, including area, RMS, ZC, turn number, peak to peak,
thickness, phase, phase complexity, phase area, shape width, turn amplitude, turn area,
length, mean frequency, and median frequency. The feature vector that was calculated
was later used as input for the clustering step. The goal of the clustering step was to find
the initial MUP templates to be used as labels to classify the candidate MUPs, using a
supervised classification technique in the third and final step.
Figure (4.2) illustrates some results from a previous study in the literature [43]; the first
column represents the candidate MUPs, the second shows valid MUPT templates, and
the last column shows the firing pattern for the MUs that was achieved by tracking the
MUP templates within the signal.
Figure 4.2 Example of valid and invalid MUPs, their MUP templates, and firing patterns [43]
90
Generally, valid MUP templates should resemble the results shown in the second
column of Figure (4.2). In this work, after applying the first two steps of the method (MUP
detection and segmentation, followed by the clustering step using the K-means method
with different values of K), invalid MUP templates were obtained, as shown in Figure (4.3).
The templates that were obtained confirmed that the applied decomposition method was
not applicable and could require EMG signals with certain characteristics that were not
available in the collected EMG dataset including high sampling frequency and isometric
muscle activity with a low level of contraction. Based on the obtained results (invalid MUP
templates), it was not possible to continue with this method, and another strategy needed
to be investigated.
(1)
(2)
91
(3)
(4)
(5)
92
(6)
Figure 4.3 From (1) to (6), different samples of the detected and clustered MUPs using the K-means clustering method
and fifteen various EMG features
4.3.2. The proposed EMG discharge properties analysis utilizing wavelet
transform decomposition and relative power
In this section, the results were presented in the form of answers to the three research
questions that were stated in Chapter 3. These questions are as follows:
1) During the pre-lesion period, do the subjects demonstrate symmetrical volitional tail
activity as measured by EMG activity?
To test the symmetry of the tail muscle activity on both sides, the RP of the EMG
signals during the pre-lesion period were analyzed using a linear mixed model. The
interaction of the frequency sub-band and the side variable in the statistical model
demonstrated a significant effect (p-value < 0.001). Given that the side*frequency
interaction is significant, Tukey’s mean comparisons were generated to identify what
means differed significantly. The results in Figure (4.4) suggest that the side variable had
a significant effect on the RP value of the D1, D2, D4, and D5 sub-bands. The estimated
93
mean of the RP for the D1, D2 and D4 sub-bands of the left side was significantly higher
than that of the right side. In contrast, the estimated mean of the RP for the D5 sub-band
on the left side was significantly lower than that of the right side. The estimated mean for
the RP across the different frequency sub-bands for both sides is illustrated in Figure
(4.4). Taken together, these results showed that the left and right tail muscles had
asymmetrical activation.
Figure 4.4 The estimated mean of the RP of seven reconstructed EMG sub-bands prior to the creation of SCI for the left and the right side of the tail. Of note for each band, there is a difference in the RP value when compared to the left and the right side of the tail. The D2 sub-band of the left and the right side has the maximum RP, and the significant difference between the two sides is at the D1, D2, D4, and D5 sub-bands. The star indicates a significant difference
[163]
In human beings, the construct of limb dominance or preference is well accepted;
i.e., most people use their right arm for functional tasks, and a minority use the left arm.
94
A select few people can use both arms equally well (ambidextrous). The question of
whether monkeys have limb dominance remains a subject of scientific inquiry. In this
analysis, there is asymmetry related to the RP of EMG data derived from the left and the
right tail muscles. These results suggest that the long-tailed macaque (Macaca
fascicularis) exhibits limb preference and/or dominance. This is also consistent with the
observations of the veterinary staff involved with this project. Generally, limb dominance
refers to the preferential use of one limb to perform functional tasks [206]. This asymmetry
in the pre-lesion (normal control) period has implications for any analysis subsequent to
the experimental SCI. Conceptually, volitional motor control involves a two-circuit
pathway; i.e., the upper motor neurons and the lower motor neurons of the corticospinal
tracts. The upper motor neurons originate in the cortex, travel the internal capsule and
pyramids, and terminate into the grey matter of the spinal cord. The lower motor neurons
originate in the grey matter of the spinal cord, exit via the nerve roots and reach the
muscles via the plexus and peripheral nerves. The data suggest that the neural network
that controls the left and right neurophysiological circuit is not symmetrical; therefore, it
stands to reason that experimental SCI will have different effects on the separate circuits.
This is consistent with human disease, to the extent that a lesion in one part of the central
nervous system may have different manifestations related to the limbs. As such, the
effects of the lesion, from a neurological perspective, were analyzed in Model Two
separately for both the left and the right side.
2) In the post-lesion period, is there a change in the EMG activity attributed to the
experimental spinal cord injury and how it could be characterized in terms of frequency
content?
95
To answer this question, the effect of the created lesion was analyzed using a linear
mixed model. The effect of the lesion was studied by testing the difference between the
EMG characteristics during the pre- and post-lesion period. The interaction of the
frequency sub-band and the lesion variable in the statistical models of both sides
demonstrated a significant effect (p-value < 0.001). Given that the lesion*frequency
interaction was significant, Tukey’s mean comparisons were generated to identify which
means differed significantly. Figures (4.5) and (4.6) summarized the estimated mean of
the RP values for different frequency sub-bands of the pre- and post-lesion group for the
two sides. On the left side, the estimated mean of the RP for the D4, D6 and A6 sub-
bands was significantly higher in the post-lesion period compared to the pre-lesion period.
On the right side, the estimated mean of the RP for the D4, D6, and A6 sub-band was
also significantly higher in the post-lesion period. In comparison, the estimated mean of
the RP for the D1 and D2 sub-bands was significantly lower in the post-lesion period
compared to the pre-lesion period. The results suggest that the created lesion had a clear
effect on the discharge properties of the MUs, and with this technique, changes in
discharge properties could be detected even when there is no clinical evidence.
96
Figure 4.5 The estimated mean of the RP of seven reconstructed EMG sub-bands prior and post to the creation of the SCI for the left side of the tail. Of note, the RP values for the frequency sub-bands (D4, D6, and A6) are significantly higher in the post-lesion period. The star indicates a significant difference [163]
In this analysis, there was a difference in the RP values when comparing the pre-
and the post-lesion data; this was statistically significant. Specifically, the low frequency
(LF) sub-bands (D4, D6, and A6) increased significantly during the post-lesion period on
both sides. As well, the high frequency (HF) sub-bands (D1 and D2) decreased
significantly during the post-lesion period on the right side. These results are consistent
with the literature [54]. As a general construct, neurological disease is associated with
abnormalities in the firing of the motor units and leads to a lower discharge frequency
[54], [55], [205]. This, in turn, can affect the distribution of the high and the low-frequency
components of the EMG signals. This shift in the RP value may be related to the
97
disorganized, spontaneous firing of single or multiple motor units, which affect the
frequency distribution. Speculatively, a component of these abnormalities was related to
spasticity; in SCI, the recruitment of Type I or Type II motor units could be preferentially
affected, which could in turn distort the frequency content. The elucidation of these
relationships requires further research. With this method, a change in the recruitment
behavior post-SCI might be detectable even before there is clinical evidence.
Figure 4.6 The estimated mean of the RP of seven reconstructed EMG sub-bands prior to and post to the creation of the SCI for the right side of the tail. Of note, the RP values for the lower frequency sub-bands (D4, D6, and A6) are significantly higher in the post-lesion period, while the higher frequency sub-bands (D1 and D2) are significantly lower in the post-lesion period. The star indicates a significant difference [163]
3) What is the difference in the EMG activity between the control and the treatment group
in the post-lesion period (treatment effect)?
98
The post-lesion data were utilized to generate two separate mixed models, one for
each side. The potential effect of the treatment was analyzed using the RP of the left and
the right sides, incorporating an analysis that compared the treatment and the control
groups. The interaction of the frequency sub-band and the treatment variables in the
models of both sides demonstrated a significant effect (p-value < 0.001). Given that the
treatment*frequency interaction was significant, Tukey’s mean comparisons were
generated to identify what means differed significantly. Figures (4.7) and (4.8) summarize
the estimated mean of the RP values for different sub-bands of the control and treatment
group for the two sides. On the left side, the D1, D2, D3, and D6 sub-bands have a
significant difference, while on the right side the effect was significant in all the sub-bands
except the D2. The results suggest that there is a significant difference in the discharge
properties of the MUs in the treatment and the control groups during the post-lesion
period.
99
Figure 4.7 The estimated mean of the RP of seven reconstructed EMG sub-bands post-lesion for the left side of the treatment (Tr) and the control (Ctrl) groups; of note, the RP values for the frequency sub-bands (D1, D2, D3, and D6) are significantly different. Subjectively, the distribution of the RP in the treatment group is similar to the RP distribution
in the pre-lesion period for all the subjects. The star indicates a significant difference [163]
100
Figure 4.8 The estimated mean of the RP of seven reconstructed EMG sub-bands post-lesion for the right side of the treatment (Tr) and the control (Ctrl) group; of note, the RP values for frequency sub-bands (D1, D3, D4, D5, D6, and A6) are significantly different. Subjectively, the distribution of the RP in the treatment group is similar to the RP
distribution in the pre-lesion period for all the subjects. The star indicates a significant difference [163]
The combination of TRH, selenium and vitamin E is associated with a treatment effect.
One of the goals of the current experiments was to gain a preliminary understanding of
the safety and efficacy of the combination of these three agents, which had not been
previously evaluated in animal models or in human studies. In a clinical trial involving
human patients who experienced SCI, TRH was administered as a 0.2-mg/kg bolus,
followed by an infusion of 0.2 mg/kg/h for 6 h [207]. For the current study, a similar weight-
based dosing paradigm was used: macaques in the treatment group received a bolus of
0.2 mg/kg with continuous infusion of 0.2 mg/kg/h for 1 h while anesthetized. Specifically,
three subjects received a combination treatment and two subjects did not receive
101
treatment. In the post-lesion period, there was a significant difference in the RP values
for most of the frequency sub-bands of the treatment and the control groups. These
results suggested there is a clear difference in the discharge properties of the MUs of the
treatment group during the post-lesion period. This finding was present for both the left
and the right sides. However, in the context of a limited number of subjects, this should
be considered as a preliminary inference. The rationale for this combination of treatment
for SCI and the implications thereof are published elsewhere [168].
4.4. SCI Classification System
EMG signals are one of the powerful biomarkers that have been used to detect and
track various neuromuscular abnormalities, such as myopathy and neuropathy disorders.
EMG signal interpretation is typically based on the visual inspection of expert clinicians.
Consequently, the diagnostic process might vary because it depends on the experience
of the clinician; therefore, developing a reliable EMG classification system may offer an
efficient assessment tool. Nonetheless, an analysis of EMG classifications may present
a robust and accurate detection method of the presence of SCI. As well, such an analysis
may expand the understanding of an SCI because a computerized analysis offers a
quantitative measure instead of a subjective visual inspection. In this study, EMG features
have been extracted using different methods, including the conventional amplitude
metrics, the Q-metric, and the RP of the WT sub-bands. All these features have been
utilized to build different classification systems using two artificial intelligence approaches:
classical machine learning techniques, which include KNN and SVM, as well as a new
application of deep learning (CNN), which has been submitted in this study.
