International Journal of Signal Processing Systems Vol. 10, No.

1, March 2022

A Review on EMG Signal Classification and

Applications
Evon Lim Wan Ting and Almon Chai
Swinburne University of Technology Sarawak Campus, Kuching, Malaysia
Email: [email protected], [email protected]

Lim Phei Chin

Universiti Malaysia Sarawak, Kuching, Malaysia
Email: [email protected]

Abstract—Electromyography (EMG) signals are muscles the muscle [1]. A report regarding recommendations on
signals that enable the identification of human movements surface EMG signals acquisition with the title “Surface
without the need of complex human kinematics calculations. EMG for Non-Invasive Assessment of Muscles
Researchers prefer EMG signals as input signals to control (SENIAM)” is published and widely used by researchers
prosthetic arms and exoskeleton robots. However, the
proper algorithm to classify human movements from raw
[2]. The report provides recommendations regarding the
EMG signals has been an interesting and challenging topic type, shape, size, materials of electrodes, and also skin
to researchers. Various studies have been carried out to preparation, inter-electrode distance and the placement of
produce EMG-based human movement classification that electrodes.
gives high accuracy and high reliability. In this paper, the EMG signals are weak and contaminated with noises.
methods used in EMG signal acquisition and processing are Noises are present in the signals even during the
reviewed. The different types of feature extraction acquisition stage. Examples of noises that pollute the
techniques preferred by researchers are also discussed, EMG signals are ambient noise, inherent noise, motion
including some combination and comparison of feature artifact and inherent instability [3]. The presence of
extraction techniques. This paper also reviews the different
types of classifiers favored by researchers to recognize
noises affects the analysis of the EMG signals, and will
human movements based on EMG signals. The current have an impact on the accuracy in the classification of
applications of EMG signals are also reviewed. human movements. Proper signal amplification,
processing and filtering are required before further
Index Terms—classification, electromyography, feature analysis of the EMG signals can be carried out for pattern
extraction, human movement classification. Therefore, a lot of research studies has
been carried out on topics regarding EMG signal
processing, filtering and analysis that can lead to high
I. INTRODUCTION accuracy in human movement classifications.
When muscles are contracted, electrical currents are The focus of this paper will be on the reviewing of the
generated. These currents are known as different kinds of methodologies regarding EMG signal
Electromyography (EMG) signals. The evaluation of acquisition, processing, feature extracting and
EMG signals allows analysis of neuromuscular activities, classification that are preferred by researchers over the
without the need of complex human kinematics past years. Besides, a review regarding the current
calculations. Hence, EMG signals are widely used by applications of EMG signals is also discussed.
researchers to study human motions or to analyse
muscular disorders. II. EMG SIGNAL ACQUISITION
Over the years, EMG-based human movements The placement of electrodes during EMG signals
classification had become an interesting and challenging acquisition is an interesting topic. Two common types of
topic to researchers. EMG signals can be collected from electrode placement techniques are observed in research
muscles via electrodes. There are two ways to acquire studies: dense sampling approach and precise anatomical
EMG signals from human muscles: (i) needle electrodes positioning approach. For dense sampling approach, no
where EMG signals are acquired invasively and (ii) specific muscle location is pointed out. Instead,
surface electrodes where EMG signals are acquired non- electrodes are equivalently placed around the limb. For
invasively. The non-invasive way is more preferred by precise anatomical positioning approach, electrodes are
researchers. The invasive way requires advice and positioned precisely at the main activity spot of those
guidance from professionals, and could be painful and chosen muscles. The muscles are usually selected based
uncomfortable since the needle needs to be inserted into on the movements of interest in which the research study
aims to classify.
Manuscript submitted October 25, 2021; revised December 7, 2021.

