Electromyography Signal Classification Using Deep

Learning
Mekia Shigute Gaso Selcuk Cankurt Abdulhamit Subasi
Department of Computer Science Department of Computer Science Institute of Biomedicine
Ala-Too International University Suleyman Demirel University University of Turku
Bishkek, Kyrgyzstan Kaskelen, Kazakhstan Turku, Finland
[email protected] Department of Computer Science [email protected]
Ala-Too International University
Bishkek, Kyrgyzstan
[email protected]
arXiv:2305.04006v1 [cs.LG] 6 May 2023

Abstract—We have implemented a deep learning model with exhibit quite noisy, complex, and high-dimensional character-
L2 regularization and trained it on Electromyography (EMG) istics [3].
data. The data comprises of EMG signals collected from control In the medical procedure when the need arise the physician
group, myopathy and ALS patients. Our proposed deep neural
network consists of eight layers; five fully connected, two batch may proceed to take the EMG signal of his subject. A muscle
normalization and one dropout layers. The data is divided showing weakness, lack of flexibility, inflicting pain during
into training and testing sections by subsequently dividing the movement are some of the symptoms that will help a physician
training data into sub-training and validation sections. Having to propose these procedure of acquiring EMG signals from
implemented this model, an accuracy of 99 percent is achieved on his patients. During these processes it is expected that the
the test data set. The model was able to distinguishes the normal
cases (control group) from the others at a precision of 100 percent electrodes will send a very small amount of electrical signals
and classify the myopathy and ALS with high accuracy of 97.4 to the patients nerves so that the later responds to it. These
and 98.2 percents, respectively. Thus we believe that, this highly data can then be transformed using computer programs into
improved classification accuracies will be beneficial for their use figures and numerical values that can later be interpreted by
in the clinical diagnosis of neuromuscular disorders. the physician [2].
Index Terms—EMG, deep learning, DNN, EMG signal classi-
fication, Myopathy, ALS.
EMG signals are affected in the processes of acquisition
of the signals from the reaction emerging from the muscles
in what is commonly known as the environmental noise.
I. I NTRODUCTION In addition to those there are effects emanating from the
intrinsic structure of the muscle. These are often related to
What is commonly refereed to as Electromyography (EMG) the interaction of the induced electric current with the muscle
is a process of obtaining and registering electrical signals from as the signal crosses different layers of the muscles as it
the musculoskeletal system in human body. Signals obtained penetrates them. The electrical changes resulted from the
in these ways can be used for various medical applications. interaction of the neuromuscular system of the muscles with
Some of these applications can generally be grouped into those that of the electrodes which is inserted in the muscles is
which are related to clinical and to those which focus on the commonly represented in what is known as the motor unit
interactions of humans with computer systems. All of these action potential (MUAP). The MUAP provides vital informa-
signals require to carefully decompose, processes and analyse tion in investigating if a subject is likely to have developed
the signal obtained in various intricate procedures to finally neuromuscular disorders. Therefore, all these collective factors
classify them into those which are showing medical disorders add up to the complex structure that is retained in the final
related to muscular functions and those which are healthy signals that is acquired, thereby introducing for a need to clean
subjects [1]. or/and de-noise the signal (see for example the discussions
The physical procedure of acquiring EMG signal involves given in [4] and references therein).
penetrating a very narrow needle which contains an electrode Resolving the above addressed issues in order to distil the
into the subject’s skin, and proceeding to record the signals most important information from EMG data requires develop-
as the muscles are contracting and relaxing following some ing various strategies for de-noising the signal. Together with
movement [2]. These signals are observed to be often very these methods, often an intermediate steps are introduced in
chaotic and to be strongly tied up to the intrinsic properties of processing the EMG data which are known as, the feature
the muscles themselves as they are related to their structures extraction methods. These methods have proved themselves
and their functionalities. As a result, these signals usually when combined with de-noising techniques of the EMG data
to have shown significantly higher classification accuracies, in

978-1-6654-0945-2/21/$31.00 ©2021 IEEE

order to classify the signals of subjects showing neuromuscular different level of contraction. Performance levels were cross
disorders from those who are not affected. For instance, checked for showing resemblance between the two algorithms
E. Gokgoz & A. Subasi in there work [4] have reported for both the experimental data there by comparing an accuracy
that using discrete wavelet transform (DWT) as a feature predicted priory for that of the simulated data. The median
extraction technique has increased the accuracy of the EMG agreement of the result between the two algorithms have been
classification. Following there suggestions we will also apply shown to be ∼96% accurate.