102
4.4.1 Classification performance evaluation
The Performance of the employed classification techniques has been analyzed using
the following approaches:
1. Confusion matrix
The performance of a classification system can be described in a matrix table known
as a confusion matrix. The typical properties and the ratios that are included in the
confusion matrix for a binary classifier is illustrated in Figure (4.9). Each row of the matrix
represents the instances number in an actual class while each column represents the
instances number in a predicted class. In this study, the positive and negative classes
refer to the post- and pre-lesion data respectively. The confusion matrix consists of four
basic measures, which are as follow:
a. True Positives (TP): the number of correctly classified post-lesion instances.
b. True Negatives (TN): the number of correctly classified pre-lesion instances.
c. False Positives (FP): the number of pre-lesion instances, which have been
incorrectly classified as a post-lesion class.
d. False Negatives (FN): the number of post-lesion instances, which have been
incorrectly classified as a pre-lesion class.
103
Predicted Class
Positive Negative
Actu
al
Cla
ss
Po
sit
ive
True Positive
TP
False Negative
FN
Neg
ati
ve
False Positive
FP
True Negative
TN
Figure 4.9 The basic components of a confusion matrix [208], [209]
Figure (4.10) represents the confusion matrix for one of the classifiers that was
employed in this study. This classifier has been used as an example to explain the main
elements of the confusion matrix. The EMG data were labelled into two classes: pre- and
post-lesion. During the training step, the classifier learned how to distinguish between the
two classes (for a binary type) and build a model; later, it used this model to predict the
labels of the unseen instances during the testing step. In Figure (4.10), 15503 out of
24,600 total post-lesion samples were correctly classified as belonging to the post-lesion
class. In comparison, 15854 out of 24,250 total pre-lesion instances were accurately
classified as belonging to the pre-lesion class. As well, the FP = 8396 and the FN =
9097samples, which represent the pre- and post-lesion instances that were misclassified
respectively. The values of FP and FN are important error metrics that need to be reduced
depending on the application itself and cannot be generalized for all applications; for
example, minimizing FN is important when there is a serious life-threatening disease that
can be treated effectively during the early stages. But then again, in the case of high costs
or a high risk of follow-up therapy, minimizing FP is important when the disease is not life-
104
threatening for the patient. Therefore, to assess the performance of the classifiers,
especially for medical applications, the metrics that consider the effect of these two errors
(FP and FN) need to be considered.
Predicted Classes
Post-lesion Pre-lesion
Ac
tual
Cla
sse
s
Post-
lesion TP= 15503 FN= 9097 24600
Pre-
lesion FP= 8396 TN= 15854 24250
Figure 4.10 An example of the confusion matrix results
2. Statistical performance metrics
To evaluate the performance of the applied KNN and CNN classification approaches,
five statistical metrics, including sensitivity (recall), specificity, precision, accuracy, and F-
measure, were used [142]. These metrics were calculated using the four basic measures
that are included in the confusion matrix of a classifier, including TP, TN, FP, and FN. The
five metrics were defined and calculated as follows:
Sensitivity (Recall): This measures the ratio of the actual positives that are correctly
identified.
𝑇𝑃
𝑇𝑃 + 𝐹𝑁𝑥 100%
Specificity: This measures the ratio of the actual negatives that are correctly identified.
105
𝑇𝑁
𝑇𝑁 + 𝐹𝑃𝑥 100%
Precision: This measures the ratio of the predicted instances that were correctly identified
as positive to the total predicted instances as positive.
𝑇𝑃
𝑇𝑃 + 𝐹𝑃𝑥 100%
Accuracy: This measure how well a classifier has classified the classes correctly.
𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝐹𝑁 + 𝑇𝑁 + 𝐹𝑃𝑥 100%
F-measure: This is the harmonic mean of sensitivity and precision.
2𝑇𝑃
2𝑇𝑃 + 𝐹𝑁 + 𝐹𝑃𝑥 100%
4.4.2. SCI classification system using classical ML classification technique
In this work, two different groups of EMG features have been utilized with the classical
ML classification techniques. The results of these different features have been illustrated
as follows:
1. SCI classification using EMG amplitude features
The four amplitude EMG features (area, RMS, turn number, and ZC number)
mentioned in Section (3.3.2.1) were combined to form one feature set. This set was used
as an input for the first SCI classification system that was developed. The most common
machine learning classification techniques in the EMG area were employed, including
106
KNN and SVM. Figures (4.11), (4.12), (4.14), and (4.15) demonstrate the confusion
matrices of the classification techniques that were employed to the EMG data regarding
the left and the right EMG sides. The x-axis and the y-axis indicate the class of data (pre-
or post-lesion), and the diagonal cells of these figures show the rate of correctly classified
samples for both classes.
Using the SVM classifier, the post-lesion samples of the left side data were classified
correctly at a higher percentage (81%) compared to the pre-lesion sample (29%), as seen
in Figure (4.11). In comparison, the KNN confusion matrix demonstrated a balanced
prediction performance in terms of the number of correctly classified instances for both
classes (63% and 65%), as seen in Figure (4.12).
Figure 4.11 Confusion matrix of the SVM classifier using EMG amplitude features / left side
107
Figure 4.12 Confusion matrix of the KNN classifier using EMG amplitude features / left side
The performance of the classifiers was also evaluated using the five statistical
metrics. sensitivity, specificity, accuracy, precision, and F-measure values of the left side
are outlined in Figure (4.13). This figure shows that KNN scored the highest values of
specificity, accuracy, and precision metrics across the two classification methods. The
KNN classified the data with an accuracy of 64.1%, and the F-measure, 63.9%; in
contrast, the SVM classified the data with an accuracy of 54.9% and the F-measure,
64.2%. The SVM method with the worst performance was found to be biased when it
classified the pre-lesion samples (negative class) with a low specificity of 28.9%, while
the post-lesion samples (positive class) were classified with a reasonable degree of
sensitivity at 80.5%. According to the findings, the KNN is a more appropriate
classification method for the left side data using the EMG amplitude features.
108
Figure 4.13 Evaluation metrics of the KNN and the SVM classifier using EMG amplitude features/ left side
Using the SVM classifier with the right side data, the pre-lesion samples were
correctly classified at a higher percentage (87%) compared to the post-lesion sample
(56%), as seen in Figure (4.14). The KNN technique also demonstrated an imbalanced
prediction performance with a rate of 68% for the post-lesion class and 87% for the pre-
lesion class, as seen in Figure (4.15).
Figure (4.16) represents the sensitivity, specificity, accuracy, precision, and F-
measure values of the SVM and KNN classification using the right side data. The KNN
scored the highest values of all the metrics (except for specificity) across the two
classification methods. The KNN classified the data with an accuracy of 77.4%, and the
F-measure, 75.1%; in contrast, the SVM classified the data with an accuracy of 71.9%
and the F-measure, 66.7%. The SVM classifier also showed some biased behavior with
the right side data when it classified the pre-lesion samples (negative class) with a
specificity of 87.3%, while the post-lesion samples (positive class) were classified with a
80.6
28.9
54.9 53.5
64.363.0 65.4 64.2 64.9 63.9
0
10
20
30
40
50
60
70
80
90
100
Sensitivity Specificity Accuracy Precision F-measure
Pe
rce
nta
ge %
Evaluation Metric
SVM KNN
109
sensitivity of 56.4%. According to the performance metrics, the KNN is the more
appropriate classification method for the right side data using the EMG amplitude
features.
Figure 4.14 Confusion matrix of the SVM classifier using EMG amplitude features / right side
110
Figure 4.15 Confusion matrix of the KNN classifier using EMG amplitude features / right side
Figure 4.16 Evaluation metrics of KNN and SVM classifier using EMG amplitude features/ right side
56.4
87.3
71.9
81.6
66.768.1
86.8
77.583.8
75.1
0
10
20
30
40
50
60
70
80
90
100
Sensitivity Specificity Accuracy Precision F-measure
Pe
rce
nta
ge %
Evaluation Metric
SVM KNN
111
2. SCI classification using Q-metric and RP features
In sections (3.3.2.2.) and (3.3.3.2.), the EMG signal was characterized using new
amplitude and frequency-time features. First, the number of peaks was obtained from
each subject as a feature to develop the newly proposed general volitional muscle metric
(Q-metric). In the second feature, the frequency content of the signal was analyzed using
the wavelet transforms, then the RP of the EMG frequency sub-bands (RPA6, RPD6,
RPD5, RPD4, RPD3, RPD2, and RPD1) were calculated. The Q-metric and the RP features
that were obtained were then combined into one dataset. The dataset was then used as
input for the SVM and the KNN classifiers. Figures (4.17), (4.18), (4.19), and (4.20)
represent the confusion matrices of the applied classifiers using the left and the right side
data respectively. According to the figures, the two classification techniques
demonstrated a high and balanced predicting performance for both classes (the dark blue
diagonal cells) with a low error rate (light blue blocks).
112
Figure 4.17 Confusion matrix of the KNN classifier using the Q-metric and the RP features of the left side EMG data
Figure 4.18 Confusion matrix of the SVM classifier using the Q-metric and the RP features of the left side EMG data
113
Figure 4.19 Confusion matrix of the KNN classifier using the Q-metric and the RP features of the right side EMG data
Figure 4.20 Confusion matrix of the SVM classifier using the Q-metric and the RP features of the right side EMG data
114
According to the results, the two methods achieved a close performance using the
left side data, as seen in Figure (4.21). The accuracy was 80.9% and 80.3% for the SVM
and the KNN method respectively. Despite the fact that the two methods had similar
values for the accuracy, the KNN outperformed the SVM in terms of the F-measure, the
precision, and the sensitivity.
Figure (4.22) shows a comparison of the classification performance metrics
between the SVM and the KNN for right side data. It can be observed that the KNN
method outperformed the SVM method by having an accuracy of 91.7% while the
classification accuracy of the SVM was only 81.6%.
Figure 4.21 Evaluation metrics of the KNN and the SVM classifier using the RPs and the peak number features/ left side
80.4 81.5 80.9 81.4 80.975.5
87.180.3
89.381.8
0
10
20
30
40
50
60
70
80
90
100
Sensitivity Specificity Accuracy Precision F-Measure
Pe
rce
nta
ge %
Evaluation Metric
SVM KNN
115
Figure 4.22 Evaluation metrics of the KNN and the SVM classifier using the RPs and the peak number features/ right side
4.4.3. SCI classification using DL technique compared to ML technique
The results of the classification comparison analysis have been presented as follows:
1. CNN hyperparameters
To design a neural network, there is a group of variables that need to be set before
optimizing the weights of the network. These variables are known as network
hyperparameters, and they include variables such as the number of layers, the number
of nodes, the learning rate, the batch size, and the number of epochs. In this study, a
manual search method was applied to find a combination of the hyperparameters that
gives the best classification performance. Various combinations of the hyperparameters
were evaluated, and the best combination was selected to perform the classification task.
The hyperparameters were optimized using training and test sets and final results have
been reported using 5-fold cross validation. The same procedure was applied to the EMG
85.678.4 81.6
76.080.5
95.188.8 91.7
88.091.4
0
10
20
30
40
50
60
70
80
90
100
Sensitivity Specificity Accuracy Precision F-Measure
Pe
rce
nta
ge %
Evaluation Metric
SVM KNN
116
data of the right and the left side. Table (4.11) summarizes the hyperparameters that were
selected.