©2022 Int. J. Sig. Process. Syst. 1

doi: 10.18178/ijsps.10.1.1-6
 International Journal of Signal Processing Systems Vol. 10, No. 1, March 2022

It is observed that, for EMG-based classification common methods to remove noises from EMG signals is
related to hand gestures, researchers mostly preferred to through signal amplification and filtering. However,
use the dense sampling approach for EMG signal proper care should be taken into consideration during the
acquisition. In [4]-[6], the teams used an eight electrodes amplification and filtering process to reduce signal
MYO armband for EMG signals acquisition. The MYO distortion in order to prevent the removal of useful
armband is placed at the forearm to collect EMG signals information from the signal.
for the classification of hand gestures. In [4], 37 subjects Researchers often prefer to use an instrumentation
participated in this study regarding the classification of amplifier with a large Common Mode Rejection Ratio
seven varieties of hand gestures. The best classification (CMRR) to remove the background noises from EMG
accuracy achieved is 93% via k-Nearest Neighbor (kNN). signals. These background noises are common mode
The research study presented by [5] managed to classify signals that reach the electrodes simultaneously and
three types of hand gestures with a recognition accuracy hence can be rejected via instrumentation amplifier.
that is greater than 90%. In [6], the EMG signals are Although the elimination of common mode noises is
collected from 21 subjects to analyse the classification of performed at the amplification stage, EMG signals can
one hand gesture. The average classification result is also be polluted with noises caused by motion artifacts,
around 99%. The research study carried out by [7] power lines as well as those at the electrode-skin junction.
acquired EMG signals through electrodes located at the Therefore, signal filtering is required. The frequencies of
lower elbow. The EMG signals collected are used to these noises are lower than 20 Hz. Hence, a band-pass
classify two hand gestures: grasp and release. The filter to remove frequency components that are lower
classification accuracy reported by the team is around than 20 Hz and higher than 500 Hz are recommended
92% to 94%. [17], [18]. However, it is also discovered that several
For studies related to movement of a specific finger or research studies proposed band-pass filter with different
the joint of the upper limb, it is spotted that the precise band-pass frequencies. The use of the notch filter to
anatomical positioning approach is more preferred by eliminate power line noise at 50 Hz is also commonly
researchers. In [8], EMG signals are collected from the seen in research studies [14], [19].
extensor digitorum communis to study the classification The study conducted to analyse the capability of
regarding the extension of three fingers: index, middle Support Vector Machine (SVM) and k-Nearest Neighbor
and ring. The classification rate reported in the study is (kNN) to recognize the three types of hand gestures
more than 90%. In [9], EMG signals are gathered from utilized a 15 Hz to 500 Hz band-pass filter and a notch
the biceps to study the classification of elbow joint. The filter for signal filtering at 50 Hz [14]. The authors of
highest classification accuracy achieved by the team is [20]-[22] applied a 10 Hz to 500 Hz band-pass filter for
around 96.4%. Similarly, the authors of [10] acquired the removal of unwanted noises from the EMG signals.
EMG signals from the biceps to evaluate the The study on angle estimation of wrist movement via
classification of elbow joint flexion to control a prosthetic EMG signals described in [11] utilized the Bessel type
arm. In another study regarding the classification of wrist filter in the EMG module to remove noises that are
joint movements, the authors selected the flexor carpi outside the ranges of 25 Hz and 500 Hz before
radialis, flexor carpi ulnaris, extensor carpi ulnaris and rectification is performed to the signal. In [23], an EMG
extensor long thumb muscles for the classification of sensor is developed, where a band pass filter is utilized to
wrist flexion, extension, abduction and adduction [11]. In filter unwanted signals outside the range of 90Hz to 450
[12], palmaris longus and extensor digitorum are chosen Hz. Besides, a notch filter is also applied to reject
to classify the flexion of the thumb, pointer finger, middle common mode noise.
finger and hand grasp. The highest classification rate
achieved in the study is around 92.64%. However, there IV. FEATURE EXTRACTION AND CLASSIFICATION
are also cases where the precise anatomical positioning
approach is used to acquire EMG signals for gesture or The extraction of features is an important step in EMG
posture based classification. In [13], the flexor carpi signal classification. Hence, it is important to select the
correct method to extract useful information from the
radialis muscle is selected for EMG signal acquisition for
processed EMG signals so as to increase the classification
the classification of eight hand postures. The
classification accuracy achieved is around 81.2%. In accuracy of EMG signals. There are numerous kinds of
another study, EMG signals are collected from the flexor feature extraction techniques that have been used or
proposed by researchers over the years. These feature
carpi ulnaris and extensor carpi radialis, longus and
extraction methods that are widely preferred by
brevis to classify three hand gestures [14].
researchers are sorted into three categories: time domain,
frequency domain as well as both time and frequency
III. EMG SIGNAL CONDITIONING AND PROCESSING
domain.
Raw EMG signal is naturally weak and contaminated Although the statistical properties of EMG signal are
with noise. Noises exist even during the signal acquisition always changing over the time, time domain features are
stage. Raw EMG signal has an amplitude that is around 0 still more preferred by researchers. This is because when
to 10 mV [15]. EMG signal has frequency that is between compared to frequency domain features, computation of
10 and 500 Hz. According to [16], the main energy of time domain features are less complex [24]. Among the
EMG signal stands between 50 and 150 Hz. One of the well-known types of time domain extraction methods that