DWT in our work which is a method by which signals are Density-based method can be used to automatically decom-
split up into multiple frequency ranges. pose a single-channel intramuscular EMG signals into their
compositions trains, the MUAP’s [9]. In this method, outliers
II. R EVIEW OF RELATED LITERATURE that belongs to the superpositions and factitious potentials are
In the following we will briefly discuss some closely related recognized and were able to be excluded in advance of the
research works with our study reported in this paper. It is com- classification process. Then the MUAP templates are pointed
mon practice to record EMG signals either with a surface elec- out by an adaptive density-based clustering procedures. These
trodes or with a needle electrodes. By using the later case in decomposition techniques have been applied on signals ac-
acquiring the EMG signal, multidimensional multiscale parser quired from lower level contractions of a maximum of 30%
(MMP) has been employed for encoding electromyographic on various types of muscles. Having compared their result
signals in the work of [5]. They have conducted experiments with an expert manual decomposition, they report that their
with a real signals which was acquired in laboratory and there algorithm have identified 80% of the total ∼230 motor unit
by showing the technique they have used is a robust one trains with > 90% accuracy.
when compared with its counterparts. They have reported the EMG decomposition system called EMGTool which is
evaluated percent root mean of encoding EMG signal which believed to be adaptive and multifaceted tool for detailed EMG
has the False Positives of 91.1% and those which are correctly analysis has been introduced by [10]. This tool has a capacity
identified of 95%. to extract the constituent MUAPs and firing patterns (FPs)
EMG signal need to be decomposed so that information which can be used for numerical analysis from the EMG signal
is extracted from it that can be used for clinical purpose. acquired with a minimal effort for clinical diagnostics. The
During the decomposition process what is known as motor authors have shown these in this work, there by successfully
unit potential trains (MUPTs) is obtained the validity of which excluding critical parameters with a fixed threshold by turning
need to be ascertained as error may be resulted in the process them into an adaptive ones. The algorithm they employed for
[6]. Having employed various techniques for this purpose they the EMG signal decomposition have 3 stages, which are ded-
author of [6] have compared the outputs on simulated as well icated for that of segmentation, clustering, and resolution of
as real data. They have reported that the adaptive gap-based compound segments. Having introduced different techniques
Duda and Hart (AGDH) method they applied have shown an for validating the results, an accuracy of ∼95% have been
improved accuracies of about 91.3% and 94.7% in classifying obtained to accurately identify the FPs.
accurately the simulated as well as the real data, respectively. There is a growing interest from those classification schemes
Similar study has been carried out in assessing the validity which makes use of machine learning to apply decision tree
of MUPTs in reference [7]. The MUPTs are acquired having algorithms in order to classify biomedical signals. The later are
decomposed an EMG signal which in the first place is obtained often required to be thoroughly de-noised and followed by the
through a needle-detected EMG signal. The authors have introduction of quite efficient feature extraction technique to
proposed two techniques to find out about the validity of the finally obtain higher level of accuracies. For instance, E. Gok-
MUPTs based on the shape information decoded in the motor goz & A. Subasi in there work of [4] has classified the acquired
unit potential (MUP). Both of the proposed methods make use EMG signals by (i) employing some de-noising techniques
of the gap statistic and jump algorithms which are applied to for cleaning the signals, (ii) applying some feature extraction
find some statistical properties of the data such as finding the techniques for processing the data in order to obtain some of its
number of groups in the data sets. Among the two methods essential features, (iii) and finally have used some classifiers on
the gap statistics has excelled to have generated an accuracy of both testing and validating data sets. Thus, for de-noising the
∼92% and ∼94% on the simulated and real data, respectively. EMG signals they have used Multiscale Principal Component
The author of reference [8] have emphasized on the de- Analysis (MSPCA) which calculates the Principal Component
composition of the EMG signals into which they are pri- Analysis (PCA) of the wavelet coefficients and combine the
marily composed of, namely the motor unit (MU) firing results at defined scales. Some of the essential feautures of the
times and action potential shapes, as these can be applied data have been extracted using Discrete Wavelet Transform
for investigating neuromuscular disorders involving various (DWT) technique which is briefly discussed in section III-B.