Table 4.6 The utilized hyperparameters
Optimizer Learning Rate Batch Size Epoch Loss Function
Adam 0.001 128 100 BC
The network was trained using the extracted EMG segments along with their
corresponding class labels (pre-lesion = Class 0, post-lesion = Class 1). The loss function
of binary cross-entropy (BC) was used as shown in Equation (4.1):
BC = −1
𝑛∑ 𝑥𝑖 . 𝑙𝑜𝑔(𝑦𝑖) + (1 − 𝑥𝑖). log (1 − 𝑦𝑖) 𝑛
𝑖=1
where yi is the predicted probability, xi is the actual probability (0 or 1), and n is the total
number of instances. All the experiments were conducted using a free cloud service
based on Jupyter Notebooks known as Google Colaboratory (Colab), which was
approved as an effective tool for deep learning in [210]. The CNN was implemented in
Keras (a Python front-end for deep learning) [211], and the Python Tensorflow library
[212] was used for the computational implementations.
2. CNN overfitting
Overfitting is a common problem in machine learning that happens when the model is
too complex and fits the training data very well (memorizing the data). To test the
performance of the proposed CNN classifier, the loss and the accuracy curves were
visualized for both sides when it was trained using 80% of the EMG segments and tested
117
using 20% of the segments. Figures (4.23), (4.24), (4.25), and (4.26) show the accuracy
and the loss curves of the CNN of the left and the right side respectively. In these figures,
both the training and the testing loss decreased continuously, and they were close
throughout the training process, which indicated that the proposed CNN did not face an
overfitting problem.
Figure 4.23 The accuracy curve for the EMG data classification of the left side using the proposed CNN architecture
118
Figure 4.24 The loss curve for the EMG data classification of the left side using the proposed CNN architecture
Figure 4.25 The accuracy curve for the EMG data classification of the right side using the proposed CNN architecture
119
Figure 4.26 The loss curve for the EMG data classification of the right side using the proposed CNN architecture
3. Performance metrics of KNN and CNN classification
Figures (4.27) and (4.28) show the five statistical metrics that were calculated for
the KNN and the CNN classifications of the EMG data for the muscles of the left and the
right side respectively. According to Figures (4.27) and (4.28), the KNN classified the
EMG data with an accuracy of 90.1% for the left side and 92.6% for the right side. As
well, the simple proposed architecture of the CNN performed reliably as it achieved 83.5%
and 87.1% for the left and the right side respectively. The CNN model had the most
competitive result for all of the evaluation metrics for both sides. This can be explained
by the advantage that the CNN has in its ability to use the temporal correlation existing
inside the EMG segments. As well, these results might indicate that the CNN has learned
the required information from the EMG signal that reflects the effect of the created lesions
on the muscle activity. These unique EMG features were learned during the successive
convolutional layers.
120
Figure 4.27 Evaluation metrics of the CNN and the KNN classifier / left side
Figure 4.28 Evaluation metrics of the CNN and the KNN classifier / right side
93.4
87.3
90.1
86.3
89.7
81.6
85.783.5
86.5
84.0
70.0
75.0
80.0
85.0
90.0
95.0
100.0
Sensitivity Specificity Accuracy Precision F-measure
Pe
rce
nta
ge %
Evaluation Metric
KNN CNN
92.293.2 92.6 93.2 92.7
88.7
85.787.1
85.186.9
70.0
75.0
80.0
85.0
90.0
95.0
100.0
Sensitivity Specificity Accuracy Precision F-measure
Pe
rce
nta
ge %
Evaluation Metric
KNN CNN
121
The main CNN power is its automated feature learning ability, which may play an
important role in the EMG signal field according to its complexity and nonstationary
nature; therefore, such an automated feature learning system might help in encoding the
hidden message of the motor control system to the skeletal muscles. As well, it may help
in characterizing the neuromuscular abnormalities (perturbations in the electrical activity
of the muscle) that occur as a result of neuromuscular diseases such as SCI.
Electromyography is a reliable and cost-efficient muscle activity assessment tool.
However, due to the complex process of generating this signal, analyzing and
understanding such a complicated signal is a challenging task. In this project particularly,
the EMG signal has more complexity because it was collected while the subjects were
performing their daily activities without any restrictions (freestyle movement). Even
though the CNN classification did not outperform the classical KNN method, it can still be
considered a reasonable method of preliminary classification. In future work, the SCI
classification system could be applied using a CNN system with a more advanced
structure.
4.5. Summary
Chapter 4 has presented a comprehensive intramuscular EMG analysis of the
NHP model of SCI. In order to examine the reliability of using dynamic EMG signals to
describe the SCI effects, this analysis considered the main characterization aspects of
EMG signals in terms of signal amplitude features, frequency content (pattern of
recruitment), and SCI classification ability using various techniques. Of note was the
challenge of using conventional EMG features to describe the SCI effects. As well, this
122
study submitted a preliminary volitional muscle activity metric (Q-metric) utilizing the peak
number of the EMG signal over a period of time. This metric could be advanced in the
future to be used as an SCI assessment tool. Moreover, this work proposed an EMG
decomposition method that included wavelet analysis and relative power calculations.
This method was able to detect the expected perturbations in the frequency components
of the EMG signals post-SCI. As well, it provided a preliminary inference that treatment
with a combination of TRH, selenium and vitamin E may improve SCI recovery. Together,
these results confirmed the reliability of using EMG signals as a measure of volitional
muscle activity for an NHP model.
An SCI diagnosis is currently based on a subjective visual assessment of different
biomarkers by specialists. The development of an SCI classification system would be an
essential means of advancing the diagnostic process because such a system would offer
an objective assessment tool using EMG signals. In chapter 3, different SCI classification
systems were developed using four selected EMG amplitude features, RP-WT features,
and Q-metric values, along with two different classification techniques (KNN and SVM).
A comparison study was performed to investigate the reliability of using a deep
learning (CNN) classifier with a raw EMG signal as an SCI classification system. The
results of the comparison indicate that a CNN with a simple architecture can provide a
level of performance close to the classical ML classification technique (KNN) for the
purpose of neuromuscular disease (SCI) classification. CNN has an advantage as it is
easier to implement in terms of the input data since it uses raw data. As well, throughout
123
the training process, it learns some representations (extracts some features
automatically) from the input data; therefore, it reduces the effort required for preparing
the engineered features.
124
Chapter 5
Conclusions and Future Work
5.1. Conclusions
• This chapter summarizes the main outcomes of the current work. The overriding goal
of this research is to build a valid NHP model for SCI with a reliable assessment tool
to characterize the injury neurophysiologically using intramuscular EMG signals. This
contribution could facilitate the development of a model to test new treatments for SCI,
and thereby assist in improving the functionality and quality of life for human beings.
• The work presents the results of experiments involving Macaca fascicularis, in which
a model of spinal cord injury (SCI) was created by using a balloon catheter inserted
into the epidural space. Prior to the creation of the lesion, an EMG telemetry device
was inserted to facilitate the measurement of tail movement and muscle activity before
and after SCI. The telemetry device was attached to two pairs of wire electrodes,
which were tunneled into the tail. This model is unique in that the impairment was
limited to the tail: the subjects did not experience limb weakness, bladder impairment,
or bowel dysfunction. The focus of the SCI model was to create a lesion that mimics
human SCI without clinical impairments.
• The tail of Macaca fascicularis can be considered a fifth limb, which is involved in the
performance of functional tasks and balance. The tail has well-developed sensory and
motor areas.
125
• The intramuscular recordings from this model are unique. Specifically, the EMG data
were obtained prior to and subsequent to SCI. Additionally, the data were obtained
while the subjects were moving freely (dynamic movement), over an extended period
of time (longitudinal).
• Some potential limitations included the limited numbers of subjects; this is, in part
related to the cost. As well, the longitudinal nature of the collection precluded a very
high sampling rate.
• Different analysis techniques were used to develop reliable EMG markers for the
detection and the classification of SCI.
• In this study, three specific objectives were set, and they were achieved using the
methodology suggested in Chapter 3. The results are summarized in the following:
1. Regarding the development of a metric for general volitional muscle activity, two
methods were examined, which included the following:
• First, four main EMG amplitude features (area, RMS, turn, and ZC) were selected
to be used as a metric of muscle activity. Based on the statistical analysis of the
results, the selected features showed inconsistent behavior regarding the effect of
the lesion that had been created and the treatment that was utilized. This could
indicate that conventional EMG metrics are overwhelmed if they are used as
characterization metrics for EMG signals that are collected in a dynamic fashion
and with different levels of contractions. A more specialized metric was therefore
needed to overcome the additional complexity of this type of EMG signal.
126
• The alternative EMG analysis paradigm that was proposed focused on the number
of peaks in the continuous signal of the EMG and it was called the Q-metric. Given
the EMG data from these experiments, the Q-metric could be used to evaluate
impairment after SCI. Specifically, the Q-metric decreased immediately after the
experimental SCI was implemented, and then increased over time. Evidence was,
however, insufficient to ascribe the changes to a treatment-associated effect.
2. In terms of the characterization of the discharge properties of the muscle motor units
during pre- and post-SCI for the treatment and the control group, and for agonist-
antagonist pairs of muscles, the following methods were examined:
• First, the most common EMG decomposition technique found in the literature was
applied to test its efficacy with the unique nature of the EMG signal in this study.
According to the acquired invalid results, this method was not applicable for such
a unique type of EMG data; therefore, another technique needed to be
investigated.
• Wavelet analysis and the relative power of EMG frequency sub-bands were
proposed as alternative decomposition methods. The RP of these sub-bands was
used as a response variable in various mixed models to analyze the effect of the
muscle side, the induced lesion, and the treatment that was used, on the collected
EMG data. In the absence of the SCI, the EMG data demonstrated an
asymmetrical activation of the two sides; this is consistent with the human
phenomenon of limb preference and/or dominance. Perturbation with an
experimental lesion resulted in a clear EMG consequence. Of note, the effect of
127
the lesion was different on the left and the right side. As well, there was a
preliminary inference that treatment with a combination of TRH, selenium and
vitamin E could improve recovery. The results indicated that the RP of the
decomposed EMG data added a new dimension to the evaluation of impairment
and recovery in this NHP model of experimental SCI. This analysis could lead to a
new assessment index for motor unit activity and progression of recovery from
SCI.
3. In terms of designing the SCI classification system:
• All the extracted EMG features in this work, including EMG amplitude features, Q-
metric, and the RP of the WT sub-bands, were utilized to develop the SCI
classification system using machine learning techniques that included SVM and
KNN.
• Using the EMG amplitude features, KNN and SVM techniques generally classified
the data with a low performance. The KNN outperformed the SVM with an F-
measure of 63.9.0% and 75.1% for the left and the right side respectively.
• In contrast, these two classification techniques showed a higher performance
using the Q-metric combined with the RP of the WT sub-bands features. Using
these features, the KNN technique outperformed the SVM method with an F-
measure of 81.8% and 91.4% for the left and the right side respectively.
• This study also presented a new application of deep learning (CNN) as an SCI
classification system using raw EMG segments as inputs and compared the results
to the classical machine learning classification (KNN).