©2022 Int. J. Sig. Process. Syst. 2

 International Journal of Signal Processing Systems Vol. 10, No. 1, March 2022

are preferred by researchers are Mean Absolute Value extracted four features (WL, SSC, MAV and ZC) from
(MAV), Simple Square Integral (SSI), Root Mean Square the EMG signals of 13 subjects. The team managed to
(RMS), Waveform Length (WL), Zero Crossing (ZC), achieve 80% and 94% accuracy from the classification of
Auto-Regressive Coefficients (AR), Mean Absolute WL and MAV features via kNN for offline and online
Value Slope (MAVS), Integrated Absolute Value (IAV), experiment respectively. However, the classification
Variance (VAR), Signal Length (SL), Difference of Mean result of this study is subject-specific as the EMG
Absolute Value (DMAV) and Integrated EMG (IEMG). samples utilized in both training and testing are from the
Fast Fourier Transform (FFT) is one type of technique same subject.
that converts EMG signals into the frequency domain Besides, there are also research studies that utilized
[25]. Features of the frequency domain have been multiple features for classification, but at the same time,
commonly obtained according to the statistical utilized Principal Component Analysis (PCA) to lessen
parameters of the Power Spectral Density (PSD). the feature to improve recognition rate. [13] extracted 16
Examples of frequency domain features are Mean time domain features from the EMG data collected from
Frequency (MNF), Median Frequency (MDF), Peak 15 subjects, aimed to classify eight types of hand postures.
Frequency (PKF), Mean Power (MNP), Power Spectrum PCA is applied to reduce the 16 features into three
Ratio (PSR) and Frequency Ratio (FR). principal components. The team managed to achieve
Human movements can be classified via recognition of 81.2% classification rate via ANN. [5] presented a
EMG patterns that are formed from features extracted sensor-assisted EMG data acquiring system that aimed to
from processed EMG signals. Most of the EMG based classify three hand gestures. A total of 48 features are
research studies preferred classifiers that can be trained produced during signal processing. Similarly, PCA is
with sample patterns. Among the common classifiers applied to select 14 features for classification. The
preferred by researchers are k-Nearest Neighbor (kNN), maximum average classification achieved via SVM is
Artificial Neural Network (ANN), Convolutional Neural more than 90%. The authors in [31] assessed the
Network (CNN), Deep Neural Network (DNN), and performance of two classifiers, namely Linear
Support Vector Machine (SVM). Discriminant Analysis (LDA) and Multinomial Logistic
In [10], a trans-humeral patient has successfully Regression (MLR) to recognize eight types of upper limb
controlled the flexion of the elbow joint of a prosthetic movements that are related to the shoulder joint. A total
arm based on RMS features extracted according to the of 26 time domain features and 10 frequency domain
EMG data collected via the biceps muscle. [26] combined features are extracted from the EMG signals to form a
the AR model and ANN in their methodology to classify feature vector with a dimension of 288. Then, PCA is
six kinds of finger movements. EMG signals are collected performed for feature reduction. It is reported that the
from four subjects. The classification rate achieved by the LDA and MLR managed to achieve classification
proposed system is more than 77% for each subject. [27] accuracy of 88.8% and 91.8% respectively.
used the wavelet transform technique to extract features In [20], various features are combined into groups to
from EMG signals to classify five finger movements via a identify the best combination of feature groups that can
DNN. The proposed method led to an average of 96.7% enhance the performance of the EMG based classifier to
recognition rate. In another study, valuable features are manipulate the prosthetic hand in real-time. Five feature
obtained through EMG data collected from four subjects groups are formed from nine time domain features to
via Discrete Wavelet Transform (DWT) to classify three classify six hand movements. The team concluded that
kinds of hand movements. The average classification rate the feature group with MAV, SSC, ZC and WL is the best
of the experiment is reported to be 93.25% [28]. The combination when classification is done with Simple
study carried out by [21] applied ANN to classify six Logistic Regression (SLR) classifier, where 91.1% of
types of forearm movements based on EMG feature accuracy is achieved for healthy subjects and 73.2% is
patterns extracted from Hilbert Huang Transform. achieved for trans-radial amputees.
Subject-specific ANN are trained and tested individually Comparison of classifiers to identify the optimal
using EMG samples from the same subject, collected classifier is also one of the interesting areas in EMG
from both the right and left hand. The overall average signal classification. In [9], both kNN and SVM classifier
accuracy of the EMG data acquired from the right hand are used to distinguish the flexion of the elbow joint. Four
(86.2%) is slightly higher in comparison to the left hand time domain features are obtained from the EMG signals
(85.8%). Nevertheless, no evaluation of data among any collected from 10 subjects. The capability of the kNN
individual has been carried out in the study. (96.4%) is better in comparison to the SVM classifier
There are also various studies that utilized multiple (85%). Similarly, [4] analysed the performance of the
time domain features for EMG classification. The study kNN and SVM classifier in recognizing seven hand
conducted by [29] applied five types of time domain gestures of 37 participants. Both classifiers are used to
features, namely MAV, SSC, RMS, SL and ZC, to classify feature patterns that are concatenated into a
classify the flexion and extension of the elbow joint, feature matrix. The classification accuracy of the kNN
pronation and supination. The authors managed to and SVM classifier is 93% and 83% respectively.
classify the four movements of eight subjects with an However, the study conducted by [14] showed that the
average accuracy of 91.78% via ANN. The study recognition ability of SVM is better than kNN. Seven
conducted by [30] described the control of a bionic hand EMG features collected from five subjects while
via four types of hand poses identification. The team performing three hand gestures are obtained via the time