clinical researches. They introduced multi-channel decompo- Lastly, the classification is performed by using decision tree
sition algorithms known as Montreal and Fuzzy Expert on algorithms such as C4.5, CART and random forests [4]. They
both simulated as well as experimental data. They analysed have reported for the robustness of these design they used
the performance of the two algorithms on the data which is in its ability to automatically classify the EMG signals into
acquired from twelve subjects of different age categories at those with neuromuscular disorder, i.e., myopathic, ALS or
those which are normal. Having compared the results through TABLE I
various performance measures they report that the overall G ENERAL INFORMATION REGARDING THE THREE GROUPS IN WHICH THE
SUBJECTS FROM WHICH AN EMG DATA IS TAKEN .
accuracy obtained in their work is ∼97% accurate. Thus they
have emphasised that their framework can seen to have the Groups and some of Control Myopathy ALS
capability on the classification of EMG signals with a good
their attributes group group group
accuracy.
Number of subjects 10 7 8
Following similar procedure, we proposed a deep neural net-
works for classifying EMG data instead of machine learning Age category 21-37 19-63 35-67
techniques. We have used the same data sets as in the work of Gender 4 male 2 Male 4 Male
the authors of [4]. We have introduced different architecture distribution 6 Female 5 Female 4 Female
which uses deep neural networks as given in Fig. 1 (see section Physical 6 very good All All
IV-C for the details) which has thus improved the accuracy. condition 3 good show show
The contribution of our work is therefore to outline that the
of the subjects 1 bad myopathy ALS
EMG signal classification accuracy has improved to be (99%)
from previous works (see for example [4] for comparison) by
making us of MSPCA de-noising methods and DWT feature
in 1984, which allows for a method of describing a generic
extraction techniques. In addition to these, this is achieved
function of spatial and temporal varying scales by braking
because of our new implementation of the deep learning model
it down into its constituents. A decomposition of a signal
that we will be presenting in the subsequent section.
The remaining parts of this paper is organized in the with a better time resolution could be obtained by making
following way: in the next part (Section III), we provide the use of the wavelet transform that splits up a signal into a
background section by introducing processing of the data with set of basic functions which are known to as a wavelets. In
a through explanation of the feature extraction techniques general, wavelet-based techniques are applicable methods in
used. In Section IV the experimental data and the subjects analyzing various types of varying signals like EMG. For
along with the implementation of the deep learning model example discrete wavelet transform (DWT) can be used to
is explained. Section V provides the results obtained from measure and to shift the so called the mother wavelet. A
our experiment with the use of various performance measures discrete-time signal can also be disintegrated into a bunch
along with the confusion matrix for both the test and validation of other smaller sets of signals by using DWT. The detailed
data sets followed by their corresponding discussions. In description of this process is outlined in greater details in
Section VI we provide the conclusion of our research work reference [14]. Following the procedure explained in reference
presented here. [4], a total of 27 features has been extracted which are used
in this work whose details are given in section IV-B.
III. BACKGROUND
Here we will provide the background of our research IV. I MPLEMENTATION OF THE D EEP L EARNING M ODEL
having introduced the methods of data processing with a brief
A. EMG Data
explanation of the feature extraction techniques used in this
work. Here we will give a brief description of the data sets
which were used in our work. We have adopted the data sets,
A. Multiscale Principal Component Analysis (MSPCA)
which were also used by E. Gokgoz & A. Subasi, which is
Multiscale Principal Component Analysis (MSPCA) com- given in reference [11]. The data is obtained from control
bines the characteristics and ability of Principal Components group, myopathy and ALS patients. We briefly summarized
Analysis (PCA) to de-correlate the variables by obtaining firm the information about the group from which the EMG signal is
inter-relationship. Wavelet analysis is employed to find quite acquired in a tabular form as in Table I. The EMG signals were
essential features and closely decorrelate the autocorrelated collected at low voluntary and constant level of contraction
values. At each scale, MSPCA calculates the PCA of the which is just above threshold by using concentric needle
wavelet coefficients and integrates the results at the marked electrode. The signals are acquired from 5 spots located at
scales. MSPCA is practicable for signal modeling which often varying location on the subjects body having inserted the
shows changes across time interval and frequency ranges. needles at a different level of penetration.