128
• The performance of the two classifiers was measured and compared using five
performance metrics. The KNN technique outperformed the CNN methods with an
F-measure of 89.7% and 92.7% for the left and the right side respectively. The
new application of combining CNN and a raw EMG signal for neuromuscular
disease classification achieved a competitive degree of classification accuracy
when compared to the classical machine learning techniques. CNN classified the
data with an F-measure of 84.0% and 86.9%. for the left and the right side, while
the KNN, using hand-crafted EMG features, achieved an F-measure of 89.7% and
92.7% for the left and the right side respectively. As the CNN technique is still a
new application, more investigations need to be performed in order to build a clear
understanding of the ability and the reliability of using this deep learning technique
for SCI detection and characterization using the automated learned features.
• In conclusion, the EMG and the histopathologic data from this study have
advanced the understanding of SCI. The additional research required may
translate into novel treatments for people and animals with spinal cord injuries.
5.2. Future Work
• Several potential alternative approaches to analyzing EMG data were found. The
EMG analysis paradigms that were used in the current study focused on hand-crafted
EMG features. An alternative approach is to employ automated learned feature
methods, such as CNN, AE, and RNN, and check their efficacy in translating the
hidden changes in the EMG signal as a result of an SCI. In this study, a preliminary
application of CNN showed reasonable results. The automated learned features
129
approach could also help to make other conservative EMG analysis methods more
applicable for dynamic EMG data; e.g. it could help to overcome the problem of
applying the conventional decomposition method (MUPT method).
130
References
[1] M. Z. Al-Faiz, A. A. Ali, and A. H. Miry, “A k-nearest neighbor based algorithm for human arm movements recognition using EMG signals,” in 2010 1st International Conference on Energy, Power and Control (EPC-IQ), 2010, pp. 159–167.
[2] A. Ameri, M. A. Akhaee, E. Scheme, and K. Englehart, “Regression convolutional neural network for improved simultaneous EMG control,” Journal of Neural Engineering, vol. 16, no. 3, p. 036015, Jun. 2019, doi: 10.1088/1741-2552/ab0e2e.
[3] S. L. James et al., “Global, regional, and national burden of traumatic brain injury and spinal cord injury, 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016,” The Lancet Neurology, vol. 18, no. 1, pp. 56–87, Jan. 2019, doi: 10.1016/S1474-4422(18)30415-0.
[4] S. Nesathurai and D. Blaustein, Eds., The Rehabilitation of People With Spinal Cord Injury 3rd Edition. Arbuckle Academic Publishing, 2013.
[5] “National Spinal Cord Injury Statistical Center, Facts and Figures at a Glance. Birmingham, AL: University of Alabama at Birmingham, 2018.”
[6] N. Seth et al., “Model of Traumatic Spinal Cord Injury for Evaluating Pharmacologic Treatments in Cynomolgus Macaques (Macaca fasicularis),” Comp. Med., vol. 68, no. 1, pp. 63–73, Feb. 2018.
[7] J. Sledge et al., “Spinal cord injury models in non human primates: Are lesions created by sharp instruments relevant to human injuries?,” Medical hypotheses, vol. 81, no. 4, pp. 747–748, 2013.
[8] M. G. Fehlings and D. C. Baptiste, “Current status of clinical trials for acute spinal cord injury,” Injury, vol. 36, no. 2, pp. S113–S122, 2005.
[9] T. Denayer, T. Stöhr, and M. Van Roy, “Animal models in translational medicine: Validation and prediction,” New Horizons in Translational Medicine, vol. 2, no. 1, pp. 5–11, 2014.
[10] W. A. Graham, D. L. Rosene, S. Westmoreland, A. Miller, E. Sejdic, and S. Nesathurai, “Humane Non-Human Primate Model of Traumatic Spinal Cord Injury Utilizing Electromyography as a Measure of Impairment and Recovery,” Open Journal of Veterinary Medicine, vol. 03, no. 01, p. 86, Mar. 2013, doi: 10.4236/ojvm.2013.31014.
[11] S. Nesathurai et al., “Model of traumatic spinal cord injury in Macaca fascicularis: similarity of experimental lesions created by epidural catheter to human spinal cord injury,” Journal of medical primatology, vol. 35, no. 6, pp. 401–404, 2006.
[12] Y. S. Nout et al., “Animal models of neurologic disorders: a nonhuman primate model of spinal cord injury,” Neurotherapeutics, vol. 9, no. 2, pp. 380–392, 2012.
[13] E. S. Rosenzweig et al., “Extensive spontaneous plasticity of corticospinal projections after primate spinal cord injury,” Nature Neuroscience, vol. 13, p. 1505, Nov. 2010.
[14] H. Friedman et al., “The Critical Role of Nonhuman Primates in Medical Research,” Pathog Immun, vol. 2, no. 3, pp. 352–365, Aug. 2017, doi: 10.20411/pai.v2i3.186.
[15] M. Jucker, “The benefits and limitations of animal models for translational research in neurodegenerative diseases,” Nature medicine, vol. 16, no. 11, pp. 1210–1214, 2010.
[16] P. Konrad, “The ABC of EMG-Practical Introduction to Kinesiological EMG, power by Noraxon INC,” USA, Version, vol. 1, p. 5, 2005.
[17] C. F. Amorim and R. A. Marson, “Application of Surface Electromyography in the Dynamics of Human Movement,” Computational Intelligence in Electromyography Analysis - A Perspective on Current Applications and Future Challenges, Oct. 2012, doi: 10.5772/52463.
[18] S. A. Go, K. Coleman-Wood, and K. R. Kaufman, “Frequency analysis of lower extremity electromyography signals for the quantitative diagnosis of dystonia,” J Electromyogr Kinesiol, vol. 24, no. 1, pp. 31–36, Feb. 2014, doi: 10.1016/j.jelekin.2013.11.002.
131
[19] J. V. Basmajian, “Muscles Alive. Their Functions Revealed by Electromyography.,” Academic Medicine, vol. 37, no. 8, p. 802, 1962.
[20] “Acute Spinal Cord Injury,” Neurology. [Online]. Available: http://www.columbianeurology.org/neurology/staywell/document.php. [Accessed: 31-Oct-2019].
[21] N. Stifani, “Motor neurons and the generation of spinal motor neuron diversity,” Front Cell Neurosci, vol. 8, Oct. 2014, doi: 10.3389/fncel.2014.00293.
[22] A. Esquenazi, “Upper Motor Neuron Syndrome,” 2009. [23] C. B. Ivanhoe and T. A. Reistetter, “Spasticity: the misunderstood part of the upper motor neuron
syndrome,” American journal of physical medicine & rehabilitation, vol. 83, no. 10, pp. S3–S9, 2004. [24] K. Javed and D. T. Daly, “Neuroanatomy, Lower Motor Neuron Lesion,” in StatPearls, Treasure Island
(FL): StatPearls Publishing, 2019. [25] “Babinski reflex: MedlinePlus Medical Encyclopedia.” [Online]. Available:
https://medlineplus.gov/ency/article/003294.htm. [Accessed: 31-Oct-2019]. [26] I. Damjanov, J. Linder, and others, Pathology: a color atlas. Mosby Inc., 2000. [27] M. Capogrosso et al., “A brainspine interface alleviating gait deficits after spinal cord injury in
primates,” Nature, vol. 539, no. 7628, pp. 284-, Nov. 2016. [28] J.-C. Norreel, J.-F. Pflieger, E. Pearlstein, J. Simeoni-Alias, F. Clarac, and L. Vinay, “Reversible
Disorganization of the Locomotor Pattern after Neonatal Spinal Cord Transection in the Rat,” The Journal of Neuroscience, vol. 23, no. 5, pp. 1924–1932, Mar. 2003, doi: 10.1523/JNEUROSCI.23-05-01924.2003.
[29] A. M. Algotar et al., “Phase 3 clinical trial investigating the effect of selenium supplementation in men at high-risk for prostate cancer,” The Prostate, vol. 73, no. 3, pp. 328–335, Feb. 2013, doi: 10.1002/pros.22573.
[30] A. R. Allen, “SURGERY OF EXPERIMENTAL LESION OF SPINAL CORD EQUIVALENT TO CRUSH INJURY OF FRACTURE DISLOCATION OF SPINAL COLUMN: A PRELIMINARY REPORT,” JAMA, vol. LVII, no. 11, pp. 878–880, Sep. 1911, doi: 10.1001/jama.1911.04260090100008.
[31] Z. Ma et al., “A controlled spinal cord contusion for the rhesus macaque monkey,” Experimental Neurology, vol. 279, pp. 261–273, May 2016, doi: 10.1016/j.expneurol.2016.02.008.
[32] A. S. Rivlin and C. H. Tator, “Effect of duration of acute spinal cord compression in a new acute cord injury model in the rat.,” Surg Neurol, vol. 10, no. 1, pp. 38–43, Jul. 1978.
[33] T. G. Brown, “The intrinsic factors in the act of progression in the mammal,” Proc. R. Soc. Lond. B, vol. 84, no. 572, pp. 308–319, Dec. 1911, doi: 10.1098/rspb.1911.0077.
[34] E. Eidelberg, J. L. Story, B. L. Meyer, and J. Nystel, “Stepping by chronic spinal cats,” Exp Brain Res, vol. 40, no. 3, pp. 241–246, 1980, doi: 10.1007/bf00237787.
[35] E. Eidelberg, J. G. Walden, and L. H. Nguyen, “Locomotor control in macaque monkeys,” Brain, vol. 104, no. Pt 4, pp. 647–663, Dec. 1981.
[36] J. F. Fulton and C. S. Sherrington, “State of the flexor reflex in paraplegic dog and monkey respectively,” J. Physiol. (Lond.), vol. 75, no. 1, pp. 17–22, May 1932.
[37] P. R. Ojha, “Tail Carriage and Dominance in the Rhesus Monkey, Macaca Mulatta,” Mammalia, vol. 38, no. 2, pp. 163–170, 2009, doi: 10.1515/mamm.1974.38.2.163.
[38] M. Sharif-Alhoseini et al., “Animal models of spinal cord injury: a systematic review,” Spinal Cord, vol. 55, no. 8, p. 714, Aug. 2017, doi: 10.1038/sc.2016.187.
[39] F. D. Carlson and D. R. Wilkie, Muscle physiology, vol. 53. Prentice-Hall Englewood Cliffs, NJ:, 1974. [40] Y. Hong and R. Bartlett, Routledge handbook of biomechanics and human movement science.
Routledge, 2008. [41] J. Carlo. "De L. Basmajian, John V., Muscle Alive. 1985.
132
[42] C. J. De Luca, “The use of surface electromyography in biomechanics,” Journal of applied biomechanics, vol. 13, pp. 135–163, 1997.
[43] H. Parsaei, “EMG Signal Decomposition Using Motor Unit Potential Train Validity,” University of Watrlooo, Waterloo, Ontario, Canada, 2011.
[44] T. Kamali, “A Multiple Instance Learning Approach to Electrophysiological Muscle Classification for Diagnosing Neuromuscular Disorders Using Quantitative EMG,” 2018.
[45] S. Rasheed, “A multiclassifier approach to motor unit potential classification for EMG signal decomposition,” University of Waterloo, 2006.