©2022 Int. J. Sig. Process. Syst. 3

 International Journal of Signal Processing Systems Vol. 10, No. 1, March 2022

domain approach. The EMG features are analysed of the ETS-MARSE monitors the EMG signals of the
separately. The SVM classified all EMG features better wearer continuously during the therapy and offers help if
than the kNN. the controller detects that the wearer is incapable of
In [32], the authors compared four types of classifier performing the necessary motion. The study conducted
and four types of feature extraction techniques to classify by [34] not only utilized EMG signals to control a robotic
five hand movements via EMG signals acquired from six arm, but also implemented an Internet-of-Things system
subjects. Four time domain features are extracted from that enables real-time remote control applications such as
the raw EMG signals via MAV, RMS, variance and SSI switching on or off a light bulb, fan or electric heater via
techniques. These EMG features are classified using the EMG-based hand gestures recognition.
Support Vector Machine (SVM), Naïve Bayes (NB), k- Researchers have been using EMG signals in the
Nearest Neighbor (k-NN) and Random Forest (RF). From diagnosis of neuromuscular diseases, including myopathy,
the outcomes presented in the study, the combination of DPN disorder and amyotrophic lateral sclerosis diseases.
SVM with MAV and RF with RMS show the highest This method of diagnosis allows neuromuscular diseases
classification accuracy, which is around 98%. to be done automatically and more accurately as
The study conducted by [12] compared three types of compared to diagnosis through human eyes. In [35], the
classifier, namely ANN, Radial Basis Function (RBF) team proposed an automated diagnosis of myopathy
and Learning Vector Quantization (LVQ). The through EMG data recognition. EMG data obtained from
performances of these three classifiers are further studied the biceps brachii are analysed via various feature
in details. The ANN is tested with different number of extraction techniques and later classified via ANN
hidden neurons; the RBF classifier with Gaussian classifier. The method introduced by the team is able to
function is tested with different spread values; while the discriminate EMG signals among healthy subjects and
LVQ classifier is evaluated by varying the number of myopathy patients, with an encouraging result, where
competitive neurons. All three classifiers are trained to 87% classification accuracy is achieved. A method on
recognize five types of movements based on EMG EMG-based automated diagnosis of amyotrophic lateral
signals collected via two electrodes channels from two sclerosis diseases is proposed by [28]. The combination
selected muscles. The classification rate of the ANN with of continuous wavelet transform and SVM classifier is
10 hidden neurons is the highest, which is approximately used to distinguish EMG signals between healthy and
93%. The average classification accuracy achieved by the unhealthy subjects. The proposed method managed to
RBF classifier with 0.7 spread value is around 84%. The distinguish 93.75% of the samples correctly.
LVQ with 28 competitive neurons has the best Since the intention of human movements can be
recognition rate, where the classification rate achieved is determined by analyzing the EMG data gathered from the
89%. muscles, researchers have been proposing and
In the study of [8], an approach is introduced to introducing approaches regarding the manipulation of
improve the Back-Propagation (BP) algorithm of the human-assistive robotics, such as powered exoskeleton,
ANN for the recognition of three finger movements. A prosthetic arm and bionic hand, via EMG signals. In [36],
correction factor is assigned to the hidden layer to enlarge the team proposed the use of a 4-channel EMG signals to
the input sensitivity to speed up the time required for the distinguish six types of hand gestures for the actuation of
training to escape the local minimum point. The proposed prosthetic drive. The classification outcomes of the EMG
method is validated with EMG signals collected from one signals are used as control signals to trigger the DC
male and one female subject, where the AR model is used motors designed to mimic the intended movements. The
to extract features for classification. The experiment team of [10] reported the successful control of an
result shows that the classification rate attained via the artificial arm model developed to demonstrate elbow
improved BP neural network is greater than 90%, which flexion after interpreting the intention of the amputee via
is better than the classification rate of a normal BP neural EMG signal of the biceps muscle. An EMG-based upper-
network. limb exoskeleton for power assist is developed in [37].
The team constructed an EMG-angle model via BPNN to
V. APPLICATIONS OF EMG SIGNALS recognize RMS features extracted from the EMG signals
collected from four muscles. The results achieved by the
Over the years, EMG signals have been applied in team indicated that predicting the intention of the wearer
various applications, such as rehabilitation therapy, through EMG signal analysis could manipulate the
power-assist exoskeleton and prosthetic hand control, exoskeleton in real time.
robotic manipulator control, personal authentication, Researchers have also proposed a EMG-based personal
automated diagnosis of neuromuscular diseases as well as authentication. In [38], EMG signals are acquired from a
ergonomics studies. proposed 2-channel EMG module to detect a specific
In rehabilitation, robot or exoskeleton has been hand gesture. A total of 100 samples are collected from
introduced to help patients to carry out the required 10 subjects. Five feature extraction techniques are used to
movement therapy. During the treatment, these robots or extract useful patterns from the EMG samples. The usage
exoskeletons are designed to passively help the patients of an ANN to recognize the EMG patterns is proposed in
to move their limbs with the aid of actuators. In [33], a 7- the methodology. The outcomes obtained via the
DOF upper-limb exoskeleton, known as ETS-MARSE, is proposed method proved the feasibility of EMG-based
designed for passive rehabilitation therapy. The controller