Thus, this characteristics of MSPCA can be taken to be the
advantage of the multiscale appraoch (see for example [4] B. Feature Extraction
for the detailed description of this approach). In this study
MSPCA is used for de-noising the raw EMG signals (see [12] EMG signals are segmented with a window length of 8192.
for the detailed discussion of this topic). Then we get 1200 instance for each of the three classes
(i.e., ALS, Myopathy and Control group), which makes the
B. Discrete Wavelet Transform (DWT) length of the dataset 3600. In this study, the EMG signals
Grossman and Morlet are the first ones in developing the are represented by using the following features of coefficients
mathematical framework describing the wavelet theory [13] given in the formulas listed below, see also reference [4].
(i) Mean of the coefficients for each sub-band:
Pn
Ci
Mean = i=1 .
n
(ii) Average power of the wavelet coefficients in each
sub-band:
PN
(Ci )2
Average = i=0 .
N
(iii) Standard deviation of the coefficients in each sub-band:
r Pn
2
i=1 (Ci − µ)
Standard deviation = .
n
(iv) Ratio of mean values of neighbouring sub-bands:
Pn
i=1 Ci
Ratio = Pn n .
j=1 Cj
n

The 1st and the 2nd features are extracted for evaluation of
the frequency distribution of the signal. The 3rd and the 4th
features are extracted for the evaluation of the changes in the
frequency distribution. Seven different features are extracted
from (i), (ii) and (iii) each; and six different features are
extracted from (iv). Hence a total of 27 features are extracted Fig. 1. The architecture of the proposed deep neural network used in this
which consist of, the mean, the average power of the wavelet work.
coefficients in each sub-band, the standard deviation, and the
ratio of the mean values of neighboring sub-bands. These
softmax activation function. The former consists of positive
extracted features are used on the wavelet coefficients so as
numbers whose sum is equal to 1.0. These numbers thus
to make them more applicable and feasible. These features
entails information which can be used as the classification
are calculated for D1–D6 and A6 frequency bands which
probabilities for the classification layer. This layer uses the
were then used as input to the classifiers. Higher classification
probabilities returned by the softmax activation function for
accuracy and lower computation cost is reported to be obtained
the corresponding input in order to assign the input to one of
when using “db4” wavelet filter [4]. Finally, we have extracted
the disjoint classes and calculates the losses. We have trained
a total of 27 features from the raw EMG signals, thus our
the deep neural network using Adam (derived from adaptive
dataset dimension became 3600 by 27.
moment estimation) optimization algorithm with a batch size
C. Classification using Deep Neural Network of 150 samples and the cross-entropy loss function by shuffling
the data at every epoch.
In this paper, we have implemented our own architecture of
deep neural network with L2 regularization (weight decay) and
V. R ESULTS AND D ISCUSSIONS
trained it on EMG data. As depicted in Fig. 1, our proposed
deep neural network contains eight layers; five fully connected, We have divided the data into training (80%) and testing
two batch normalization and one dropout layers. The overall (20%) sections. The training data is further divided into sub-
design of the network architecture consists of three main training (90%) and validation (10%) sections. We have used,
sections: (1) An initial section which involves a single layer to the sub-training data to train the network and update the
present the feature vector of EMG signals to the network, (2) weights, and the validation data to compute the accuracies at
four stages of fully connected layers with different number of the regular intervals during training. The deep neural network’s
hidden neurons (120, 90, 30, 5 nodes), and (3) a final section learning curves plotted in Fig. 2 illustrate the training progress
of a fully connected layer with 3 nodes and softmax activation and show the mini-batch loss and accuracy together with the
function. The last dense layer is used to combine the features validation loss and accuracy. To estimate the final accuracy of
which are used to classify the EMG signals. As a result, the the implemented deep learning model we have employed the
number of the nodes in the last dense layer should match the test data. Our model has achieved 99% accuracy on the test
number of classes in our data set. In this case, the output size data set.