[46] S. Rasheed, D. W. Stashuk, and M. S. Kamel, “Integrating Heterogeneous Classifier Ensembles for EMG Signal Decomposition Based on Classifier Agreement,” IEEE Transactions on Information Technology in Biomedicine, vol. 14, no. 3, pp. 866–882, May 2010, doi: 10.1109/TITB.2008.2010552.
[47] A. H. M. Almanji, “Wavelet signal processing of human muscle electromyography signals: a thesis in partial fulfilment of the requirement for the degree of Masters of Engineering in Mechatronics, Massey University, Albany, New Zealand,” 2010.
[48] M. Abdelmaseeh, B. Smith, and D. Stashuk, “Feature selection for motor unit potential train characterization,” Muscle & nerve, vol. 49, no. 5, pp. 680–690, 2014.
[49] M. N. Hadley et al., “Clinical assessment following acute cervical spinal cord injury,” Neurosurgery, vol. 72 Suppl 2, pp. 40–53, Mar. 2013, doi: 10.1227/NEU.0b013e318276edda.
[50] K. D. Anderson, A. Gunawan, and O. Steward, “Quantitative assessment of forelimb motor function after cervical spinal cord injury in rats: relationship to the corticospinal tract,” Experimental neurology, vol. 194, no. 1, pp. 161–174, 2005.
[51] Y. S. Nout et al., “Methods for functional assessment after C7 spinal cord hemisection in the rhesus monkey,” Neurorehabilitation and neural repair, vol. 26, no. 6, pp. 556–569, 2012.
[52] B. Calancie, M. R. Molano, and J. G. Broton, “EMG for assessing the recovery of voluntary movement after acute spinal cord injury in man,” Clinical Neurophysiology, vol. 115, no. 8, pp. 1748–1759, Aug. 2004, doi: 10.1016/j.clinph.2004.03.002.
[53] J. P. Lewko, “Assessment of muscle electrical activity in spinal cord injury subjects during quiet standing,” Paraplegia, vol. 34, no. 3, pp. 158–163, Mar. 1996.
[54] A. W. Wiegner, M. M. Wierzbicka, L. Davies, and R. R. Young, “Discharge properties of single motor units in patients with spinal cord injuries,” Muscle Nerve, vol. 16, no. 6, pp. 661–671, Jun. 1993, doi: 10.1002/mus.880160613.
[55] B. T. Shahani, M. M. Wierzbicka, and S. W. Parker, “Abnormal single motor unit behavior in the upper motor neuron syndrome,” Muscle Nerve, vol. 14, no. 1, pp. 64–69, Jan. 1991, doi: 10.1002/mus.880140111.
[56] P. Dolan, A. F. Mannion, and M. A. Adams, “Fatigue of the Erector Spinae Muscles: A Quantitative Assessment Using" Frequency Banding" of the Surface Electromyography Signal.,” Spine, vol. 20, no. 2, pp. 149–159, 1995.
[57] A. Phinyomark, C. Limsakul, and P. Phukpattaranont, “Application of wavelet analysis in EMG feature extraction for pattern classification,” Measurement Science Review, vol. 11, no. 2, pp. 45–52, 2011.
[58] A. Krasoulis, S. Vijayakumar, and K. Nazarpour, “Evaluation of regression methods for the continuous decoding of finger movement from surface EMG and accelerometry,” in Neural Engineering (NER), 2015 7th International IEEE/EMBS Conference on, 2015, pp. 631–634.
[59] Y. Wu, M. Á. M. Martínez, and P. O. Balaguer, “Overview of the Application of EMG Recording in the Diagnosis and Approach of Neurological Disorders,” 2013.
[60] D. Zennaro, P. Wellig, V. M. Koch, G. S. Moschytz, and T. Läubli, “A software package for the decomposition of long-term multichannel EMG signals using wavelet coefficients,” Biomedical Engineering, IEEE Transactions on, vol. 50, no. 1, pp. 58–69, 2003.
133
[61] A. Subasi, “Classification of EMG signals using combined features and soft computing techniques,” Applied soft computing, vol. 12, no. 8, pp. 2188–2198, 2012.
[62] L. Daniuseviciute et al., “Wavelet-based entropy analysis of electromyography during 100 jumps,” Elektronika ir Elektrotechnika, vol. 104, no. 8, pp. 93–96, 2015.
[63] M. R. Ahsan, M. I. Ibrahimy, and O. O. Khalifa, “Electromygraphy (EMG) signal based hand gesture recognition using artificial neural network (ANN),” in 2011 4th International Conference on Mechatronics (ICOM), 2011, pp. 1–6, doi: 10.1109/ICOM.2011.5937135.
[64] M. A. Oskoei and H. Hu, “Support vector machine-based classification scheme for myoelectric control applied to upper limb,” Biomedical Engineering, IEEE Transactions on, vol. 55, no. 8, pp. 1956–1965, 2008.
[65] K. Peter, “the abc of EMG,” A practical introduction to kinesiological Electromyography, version, vol. 1, 2005.
[66] J. Wang, L. Tang, and J. E. Bronlund, “Surface EMG signal amplification and filtering,” International Journal of Computer Applications, vol. 82, no. 1, pp. 15–22, 2013.
[67] C. J. De Luca, L. D. Gilmore, M. Kuznetsov, and S. H. Roy, “Filtering the surface EMG signal: Movement artifact and baseline noise contamination,” Journal of biomechanics, vol. 43, no. 8, pp. 1573–1579, 2010.
[68] A. Phinyomark, F. Quaine, S. Charbonnier, C. Serviere, F. Tarpin-Bernard, and Y. Laurillau, “EMG feature evaluation for improving myoelectric pattern recognition robustness,” Expert Systems with Applications, vol. 40, no. 12, pp. 4832–4840, Sep. 2013, doi: 10.1016/j.eswa.2013.02.023.
[69] A. Phinyomark, A. Nuidod, P. Phukpattaranont, and C. Limsakul, “Feature Extraction and Reduction of Wavelet Transform Coefficients for EMG Pattern Classification,” Elektronika ir Elektrotechnika, vol. 122, no. 6, pp. 27–32, Jul. 2012, doi: 10.5755/j01.eee.122.6.1816.
[70] M. Asghari Oskoei and H. Hu, “Myoelectric control systems—A survey,” Biomedical Signal Processing and Control, vol. 2, no. 4, pp. 275–294, Oct. 2007, doi: 10.1016/j.bspc.2007.07.009.
[71] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, “Feature reduction and selection for EMG signal classification,” Expert Systems with Applications, vol. 39, no. 8, pp. 7420–7431, 2012.
[72] R. N. Khushaba, A. H. Al-Timemy, A. Al-Ani, and A. Al-Jumaily, “A framework of temporal-spatial descriptors-based feature extraction for improved myoelectric pattern recognition,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 25, no. 10, pp. 1821–1831, 2017.
[73] G. R. Naik, “A comparison of ICA algorithms in surface EMG signal processing,” International Journal of Biomedical Engineering and Technology, vol. 6, no. 4, pp. 363–374, 2011.
[74] G. R. Naik, D. K. Kumar, and M. Palaniswami, “Signal Processing Evaluation of Myoelectric Sensor Placement in Low-level Gestures: Sensitivity Analysis Using Independent Component Analysis,” Expert Sys: J. Knowl. Eng., vol. 31, no. 1, pp. 91–99, Feb. 2014, doi: 10.1111/exsy.12008.
[75] S. P. Arjunan, D. K. Kumar, and G. Naik, “Computation and evaluation of features of surface electromyogram to identify the force of muscle contraction and muscle fatigue,” BioMed research international, vol. 2014, 2014.
[76] Y. Guo, G. R. Naik, S. Huang, A. Abraham, and H. T. Nguyen, “Nonlinear multiscale Maximal Lyapunov Exponent for accurate myoelectric signal classification,” Applied Soft Computing, vol. 36, pp. 633–640, 2015.
[77] G. R. Naik and H. T. Nguyen, “Nonnegative matrix factorization for the identification of EMG finger movements: Evaluation using matrix analysis,” IEEE journal of biomedical and health informatics, vol. 19, no. 2, pp. 478–485, 2014.
[78] H. Parsaei and D. W. Stashuk, “EMG Signal Decomposition Using Motor Unit Potential Train Validity,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 21, no. 2, pp. 265–274, Mar. 2013, doi: 10.1109/TNSRE.2012.2218287.
134
[79] H. Parsaei, D. W. Stashuk, S. Rasheed, C. Farkas, and A. Hamilton-Wright, “Intramuscular EMG Signal Decomposition,” CRB, vol. 38, no. 5, 2010, doi: 10.1615/CritRevBiomedEng.v38.i5.20.
[80] T. J. Doherty and D. W. Stashuk, “Decomposition-based quantitative electromyography: methods and initial normative data in five muscles,” Muscle Nerve, vol. 28, no. 2, pp. 204–211, Aug. 2003, doi: 10.1002/mus.10427.
[81] E. R. Kandel et al., Principles of neural science, vol. 4. McGraw-hill New York, 2000. [82] H. Parsaei, D. W. Stashuk, and T. M. Adel, “Decomposition of intramuscular EMG signals using a
knowledge -based certainty classifier algorithm,” in 2012 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2012, pp. 6208–6211, doi: 10.1109/EMBC.2012.6347412.
[83] D. Stashuk and G. M. Paoli, “Robust supervised classification of motor unit action potentials,” Medical and Biological Engineering and Computing, vol. 36, no. 1, pp. 75–82, 1998.
[84] A. Phinyomark, S. Thongpanja, H. Hu, P. Phukpattaranont, and C. Limsakul, “The Usefulness of Mean and Median Frequencies in Electromyography Analysis,” 2012, doi: 10.5772/50639.
[85] B. Hudgins, P. Parker, and R. Scott, “A new strategy for multifunction myoelectric control,” IEEE transactions on bio-medical engineering, vol. 40, pp. 82–94, Feb. 1993, doi: 10.1109/10.204774.
[86] J. Rafiee, M. A. Rafiee, F. Yavari, and M. P. Schoen, “Feature extraction of forearm EMG signals for prosthetics,” Expert Systems with Applications, vol. 38, no. 4, pp. 4058–4067, Apr. 2011, doi: 10.1016/j.eswa.2010.09.068.
[87] N. Burhan, M. Kasno, and R. Ghazali, “Feature extraction of surface electromyography (sEMG) and signal processing technique in wavelet transform: A review,” in 2016 IEEE International Conference on Automatic Control and Intelligent Systems (I2CACIS), 2016, pp. 141–146, doi: 10.1109/I2CACIS.2016.7885304.
[88] N. T. Artuğ, İ. Göker, B. Bolat, G. Tulum, O. Osman, and M. B. Baslo, “Feature extraction and classification of neuromuscular diseases using scanning EMG,” in 2014 IEEE International Symposium on Innovations in Intelligent Systems and Applications (INISTA) Proceedings, 2014, pp. 262–265, doi: 10.1109/INISTA.2014.6873628.
[89] L. J. Pino, “Neuromuscular clinical decision support using motor unit potentials characterized by ‘pattern discovery,’” University of Waterloo, 2008.
[90] V. Koch and H.-A. Loeliger, “EMG signal decomposition by loopy belief propagation.,” in ICASSP (5), 2005, pp. 397–400.