©2022 Int. J. Sig. Process. Syst. 4

 International Journal of Signal Processing Systems Vol. 10, No. 1, March 2022

personal authentication, where a 95% recognition rate is ACKNOWLEDGMENT

achieved.
The authors would like to extend appreciation and
EMG signals are also used in ergonomics study. The
support to the School of Research, Swinburne University
authors in [39] conducted a study on the evaluation of
of Technology Sarawak Campus.
occupational injuries based on EMG signals. Ten subjects
are invited to participate in the study. The EMG signals
from the biceps brachii are collected from the participants REFERENCES
while performing a specific task that is suspected to cause [1] S. A. Raurale, “Acquisition and processing real-time EMG signals
Musculoskeletal Disorder (MSD). The team concluded for prosthesis active hand movements,” in Proc. Int. Conf. Green
Comput. Commun. Electr. Eng., 2014, pp. 1-6.
from their study that the frequency domain features [2] H. J. Hermens, B. Freriks, R. Merletti, G. G. Hagg, D. Stegeman, J.
obtained from the EMG data are able to give more Blok, G. Rau, and C. Disselhorst-Klug, “European
information, but time domain features perform better in recommendations for surface electromyography,” Roessingh
classification. In another study, the posture comfort of the Research and Development, vol. 8, pp. 13-54, 1999.
[3] M. B. I. Reaz, M. S. Hussain, and F. Mohd-Yasin, “Techniques of
firefighter is assessed via EMG approach [40]. EMG EMG signal analysis: Detection, processing, classification and
signals are acquired while the firefighter performs certain applications,” Biol. Proced. Online, vol. 8, no. 1, pp. 11-35, 2006.
postures that are common during firefighting activities to [4] A. Devaraj and A. K. Nair, “Hand gesture signal classification
evaluate the comfort level of the postures. The research using machine learning,” in Proc. Int. Conf. Commun. Signal
results can act as a reference on operating posture for Process., 2020, pp. 0390-0394.
[5] M. Kunapipat, P. Phukpattaranont, P. Neranon, and K. Thongpull,
firefighter to follow and practice to avoid discomfort “Sensor-assisted EMG data recording system,” in Proc. 15th Int.
during firefighting operations. Conf. Electr. Eng./Electron. Comput. Telecommun. Inf. Technol.,
2018, pp. 772-775.
[6] L. Lu, J. Mao, W. Wang, G. Ding, and Z. Zhang, “A study of
VI. CONCLUSION personal recognition method based on EMG signal,” IEEE Trans.
The classification of EMG signals to predict human Biomed. Circuits Syst., vol. 14, no. 4, pp. 681-691, 2020.
[7] L. S. Praveen, S. N. Nagananda, and P. Shankapal, “Design and
movements allows the control of the prosthetic arm and development of real time bionic hand control using EMG signal,”
exoskeleton robot to be done based on the intention of the in Proc. IEEE Int. Conf. Electr. Comput. Commun. Technol., 2018,
human operator. Current research studies focused more pp. 1-4.
on selecting the appropriate feature extraction techniques [8] F. Li, Y. Zhang, and K. Gao, “Pattern recognition of finger
motion's EMG signal based on improved BP neural networks,” in
and classifiers to improve classification accuracy. Proc. Int. Conf. Comput. Sci. Netw. Technol., 2011, pp. 1266-1269.
Various research studies have proposed the use of more [9] I. M. Khairuddin, S. N. Sidek, A. P. P. A. Majeed, and A. A. Puzi,
than one feature extraction techniques. Researchers have “Classifying motion intention from EMG signal: A k-NN