is set to be 3 as in the number of the classes. Fig. 3 displays the confusion matrix for the test data set
The Leaky ReLU and the ReLU activation functions, which by using column and row summaries. The trained deep neural
introduces the non-linearity to layers is employed to generate network model distinguishes the normal cases from the others
the activations of fully connected layers. The outputs of without confusing them (i.e., with the precision of 100%)
the fully connected layer are normalised with the use of and classify the Myopathy and ALS with high accuracies of
 network but also improve the model’s accuracy. We have
used the L2 of weight and bias regularization on the first,
second and third dense layers, and carefully tuned the L2
regularization hyperparameter of weight decay to 0.000001.
We have found that using this rate, instead of setting it to its
default value of 0.01, is quite optimal in order to minimize
the overfitting of the implemented model.
When the weights are updated after each mini-batch, the
distribution of the inputs to layers may change, which is called
internal covariate shift. Batch normalization standardizes the
distribution of the inputs through the layers for each mini-
batch [16]. In our design we have added batch normalization
layers between the dense layers and Relu activations. We
observed that adding batch normalization layer stabilizes the
training process and improves the training accuracy. Dropout
layer avoids the neurons from co-adapting extremely and pre-
vents the overfitting [17]. Using both batch normalization and
Fig. 2. A graph of learning curves showing training progresses for various dropout techniques in the same dense layer is reported as not
epochs with the losses and accuracies together with their corresponding relevant [16]. Following reference [16], we have regularized
accuracies of the validation data sets. the first and the third dense layers using batch normalization
and the second dense layer using the dropout layer.
97.4% and 98.2%, respectively. Training a deep network with The following are the computational environments that
eight layers is challenging as they can be sensitive to the were used in this work. A deep-learning packages Tensorflow
initial configuration of the network and learning algorithm. (v2.3.1) [18] back-end and Keras API (v2.4.0) [19] have
Overfitting, which is simply memorization of the training data been used to build deep neural networks. We have trained
is a common problem for the traditional feed forward neural our proposed model on a PC system with Inter Core i5
networks including the deep neural networks. This problem 9th generation at 2.40 GHz CPU, GeForce GTX 1650 GPU
avoids the network to generalize the learning process and make and 16 GB RAM.
accurate predictions for new data. Overfitting of a network VI. C ONCLUSION
can be monitored using the plot of learning curves during the
In this work we have implemented a deep learning model
training process. To reduce overfitting we have used, the L2
with L2 regularization and trained it on EMG data. The data
regularization at the first, second and third fully connected
comprises of signals collected from control group, myopathy
layers, two batch normalization layers after the first and third
and ALS patients. We find a strong relationship between our
dense layers, and one dropout layer after the second dense
implementation of the deep neural network architecture, and
layer. There are several weight regularization techniques, such
de-noising and decomposition methods on the data, as results
as L1 and L2 regularizations, and each has a hyperparameter
obtained in these work have revealed. The contribution of this
that must be configured. In reference [15] it has been reported
study is therefore rests on developing an effective and efficient
that small amount of weight decay not only regularize the
EMG classification algorithms for intramuscular EMG signals.
Our proposed deep neural network contains eight layers; five
fully connected, two batch normalization and one dropout
layers. The data is divided into training and testing sections by
subsequently dividing the training data into sub-training and
validation sections. An accuracy of 99% is achieved on the
test data set. The model was able to distinguishes the normal
cases (control group) from the others at a precision of 100%
and classify the myopathy and ALS with high accuracy of
97.4% and 98.2%, respectively. Therefor, we believe that these
highly improved classification accuracies will be beneficial for
the clinical diagnosis of neuromuscular disorders.
ACKNOWLEDGMENT
M. S. G would like to give thanks to Ala-Too International
University for the scholarship opportunity which is given to
Fig. 3. Confusion matrix for the test data set. The three classes namely, the her by Ala-Too International University, and Dr. S. Cankurt
normal, mypathy and ALS are represented by ‘0’, ‘1’ and ‘2’, respectively. for the continuous guidance and for the wonderful supervision
on her M.Sc. research and Dr. R. R. Mekuria for the fruitful
discussions on some aspects of the writing up process of this
research paper.