[91] P. Wellig and G. S. Moschytz, “Electromyogram decomposition using the single-lineage clustering algorithm and wavelets,” in Electronics, Circuits and Systems, 1999. Proceedings of ICECS’99. The 6th IEEE International Conference on, 1999, vol. 1, pp. 537–540.
[92] G. Kaur, A. S. Arora, and V. K. Jain, “Comparison of the techniques used for segmentation of EMG signals,” in Proceedings of the 11th WSEAS international conference on Mathematical and computational methods in science and engineering, 2009, pp. 124–129.
[93] D. Stashuk, “EMG signal decomposition: how can it be accomplished and used?,” Journal of Electromyography and Kinesiology, vol. 11, no. 3, pp. 151–173, 2001.
[94] R. S. Lefever and C. J. D. Luca, “A Procedure for Decomposing the Myoelectric Signal Into Its Constituent Action Potentials - Part I: Technique, Theory, and Implementation,” IEEE Transactions on Biomedical Engineering, vol. BME-29, no. 3, pp. 149–157, Mar. 1982, doi: 10.1109/TBME.1982.324881.
[95] D. W. Stashuk, D. Farina, and K. Søgaard, “Decomposition of intramuscular EMG signals,” Electromyography: Physiology, Engineering, and Noninvasive Applications, pp. 47–80, 2005.
[96] S. H. Nawab, R. P. Wotiz, and C. J. De Luca, “Decomposition of indwelling EMG signals,” Journal of Applied Physiology, vol. 105, no. 2, pp. 700–710, 2008.
135
[97] K. C. McGill, Z. C. Lateva, and H. R. Marateb, “EMGLAB: An interactive EMG decomposition program,” Journal of Neuroscience Methods, vol. 149, no. 2, pp. 121–133, Dec. 2005, doi: 10.1016/j.jneumeth.2005.05.015.
[98] K. C. McGill and L. J. Dorfman, “Automatic decomposition electromyography (ADEMG): validation and normative data in brachial biceps,” Electroencephalography and clinical neurophysiology, vol. 61, no. 5, pp. 453–461, 1985.
[99] J. Fang, G. C. Agarwal, and B. T. Shahani, “Decomposition of multiunit electromyographic signals,” IEEE Trans Biomed Eng, vol. 46, no. 6, pp. 685–697, Jun. 1999, doi: 10.1109/10.764945.
[100] X. Ren, Z. Wang, and X. Hu, “EMG signal decomposition based on wavelet transform and ICA method.,” Shuju Caiji yu Chuli(Journal of Data Acquisition & Processing), vol. 21, no. 3, pp. 272–276, 2006.
[101] C. I. Christodoulou and C. S. Pattichis, “Unsupervised pattern recognition for the classification of EMG signals,” IEEE Transactions on Biomedical Engineering, vol. 46, no. 2, pp. 169–178, Feb. 1999, doi: 10.1109/10.740879.
[102] D. Stashuk and Y. Qu, “Adaptive motor unit action potential clustering using shape and temporal information,” Med. Biol. Eng. Comput., vol. 34, no. 1, pp. 41–49, Jan. 1996, doi: 10.1007/BF02637021.
[103] X. Ren, X. Hu, Z. Wang, and Z. Yan, “MUAP extraction and classification based on wavelet transform and ICA for EMG decomposition,” Med Bio Eng Comput, vol. 44, no. 5, p. 371, Apr. 2006, doi: 10.1007/s11517-006-0051-3.
[104] P. Wellig, G. S. Moschytz, and T. Liiubli, “Decomposition of EMG signals using time-frequency features,” in Engineering in Medicine and Biology Society, 1998. Proceedings of the 20th Annual International Conference of the IEEE, 1998, vol. 3, pp. 1497–1500.
[105] P. Wellig, G. S. Moschytz, and T. Läubli, “Decomposition of electromyogram long-term recordings,” in [Engineering in Medicine and Biology, 1999. 21st Annual Conference and the 1999 Annual Fall Meetring of the Biomedical Engineering Society] BMES/EMBS Conference, 1999. Proceedings of the First Joint, 1999, vol. 1, pp. 578–vol.
[106] D. W. Stashuk, “Decomposition and quantitative analysis of clinical electromyographic signals,” Med Eng Phys, vol. 21, no. 6–7, pp. 389–404, Sep. 1999, doi: 10.1016/s1350-4533(99)00064-8.
[107] Z. Erim and W. Lin, “Decomposition of intramuscular EMG signals using a heuristic fuzzy expert system,” Biomedical Engineering, IEEE Transactions on, vol. 55, no. 9, pp. 2180–2189, 2008.
[108] A. E. Hramov, A. A. Koronovskii, V. A. Makarov, A. N. Pavlov, and E. Sitnikova, Wavelets in neuroscience. Springer, 2015.
[109] A. Shoeb and G. Cliord, “Chapter 16—wavelets; multiscale activity in physiological signals,” Biomedical Signal and Image Processing; Springer: New York, NY, USA, 2005.
[110] G. Wang, Y. Zhang, and J. Wang, “The Analysis of Surface EMG Signals with the Wavelet-Based Correlation Dimension Method,” Computational and Mathematical Methods in Medicine, 2014. [Online]. Available: https://www.hindawi.com/journals/cmmm/2014/284308/. [Accessed: 29-Nov-2017].
[111] I. Hostens, J. Seghers, A. Spaepen, and H. Ramon, “Validation of the wavelet spectral estimation technique in biceps brachii and brachioradialis fatigue assessment during prolonged low-level static and dynamic contractions,” J Electromyogr Kinesiol, vol. 14, no. 2, pp. 205–215, Apr. 2004, doi: 10.1016/S1050-6411(03)00101-9.
[112] R. C. H. So, J. K.-F. Ng, R. W. K. Lam, C. K. K. Lo, and G. Y. F. Ng, “EMG wavelet analysis of quadriceps muscle during repeated knee extension movement,” Med Sci Sports Exerc, vol. 41, no. 4, pp. 788–796, Apr. 2009, doi: 10.1249/MSS.0b013e31818cb4d0.
136
[113] M. González-Izal et al., “sEMG wavelet-based indices predicts muscle power loss during dynamic contractions,” J Electromyogr Kinesiol, vol. 20, no. 6, pp. 1097–1106, Dec. 2010, doi: 10.1016/j.jelekin.2010.05.010.
[114] J. Fang, G. C. Agarwal, and B. T. Shahani, “Decomposition of EMG signal by wavelet spectrum matching,” in Proceedings of the 19th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 1997, 1997, vol. 3, pp. 1253–1256 vol.3, doi: 10.1109/IEMBS.1997.756598.
[115] A. Merlo, D. Farina, and R. Merletti, “A fast and reliable technique for muscle activity detection from surface EMG signals,” IEEE Transactions on Biomedical Engineering, vol. 50, no. 3, pp. 316–323, Mar. 2003, doi: 10.1109/TBME.2003.808829.
[116] S. Sharma, Wavelet analysis based feature extraction for pattern classification from Singlechannel acquired EMG signal, vol. 50. 2012.
[117] T. Sharma and K. Veer, “EMG classification using wavelet functions to determine muscle contraction,” J Med Eng Technol, vol. 40, no. 3, pp. 99–105, 2016, doi: 10.3109/03091902.2016.1139202.
[118] R. Vetter, J. Schild, A. Kuhn, and L. Radlinger, “Discrimination of Healthy and Post-partum Subjects using Wavelet Filterbank and Auto-regressive Modelling,” Proceedings of the International Conference on Bio-inspired Systems and Signal Processing.
[119] J. M. Wakeling, S. S. M. Lee, A. S. Arnold, M. de Boef Miara, and A. A. Biewener, “A Muscle’s Force Depends on the Recruitment Patterns of Its Fibers,” Annals of Biomedical Engineering, vol. 40, no. 8, pp. 1708–1720, Aug. 2012, doi: 10.1007/s10439-012-0531-6.
[120] R. Yamada, J. Ushiba, Y. Tomita, and Y. Masakado, “Decomposition of electromyographic signal by principal component analysis of wavelet coefficients,” in IEEE EMBS Asian-Pacific Conference on Biomedical Engineering, 2003., 2003, pp. 118–119, doi: 10.1109/APBME.2003.1302612.
[121] R. Ali, “Denoising Electromyographic Signals via Stationary Wavelet Decomposition and Filtering,” The Tower Undergraduate Research Journal, vol. 6, no. 2, Oct. 2014.
[122] M. S. HUSSAIN, M. B. I. Reaz, F. Mohd-Yasin, and M. K. KHAW, Denoising and Analysis of Surface EMG Signals. 2006.
[123] D. Gradolewski, P. Tojza, J. Jaworski, D. Ambroziak, G. Redlarski, and M. Krawczuk, Arm EMG wavelet-based denoising system, vol. 317. 2015.
[124] M. R. Canal, “Comparison of wavelet and short time Fourier transform methods in the analysis of EMG signals,” J Med Syst, vol. 34, no. 1, pp. 91–94, Feb. 2010.
[125] P. Bartuzi and D. Roman-Liu, “Assessment of muscle load and fatigue with the usage of frequency and time-frequency analysis of the EMG signal,” Acta Bioeng Biomech, vol. 16, no. 2, pp. 31–39, 2014.
[126] L. A. Kallenberg and H. J. Hermens, “Behaviour of motor unit action potential rate, estimated from surface EMG, as a measure of muscle activation level,” J Neuroengineering Rehabil, vol. 3, p. 15, Jul. 2006, doi: 10.1186/1743-0003-3-15.
[127] R. Raut and D. A. A. Gurjar, “Bio-medical (EMG) Signal Analysis and Feature Extraction Using Wavelet Transform,” 2015. [Online]. Available: http://uoguelph.library.ingentaconnect.com/content/doaj/22489622/2015/00000005/00000003/art00019. [Accessed: 18-Dec-2017].
[128] A. R. Ismail and S. S. Asfour, “Continuous wavelet transform application to EMG signals during human gait,” in Conference Record of Thirty-Second Asilomar Conference on Signals, Systems and Computers (Cat. No.98CH36284), 1998, vol. 1, pp. 325–329 vol.1, doi: 10.1109/ACSSC.1998.750880.
[129] M. Khan, J. K. Singh, and M. Tiwari, “A Multi-Classifier Approach of EMG signal classification for Diagnosis of Neuromuscular Disorders,” p. 6, 2016.
137
[130] J. Kilby and K. Prasad, Continuous Wavelet Analysis and Classification of Surface Electromyography Signals. 2013.
[131] M. O. Diab, B. Moslem, M. Khalil, and C. Marque, “Classification of uterine EMG signals by using Normalized Wavelet Packet Energy,” in 2012 16th IEEE Mediterranean Electrotechnical Conference, 2012, pp. 335–338, doi: 10.1109/MELCON.2012.6196443.
[132] M. Hussain and Md. Mamun, “Effectiveness of the Wavelet Transform on the Surface EMG to Understand the Muscle Fatigue During Walk,” Measurement Science Review, vol. 12, no. 1, pp. 28–33, 2012, doi: 10.2478/v10048-012-0005-x.