also been finding appropriate combination of feature approach,” in Proc. 7th Int. Conf. Mechatron. Eng., 2019, pp. 1-4.
[10] R. Ahmed, R. Halder, M. Uddin, P. C. Mondal, and A. K.
extraction approaches that could further enhance the Karmaker, “Prosthetic arm control using electromyography
classification accuracy. The types of classifiers that are signal,” in Proc. Int. Conf. Adv. Electr. Electron. Eng., 2018, pp.
most preferred by researchers are kNN, SVM and ANN. 1-4.
Comparison between the performances of different [11] A. F. Flores, G. A. Verdezoto, L. M. Tobar, I. Iglesias, and S. G.
Guerra, “Angle estimation of wrist movements through surface
classifiers has been carried out in several studies. EMG signals,” in Proc. IEEE Fourth Ecuador Technical Chapters
From the studies reviewed, most of the studies only Meeting (ETCM), 2019, pp. 1-6.
focus on classifying small amount of human movements. [12] G. Tsenov, A. H. Zeghbib, F. Palis, N. Shoylev, and V. Mladenov,
The number of subjects participating in the studies is also “Neural networks for online classification of hand and finger
small. The number of EMG samples should be increased movements using surface EMG signals,” in Proc. 8th Seminar on
Neural Network Applications in Electrical Engineering, 2006, pp.
to enlarge the database that allows more comparison 167-171.
between subjects to be done so as to upsurge the [13] W. Caesarendra, S. U. Lekson, K. A. Mustaqim, A. R. Winoto,
reliability of the research outcomes. There are also and A. Widyotriatmo, “A classification method of hand EMG
studies without inter-subject comparison. Although the signals based on principal component analysis and artificial neural
network,” in Proc. Int. Conf. Instrum. Control Autom., 2016, pp.
accuracy achieved is high, the suitability of the proposed 22-27.
algorithm for generic applications cannot be justified. [14] Y. Paul, V. Goyal, and R. A. Jaswal, “Comparative analysis
Besides, the EMG data used in most research studies are between SVM & KNN classifier for EMG signal classification on
not published. Researchers have proposed different elementary time domain features,” in Proc. 4th Int. Conf. Signal
Process. Comput. Control, 2017, pp. 169-175.
methods, but these methods cannot be compared since the [15] C. J. D. Luca, “Surface electromygraphy: Detection and
EMG data used in their studies are different. The amount recording”, DelSys Incorporated, vol. 10, no. 2, pp. 1-10, 2002.
of EMG data used in each study should be enlarged to [16] J. Wang, L. Tang, and J. E. Bronlund, “Surface EMG signal
increase the reliability of a proposed algorithm. amplification and filtering,” Int. J. Comput. Appl., vol. 82, no. 1,
pp. 15-22, 2013.
[17] S. Ullah and K. Iqbal, “A preliminary review on EMG signals for
CONFLICT OF INTEREST assessment of diabetic peripheral neuropathy disorder,” in Proc.
7th Int. Conf. Electr. Electron. Eng., 2020, pp. 42-46, 2020.
The authors declare no conflict of interest. [18] M. R. Ahsan, M. I. Ibrahimy, and O. O. Khalifa, “Design and
optimization of Levenberg-Marquardt based neural network
AUTHOR CONTRIBUTIONS classifier for EMG signals to identify hand motions,” Meas. Sci.
Rev., vol. 13, no. 3, pp. 142-151, 2013.
All authors have contributed equally to this work; all [19] A. Saikia, N. M. Kakoty, N. Phukan, M. Balakrishnan, N. Sahai, S.
authors had approved the final version. Paul, and D. Bhatia, “Combination of EMG features and stability