R EFERENCES
[1] Reaz, Mamun Bin Ibne, M Sazzad Hussain, and Faisal Mohd-Yasin
(2006). “Tech-niques of EMG signal analysis: detection, processing,
classification and applications”. In:Biological procedures online8.1, pp.
11–35.
[2] William Morrison M.D. — Written by Danielle Moores — Updated
on March 20,2018 Medically reviewed by (2018). Electromyogra-
phy(EMG).URL:https://www.healthline.com/health/electromyography
(visited on 05/24/2021).
[3] Lashgari, Elnaz and Uri Maoz (2020). “Electromyography Classification
during Reach-to-Grasp Motion using Manifold Learning”. PLOS ONE
doi: 10.1371/journal.pone.0255926.
[4] E. Gokgoz, A. Subasi, Comparison of decision tree algorithms for
EMG signal classification Biomedical Signal Processing and Control,
18 (2015), 138–144.
[5] Eddie Filho, BL, Eduardo AB da Silva, Murilo B de Carvalho, (2008).
“On EMG signal compression with recurrent patterns”. In:IEEE trans-
actions on biomedicalengineering55.7, pp. 1920–1923.
[6] Parsaei, Hossein, and Daniel W. Stashuk. ”Adaptive motor unit potential
train validation using MUP shape information.” Medical engineering &
physics 33.5 (2011): 581-589.
[7] Parsaei, Hossein and Daniel W Stashuk (2010). “Evaluating a motor unit
potential train using cluster validation methods”. In:Canadian Student
Conference onBiomedical Computing and Engineering, pp. 31–35
[8] Li, Yejin et al. (2013). “Cross-comparison between two multi-channel
EMG decomposition algorithms assessed with experimental and simu-
lated data”. In:201339th Annual Northeast Bioengineering Conference.
IEEE, pp. 191–192
[9] Marateb, H. R., Muceli, S., McGill, K. C., Merletti, R., & Farina, D.
(2011). “Robust decomposition of single-channel intramuscular EMG
signals at low force levels”. In:Journal of neural engineering8.6, p.
066015
[10] Nikolic, Miki and Christian Krarup (2010). “EMGTools, an adaptive
and versatiletool for detailed EMG analysis”. In:IEEE transactions on
biomedical engineering58.10, pp. 2707–2718.
[11] Nikolic M. Detailed Analysis of Clinical Electromyography Signals
EMG Decomposition, Findings and Firing Pattern Analysis in Controls
and Patients with Myopathy and Amytrophic Lateral Sclerosis. PhD
Thesis, Faculty of Health Science, University of Copenhagen, 2001.
[The data are available as dataset N2001 at http://www.emglab.net]
[12] Gokgoz E, Subasi A. Effect of multiscale PCA de-noising on EMG
signal classification for diagnosis of neuromuscular disorders. J Med
Syst. 2014 Apr;38(4):31. doi: 10.1007/s10916-014-0031-3. Epub 2014
Apr 3. PMID: 24696395.
[13] Grossmann, A. and J. Morlet, “Decomposition of Hardy functions into
square integrable wavelets of constant shape”, SIAM J. Anal. 15, pp.
723-736, 1984.
[14] Subasi, Abdulhamit, and Emine Yaman. ”EMG signal classification
using discrete wavelet transform and rotation forest.” International
Conference on Medical and Biological Engineering. Springer, Cham,
2019.
[15] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classi-
fication with Deep Convolutional Neural Networks. Advances in Neural
Information Processing Systems 25.
[16] Ioffe, S., & Szegedy, C. (2015). Batch Normalization: Accelerating Deep
Network Training by Reducing Internal Covariate Shift. Proceedings
of the 32nd International Conference on International Conference on
Machine Learning, (pp. 448–456).
[17] Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdi-
nov, R. (2014). Dropout: a simple way to prevent neural networks from
overfitting. The Journal of Machine Learning Research, 1929–1958.
[18] Abadi, Martın, et al. ”TensorFlow: Large-scale machine learning on
heterogeneous systems, software available from tensorflow.org (2015).”
URL https://www.tensorflow.org (2015).
[19] Chollet, F. & others, 2015. Keras. Available at:
https://github.com/fchollet/keras.