[133] F. A. Mahdavi, S. A. Ahmad, M. H. Marhaban, and M. R. Akbarzadeh Totonchi, “Surface Electromyography Feature Extraction Based on Wavelet Transform,” International Journal of Integrated Engineering, vol. 4, Dec. 2012.
[134] M. Ozsert, O. Yavuz, and L. Durak-Ata, “Analysis and classification of compressed EMG signals by wavelet transform via alternative neural networks algorithms,” Comput Methods Biomech Biomed Engin, vol. 14, no. 6, pp. 521–525, Jun. 2011, doi: 10.1080/10255842.2010.485130.
[135] N. Gnanassegarane, “Classification of EMG Signals using Wavelet Features and Fuzzy Logic Classifiers,” National University of Singapore, 2009.
[136] J. Salvador, “Application of the Wavelet Transform for EMG M-Wave Pattern Recognition,” PhD Thesis, 2006.
[137] J. Yousefi and A. Hamilton-Wright, “Characterizing EMG data using machine-learning tools,” Computers in biology and medicine, vol. 51, pp. 1–13, 2014.
[138] R. N. Khushaba, S. Kodagoda, M. Takruri, and G. Dissanayake, “Toward improved control of prosthetic fingers using surface electromyogram (EMG) signals,” Expert Systems with Applications, vol. 39, no. 12, pp. 10731–10738, Sep. 2012, doi: 10.1016/j.eswa.2012.02.192.
[139] S. Koçer, “Classifying Myopathy and Neuropathy neuromuscular Diseases Using Artificial Neural Networks,” IJPRAI, vol. 24, pp. 791–807, 2010, doi: 10.1142/S0218001410008184.
[140] S. Koçer and A. E. Tümer, “Classifying neuromuscular diseases using artificial neural networks with applied Autoregressive and Cepstral analysis,” Neural Computing and Applications, vol. 28, no. S1, pp. 945–952, Dec. 2017, doi: 10.1007/s00521-016-2383-8.
[141] D. H. Hawkes, O. A. Khaiyat, A. J. Howard, G. J. Kemp, and S. P. Frostick, “Patterns of muscle coordination during dynamic glenohumeral joint elevation: An EMG study,” PLOS ONE, vol. 14, no. 2, p. e0211800, Feb. 2019, doi: 10.1371/journal.pone.0211800.
[142] A. Subasi, “Classification of EMG signals using PSO optimized SVM for diagnosis of neuromuscular disorders,” Comput. Biol. Med., vol. 43, no. 5, pp. 576–586, Jun. 2013, doi: 10.1016/j.compbiomed.2013.01.020.
[143] N. F. Güler and S. Koçer, “Classification of EMG Signals Using PCA and FFT,” J Med Syst, vol. 29, no. 3, pp. 241–250, Jun. 2005, doi: 10.1007/s10916-005-5184-7.
[144] G. Kaur, A. S. Arora, and V. K. Jain, “EMG diagnosis via AR modeling and binary support vector machine classification,” Int. J. Eng. Sci. Technol, vol. 2, no. 6, pp. 1767–1772, 2010.
[145] I. Elamvazuthi, N. H. X. Duy, Z. Ali, S. W. Su, M. K. A. A. Khan, and S. Parasuraman, “Electromyography (EMG) based Classification of Neuromuscular Disorders using Multi-Layer Perceptron,” Procedia Computer Science, vol. 76, pp. 223–228, Jan. 2015, doi: 10.1016/j.procs.2015.12.346.
[146] E. Gokgoz and A. Subasi, “Effect of multiscale PCA de-noising on EMG signal classification for diagnosis of neuromuscular disorders,” J Med Syst, vol. 38, no. 4, p. 31, Apr. 2014, doi: 10.1007/s10916-014-0031-3.
138
[147] R. Bose, K. Samanta, and S. Chatterjee, “Cross-correlation based feature extraction from EMG signals for classification of neuro-muscular diseases,” in 2016 International Conference on Intelligent Control Power and Instrumentation (ICICPI), 2016, pp. 241–245, doi: 10.1109/ICICPI.2016.7859710.
[148] M. Atzori, M. Cognolato, and H. Müller, “Deep Learning with Convolutional Neural Networks Applied to Electromyography Data: A Resource for the Classification of Movements for Prosthetic Hands,” Front Neurorobot, vol. 10, Sep. 2016, doi: 10.3389/fnbot.2016.00009.
[149] A. Alhyari, A. Shamli, Z. Abuwaimer, S. Areibi, and G. Grewal, “A Deep Learning Framework to Predict Routability for FPGA Circuit Placement,” in 2019 29th International Conference on Field Programmable Logic and Applications (FPL), 2019, pp. 334–341, doi: 10.1109/FPL.2019.00060.
[150] S. Lawrence, C. L. Giles, Ah Chung Tsoi, and A. D. Back, “Face recognition: a convolutional neural-network approach,” IEEE Transactions on Neural Networks, vol. 8, no. 1, pp. 98–113, Jan. 1997, doi: 10.1109/72.554195.
[151] X. Glorot, A. Bordes, and Y. Bengio, “Deep Sparse Rectifier Neural Networks,” in Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, 2011, pp. 315–323.
[152] X. Liu, H. Li, C. Lou, T. Liang, X. Liu, and H. Wang, “A New Approach to Fall Detection Based on Improved Dual Parallel Channels Convolutional Neural Network,” Sensors (Basel), vol. 19, no. 12, Jun. 2019, doi: 10.3390/s19122814.
[153] T. Jing Wei, A. R. B. Abdullah, N. B. Mohd Saad, N. B. Mohd Ali, and T. N. S. B. Tengku Zawawi, “Featureless EMG pattern recognition based on convolutional neural network,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 14, no. 3, p. 1291, Jun. 2019, doi: 10.11591/ijeecs.v14.i3.pp1291-1297.
[154] A. Hartwell, V. Kadirkamanathan, and S. R. Anderson, “Compact Deep Neural Networks for Computationally Efficient Gesture Classification From Electromyography Signals,” arXiv:1806.08641 [cs], Jun. 2018.
[155] U. Côté-Allard et al., “Deep Learning for Electromyographic Hand Gesture Signal Classification Using Transfer Learning,” arXiv:1801.07756 [cs, stat], Jan. 2018.
[156] U. Cote-Allard, C. L. Fall, A. Campeau-Lecours, C. Gosselin, F. Laviolette, and B. Gosselin, “Transfer Learning for sEMG Hand Gestures Recognition Using Convolutional Neural Network,” p. 6.
[157] X. Zhai, B. Jelfs, R. H. M. Chan, and C. Tin, “Self-Recalibrating Surface EMG Pattern Recognition for Neuroprosthesis Control Based on Convolutional Neural Network,” Front Neurosci, vol. 11, p. 379, 2017, doi: 10.3389/fnins.2017.00379.
[158] Y. Du, W. Jin, W. Wei, Y. Hu, and W. Geng, “Surface EMG-Based Inter-Session Gesture Recognition Enhanced by Deep Domain Adaptation,” Sensors (Basel), vol. 17, no. 3, Feb. 2017, doi: 10.3390/s17030458.
[159] W. Geng, Y. Du, W. Jin, W. Wei, Y. Hu, and J. Li, “Gesture recognition by instantaneous surface EMG images,” Sci Rep, vol. 6, p. 36571, 15 2016, doi: 10.1038/srep36571.
[160] U. Cote Allard et al., “A convolutional neural network for robotic arm guidance using sEMG based frequency-features,” in 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Daejeon, South Korea, 2016, pp. 2464–2470, doi: 10.1109/IROS.2016.7759384.
[161] T. I. Grand, “A mechanical interpretation of terminal branch feeding,” Journal of Mammalogy, vol. 53, no. 1, pp. 198–201, 1972.
[162] P. S. Rodman, “Skeletal differentiation of Macaca fascicularis and Macaca nemestrina in relation to arboreal and terrestrial quadrupedalism,” American Journal of Physical Anthropology, vol. 51, no. 1, pp. 51–62, 1979, doi: 10.1002/ajpa.1330510107.
[163] F. Masood et al., “Neurophysiological Characterization of a Non-Human Primate Model of Traumatic Spinal Cord Injury Utilizing Fine-Wire EMG Electrodes,” Sensors, vol. 19, no. 15, p. 3303, Jan. 2019, doi: 10.3390/s19153303.
139
[164] M. B. I. Raez, M. S. Hussain, and F. Mohd-Yasin, “Techniques of EMG signal analysis: detection, processing, classification and applications,” Biol Proced Online, vol. 8, pp. 11–35, Mar. 2006, doi: 10.1251/bpo115.
[165] K. R. Mills, “The basics of electromyography,” Journal of Neurology, Neurosurgery & Psychiatry, vol. 76, no. suppl_2, pp. ii32–ii35, Jun. 2005, doi: 10.1136/jnnp.2005.069211.
[166] P. Bonato, S. H. Roy, M. Knaflitz, and C. J. de Luca, “Time-frequency parameters of the surface myoelectric signal for assessing muscle fatigue during cyclic dynamic contractions,” IEEE Transactions on Biomedical Engineering, vol. 48, no. 7, pp. 745–753, Jul. 2001, doi: 10.1109/10.930899.
[167] T. Farrell, “Multifunctional Prosthesis Control: The Effects of Targeting Surface vs. Intramuscular Electrodes on Classification Accuracy and the Effect of Controller Delay on Prosthesis Performance.” Unpublished, 2007, doi: 10.13140/rg.2.1.5024.6000.
[168] M. A. Oskoei and H. Hu, Author’s personal copy Review Myoelectric control systems—A survey. 2007.
[169] K. Englehart and B. Hudgins, “A robust, real-time control scheme for multifunction myoelectric control,” IEEE Transactions on Biomedical Engineering, vol. 50, no. 7, pp. 848–854, Jul. 2003, doi: 10.1109/TBME.2003.813539.
[170] Y. Huang, K. B. Englehart, B. Hudgins, and A. D. C. Chan, “A Gaussian mixture model based classification scheme for myoelectric control of powered upper limb prostheses,” IEEE Trans Biomed Eng, vol. 52, no. 11, pp. 1801–1811, Nov. 2005, doi: 10.1109/TBME.2005.856295.
[171] L. A. Green, A. Christie, and D. A. Gabriel, “Spike shape analysis for the surface and needle electromyographic interference pattern,” Biomedical Signal Processing and Control, vol. 36, pp. 1–10, 2017.
[172] A. Magora and B. Gonen, “Computer analysis of the shape of spike from electromyographic interference pattern,” Electromyography, vol. 10, no. 3, pp. 261–271, Oct. 1970.
[173] R. Merletti, P. A. Parker, and P. J. Parker, Electromyography: Physiology, Engineering, and Non-Invasive Applications. John Wiley & Sons, 2004.
[174] C. J. D. Luca, “Control properties of motor units,” Journal of Experimental Biology, vol. 115, no. 1, pp. 125–136, Mar. 1985.
[175] F. Negro, S. Muceli, A. M. Castronovo, A. Holobar, and D. Farina, “Multi-channel intramuscular and surface EMG decomposition by convolutive blind source separation,” J. Neural Eng., vol. 13, no. 2, p. 026027, Feb. 2016, doi: 10.1088/1741-2560/13/2/026027.
[176] F. et al. Charles, “A review of clinical quantitative electromyography.,” Critical ReviewsTM in Biomedical Engineering, vol. 5, no. 38, 2010.