©2022 Int. J. Sig. Process. Syst. 5

 International Journal of Signal Processing Systems Vol. 10, No. 1, March 2022

index for finger movements recognition,” Procedia Comput. Sci., [36] P. Geethanjali, K. K. Ray, and P. V. Shanmuganathan, “Actuation
vol. 133, pp. 92-98, 2018. of prosthetic drive using EMG signal,” in Proc. IEEE Region 10
[20] G. Purushothaman and K. K. Ray, “EMG based man–machine Conf., 2009, pp. 1-5.
interaction — A pattern recognition research platform,” Rob. [37] Z. Tang, K. Zhang, S. Sun, Z. Gao, L. Zhang, and Z. Yang, “An
Auton. Syst., vol. 62, no. 6, pp. 864-870, 2014. upper-limb power-assist exoskeleton using proportional
[21] A. Kukker, R. Sharma, and H. Malik, “Forearm movements myoelectric control,” Sensors (Basel, Switzerland), vol. 14, no. 4,
classification of EMG signals using Hilbert Huang transform and pp. 6677-6694, 2014.
artificial neural networks,” in Proc. IEEE 7th Power India Int. [38] S. Shin, J. Jung, M. Kang, and Y. T. Kim, “A Study on EMG
Conf. (PIICON), 2016, pp. 1-6. signal acquisition modules and artificial neural networks for
[22] R. Jia, L. Yang, Y. Li, and Z. Xin, “Gestures recognition of sEMG personal authentication,” in Proc. 2018 Int. Conf. Comput. Sci.
signal based on Random Forest,” in Proc. 2021 IEEE 16th Conf. Comput. Intell., 2018, pp. 1458-1459.
Ind. Electron. Appl. (ICIEA), 2021, pp. 1673-1678. [39] T. N. S. T. Zawawi, A. R. Abudllah, W. T. Jin, R. Sudirman, and
[23] A. Marinelli, et al., “Miniature EMG sensors for prosthetic N. M. Saad, “Electromyography signal analysis using time and
applications,” in Proc. 2021 10th Int. IEEE/EMBS Conf. Neural frequency domain for health screening system task,” Int. J. Hum.
Eng. (NER), 2021, pp. 1022-1025. Technol. Interact., vol. 2, no. 1, pp. 35-43, 2018.
[24] A. Phinyomark, P. Phukpattaranont, and C. Limsakul, “Feature [40] L. Yang, B. Kang, T. Wang, and T. Zhao, “An ergonomic study of
reduction and selection for EMG signal classification,” Expert Syst. firefighters’ postural comfort evaluation based on EMG method,”
Appl., vol. 39, no. 8, pp. 7420-7431, 2012. in Proc. Hum. Factors Ergon. Soc. Annu. Meet, vol. 58, no. 1, pp.
[25] J. Too, A. R. Abdullah, T. N. S. T. Zawawi, N. M. Saad, and H. 2310-2314, 2014.
Musa, “Classification of EMG signal based on time domain and
frequency domain features,” Int. J. Hum. Techol. Interact., vol. 1, Copyright © 2022 by the authors. This is an open access article
no. 1, pp. 25-29, 2017. distributed under the Creative Commons Attribution License (CC BY-
[26] J. Z. Wang, R. C. Wang, F. Li, M. W. Jiang, and D. W. Jin, “EMG NC-ND 4.0), which permits use, distribution and reproduction in any
signal classification for myoelectric teleoperating a dexterous medium, provided that the article is properly cited, the use is non-
robot hand,” in Proc. IEEE Eng. Med. Biol. 27th Annual Conf., commercial and no modifications or adaptations are made.
2005, pp. 5931-5933.
[27] J. Ahmad, A. M. Butt, M. Hussain, M. A. Akbar, and W. U. Evon Lim Wan Ting received the Bachelor
Rehman, “The deep neural network based classification of fingers of Engineering in Robotics and Mechatronics