[177] G. H. Loudon, N. B. Jones, and A. S. Sehmi, “New signal processing techniques for the decomposition of EMG signals,” Medical and Biological Engineering and Computing, vol. 30, no. 6, pp. 591–599, 1992.
[178] C. D. Katsis, T. P. Exarchos, C. Papaloukas, Y. Goletsis, D. I. Fotiadis, and I. Sarmas, “A two-stage method for MUAP classification based on EMG decomposition,” Computers in Biology and Medicine, vol. 37, no. 9, pp. 1232–1240, 2007.
[179] H. Parsaei and D. W. Stashuk, “EMG signal decomposition using motor unit potential train validity,” Neural Systems and Rehabilitation Engineering, IEEE Transactions on, vol. 21, no. 2, pp. 265–274, 2013.
[180] A. O. Andrade, A. B. Soares, P. J. Kyberd, and S. J. Nasuto, EMG decomposition and artefact removal. INTECH Open Access Publisher, 2012.
[181] D. Stashuk and H. de Bruin, “Automatic decomposition of selective needle-detected myoelectric signals,” IEEE Trans Biomed Eng, vol. 35, no. 1, pp. 1–10, Jan. 1988, doi: 10.1109/10.1330.
140
[182] M. H. Hassoun, C. Wang, and others, “NNERVE: neural network extraction of repetitive vectors for electromyography. II. Performance analysis,” Biomedical Engineering, IEEE Transactions on, vol. 41, no. 11, pp. 1053–1061, 1994.
[183] C. D. Katsis, Y. Goletsis, A. Likas, D. I. Fotiadis, and I. Sarmas, “A novel method for automated EMG decomposition and MUAP classification,” Artificial Intelligence in Medicine, vol. 37, no. 1, pp. 55–64, 2006.
[184] M. Nikolic and C. Krarup, “EMGTools, an adaptive and versatile tool for detailed EMG analysis,” Biomedical Engineering, IEEE Transactions on, vol. 58, no. 10, pp. 2707–2718, 2011.
[185] Y. Wu, M. Á. M. Martínez, and P. OrizaolaBalaguer, “Overview of the Application of EMG Recording in the Diagnosis and Approach of Neurological Disorders,” 2013, doi: 10.5772/56030.
[186] S. R. Messer, J. Agzarian, and D. Abbott, “Optimal wavelet denoising for phonocardiograms,” Microelectronics Journal, vol. 32, no. 12, pp. 931–941, Dec. 2001, doi: 10.1016/S0026-2692(01)00095-7.
[187] N. K. Al-Qazzaz, S. Hamid Bin Mohd Ali, S. A. Ahmad, M. S. Islam, and J. Escudero, “Selection of Mother Wavelet Functions for Multi-Channel EEG Signal Analysis during a Working Memory Task,” Sensors, vol. 15, no. 11, pp. 29015–29035, Nov. 2015, doi: 10.3390/s151129015.
[188] C. J. D. Luca, “Physiology and Mathematics of Myoelectric Signals,” IEEE Transactions on Biomedical Engineering, vol. BME-26, no. 6, pp. 313–325, Jun. 1979, doi: 10.1109/TBME.1979.326534.
[189] N. K. Al-Qazzaz, S. H. M. Ali, S. Islam, S. A. Ahmad, and J. Escudero, “EEG Wavelet Spectral Analysis During a Working Memory Tasks in Stroke-Related Mild Cognitive Impairment Patients,” in International Conference for Innovation in Biomedical Engineering and Life Sciences, 2015, pp. 82–85, doi: 10.1007/978-981-10-0266-3_17.
[190] N. Mammone, F. L. Foresta, and F. C. Morabito, “Automatic Artifact Rejection From Multichannel Scalp EEG by Wavelet ICA,” IEEE Sensors Journal, vol. 12, no. 3, pp. 533–542, Mar. 2012, doi: 10.1109/JSEN.2011.2115236.
[191] G. Inuso, F. L. Foresta, N. Mammone, and F. C. Morabito, “Wavelet-ICA methodology for efficient artifact removal from Electroencephalographic recordings,” in 2007 International Joint Conference on Neural Networks, 2007, pp. 1524–1529, doi: 10.1109/IJCNN.2007.4371184.
[192] X. Hu, Z. Wang, and X. Ren, “Classification of surface EMG signal using relative wavelet packet energy,” Computer Methods and Programs in Biomedicine, vol. 79, no. 3, pp. 189–195, Sep. 2005, doi: 10.1016/j.cmpb.2005.04.001.
[193] K.-E. Ko, H.-C. Yang, and K.-B. Sim, “Emotion recognition using EEG signals with relative power values and Bayesian network,” Int. J. Control Autom. Syst., vol. 7, no. 5, pp. 865, Oct. 2009, doi: 10.1007/s12555-009-0521-0.
[194] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: Synthetic Minority Over-sampling Technique,” 1, vol. 16, pp. 321–357, Jun. 2002, doi: 10.1613/jair.953.
[195] V. López, A. Fernández, S. García, V. Palade, and F. Herrera, “An insight into classification with imbalanced data: Empirical results and current trends on using data intrinsic characteristics,” Information Sciences, vol. 250, pp. 113–141, Nov. 2013, doi: 10.1016/j.ins.2013.07.007.
[196] I. Guler and E. D. Ubeyli, “Multiclass support vector machines for EEG-signals classification,” IEEE Transactions on Information Technology in Biomedicine, vol. 11, no. 2, pp. 117–126, 2007.
[197] S. Amarappa and S. V. Sathyanarayana, “Data classification using Support vector Machine (SVM), a simplified approach,” International Journal of Electronics and Computer Science Engineering. ISSN-2277-1956, 2014.
[198] C. J. C. Burges, “A Tutorial on Support Vector Machines for Pattern Recognition,” pp. 43.
141
[199] B. Üstün, W. J. Melssen, and L. M. C. Buydens, “Facilitating the application of Support Vector Regression by using a universal Pearson VII function based kernel,” Chemometrics and Intelligent Laboratory Systems, vol. 81, no. 1, pp. 29–40, Mar. 2006, doi: 10.1016/j.chemolab.2005.09.003.
[200] K. A. Abakar and C. Yu, “Performance of SVM based on PUK kernel in comparison to SVM based on RBF kernel in prediction of yarn tenacity,” IJFTR Vol.39(1) [March 2014], Mar. 2014.
[201] D. J. Hand, H. Mannila, and P. Smyth, Principles of data mining. Cambridge, Mass: MIT Press, 2001. [202] Z. R. Yang, Machine Learning Approaches to Bioinformatics, 1st ed. River Edge, NJ, USA: World
Scientific Publishing Co., Inc., 2010. [203] G. Guo, H. Wang, D. Bell, Y. Bi, and K. Greer, “KNN model-based approach in classification,” in
OTM Confederated International Conferences" On the Move to Meaningful Internet Systems", 2003, pp. 986–996.
[204] R. Aldea, M. Fira, and A. Lazăr, “Classifications of motor imagery tasks using k-nearest neighbors,” in 12th Symposium on Neural Network Applications in Electrical Engineering (NEUREL), 2014, pp. 115–120.
[205] M. de Carvalho, A. Eisen, C. Krieger, and M. Swash, “Motoneuron firing in amyotrophic lateral sclerosis (ALS),” Front Hum Neurosci, vol. 8, Sep. 2014, doi: 10.3389/fnhum.2014.00719.
[206] H. Sadeghi, P. Allard, F. Prince, and H. Labelle, “Symmetry and limb dominance in able-bodied gait: a review,” Gait Posture, vol. 12, no. 1, pp. 34–45, Sep. 2000.
[207] L. H. Pitts, A. Ross, G. A. Chase, and A. I. Faden, “Treatment with thyrotropin-releasing hormone (TRH) in patients with traumatic spinal cord injuries,” J. Neurotrauma, vol. 12, no. 3, pp. 235–243, Jun. 1995, doi: 10.1089/neu.1995.12.235.
[208] T. Fawcett, “An introduction to ROC analysis,” Pattern Recognition Letters, vol. 27, no. 8, pp. 861–874, Jun. 2006, doi: 10.1016/j.patrec.2005.10.010.
[209] D. Powers, “Evaluation: From Precision, Recall and F-Factor to ROC, Informedness, Markedness & Correlation,” pp. 24.
[210] T. Carneiro, R. V. M. D. Nóbrega, T. Nepomuceno, G. Bian, V. H. C. D. Albuquerque, and P. P. R. Filho, “Performance Analysis of Google Colaboratory as a Tool for Accelerating Deep Learning Applications,” IEEE Access, vol. 6, pp. 61677–61685, 2018, doi: 10.1109/ACCESS.2018.2874767.
[211] F. Chollet, "Keras Documentation". keras. io. 2015. [212] M. Abadi et al., “TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed
Systems,” Mar. 2016.
142
List of Publications Some of the work described in this thesis previously appeared in:
1- Masood, F., Abdullah, H.A., Seth, N., Simmons, H., Brunner, K., Sejdic, E., & Schalk, D.R., et al. "Neurophysiological Characterization of a Non-Human Primate Model of Traumatic Spinal Cord Injury Utilizing Fine-Wire EMG Electrodes." Sensors 19, no. 15 (2019): 3303.
2- Seth, N., Simmons, H.A., Masood, F., Graham, W.A., Rosene, D.L., Westmoreland, S.V., & Cummings, S.M. et al. "Model of Traumatic Spinal Cord Injury for Evaluating Pharmacologic Treatments in Cynomolgus Macaques (Macaca fasicularis)." Comparative Medicine 68, no. 1 (2018): 63-73.
3- Masood, F., Farzana, M., Nesathurai, S., & Abdullah, H., "Comparison Study of Classification Methods of Intramuscular EMG Data for Non-Human Primate Model of Traumatic Spinal Cord Injury." Journal of Engineering in Medicine (under review)
4- Masood, F., Abdullah, H. A., Sharma, M., Mand, D., Simmons, H., Brunner, K., Schalk, D., Sledge, J., & Nesathurai, S., “A Novel Application of Deep Learning (Convolutional Neural Network) for Spinal Cord Injury Classification Using Raw EMG Signal.” (under review)
5- Seth, N., Masood, F., Sledge, J.B., Graham, W.A., Rosene, D.L., Westmoreland, S., & Abdullah, H.A. “Humane Non-human Primate Model of Traumatic Spinal Cord Injury: Quantitative Analysis of Electromyographic Data.” Open Journal of Veterinary Medicine, 5(07), (2015): p.161.
6- Gwardjan, B., Nitin, S., Masood, F., Sledge, J., Rosene, D., Westmoreland, S., & Abdullah, H. “Treating Acute Traumatic Spinal Cord Injury with Combination Therapy of Thyrotropin Releasing Hormone, Selenium and Vitamin E.” Journal of Neuropathology and Experimental Neurology Vol. 75, No. (6), (2016, June): pp. 605-606.
7- Fayez, O., Johnson, K., Simmons, H., Brunner, K., Schalk, D., Mensinkai, A., & Sedjic, E. “Pathological and Radiological Features of Traumatic Spinal Cord Injury”. In Journal of Neuropathology and Experimental Neurology 78(6), (2019, June): pp. 579-579.