pattern using electromyography,” in Proc. IEEE Adv. Inf. Manag. from Swinburne University of Technology,
Commun. Electron. Autom. Control Conf., 2018, pp. 455-461. Sarawak Campus, in 2014 and the Ph. D.
[28] N. Sengar, M. K. Dutta, and R. Burget, “Detection of neuro degree in Engineering from Swinburne
mascular disease using EMG signals in wavelet domain,” in Proc. University of Technology in 2020. She is
4th IEEE Uttar Pradesh Sect. Int. Conf. Electr. Comput. Electron., currently a post-doctoral research fellow with
2017, pp. 624-627. the School of Research, Swinburne University
[29] J. Wang and W. Wichakool, “Artificial elbow joint classification of Technology Sarawak Campus. Her
using upper arm based on surface-EMG signal,” in Proc. IEEE research interest includes the analytics and
3rd Int. Conf. Eng. Technol. Social Sci., 2017, pp. 1-4. optimization on classification and application
[30] W. T. Shi, Z. J. Lyu, S. T. Tang, T. L. Chia, and C. Y. Yang, “A of EMG signals related to movement control and manipulation of
bionic hand controlled by hand gesture recognition based on exoskeleton.
surface EMG signals: A preliminary study,” Biocybern. Biomed.
Eng., vol. 38, no. 1, pp. 126-135, 2018. Almon Chai received his bachelor’s degree in
[31] A. Tigrini, L. A. Pettinari, F. Verdini, S. Fioretti, and A. Electrical and Electronics engineering in 2000
Mengarelli, “Shoulder motion intention detection through and PhD in Manufacturing specializing in
myoelectric pattern recognition,” IEEE Sens. Lett., vol. 5, no. 8, thermofluids in 2007. He is currently an
pp. 1-4, Aug. 2021. academic with Faculty of Engineering,
[32] B. Kundu and D. Subarram Naidu, “Classification and feature Computing and Science at Swinburne
extraction of different hand movements from EMG signal using University of Technology Sarawak Campus.
machine leaning based algorithms,” in Proc. 2021 Int. Conf. Electr His research interest included computational
Commun. Comput. Eng. (ICECCE), 2021, pp. 1-5. techniques, simulation, human factors and
[33] M. H. Rahman, M. J. Rahman, M. Saad, C. Ochoa-Luna, S. B. ergonomics.
Ferrer, and P. S. Archambault, “Control of an upper extremity
exoskeleton robot to provide active assistive therapy,” in Proc. 5th Lim Phei Chin is a lecturer at the Faculty of
Int. Conf. Model. Identif. Control, 2013, pp. 105-110. Computer Science and Information
[34] M. Nguyen, T. N. Gia, and T. Westerlund, “EMG-based IoT Technology, Universiti Malaysia Sarawak.
system using hand gestures for remote control applications,” in She has been involved in various research
Proc. 2021 IEEE 7th World Forum on Internet of Things (WF- projects since completing her PHD in Text
IoT), 2021, pp. 911-912. and Data Mining in 2016. Her research
[35] A. Singh, M. K. Dutta, and C. M. Travieso, “Analysis of EMG interests are in image processing, statistical
signals for automated diagnosis of myopathy,” in Proc. 4th IEEE analysis, machine learning and recently
Uttar Pradesh Sect. Int. Conf. Electr. Comput. Electron. Eng., ventured into web security.
2017, pp. 628-631.

©2022 Int. J. Sig. Process. Syst. 6