IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 4, December 2023, pp. 1735~1743

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i4.pp1735-1743  1735

Comparative analysis of machine learning algorithms on

myoelectric signal from intact and transradial amputated limbs

Dhirgaam A. Kadhim, Mithaq N. Raheema, Jabbar S. Hussein

Department of Prosthetic & Orthotic Engineering, Faculty of Engineering, Kerbala University, Kerbala, Iraq

Article Info ABSTRACT

Article history: Control strategies of smart hand prosthesis-based myoelectric signals in
recent years don't provide the patients with the sensation of biological
Received Aug 31, 2022 control of prostheses hand fingers. Therefore, in current work
Revised Jan 14, 2023 hyperparameters optimization in machine learning algorithm and hand
Accepted Feb 9, 2023 gesture recognition techniques were applied to the myoelectric signal-based
on residual muscles contraction of the amputees corresponding to intact
forearm limb movement to improve their biological control. In this paper,
Keywords: myoelectric signals are extracted using the MYO armband to recognize ten
gestures from ten volunteers (healthy and transradial amputation) on the
Comparative analysis forearm, thereafter the noise of myoelectric signals using a notch filter (NF)
Hand gestures is removed. The proposed classification system involved two machine
Hyperparameters optimization learning algorithms: (1) the decision tree (DT), tri-layered neural network
Machine learning (TLNN), k-nearest-neighbor (KNN), support vector machine (SVM) and
Myoelectric armband ensemble boosted tree (EBT) classifiers. (2) the optimized machine learning
classifiers, i.e., OKNN, OSVM, OEBT with optical diffraction tomography
(ODT) and ommatidia detecting algorithm (ODA). The experimental results
of classifiers comparison pointed out an algorithm that outperformed with
high accuracy is OEBT closely followed by OKNN achieves an accuracy of
97.8% and 97.1% for intact forearm limb, while for transradial amputation
with an accuracy of 91.9% and 91.4%, respectively.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Dhirgaam A. Kadhim
Department of Prosthetic & Orthotic Engineering, Faculty of Engineering, Kerbala University
Kerbala, Iraq
Email: [email protected]

1. INTRODUCTION
The bundles of specialized cells that constitute muscles are able to contract and relax. In the human
body, a bioelectrical signal corresponding to muscle activity is produced during muscular contraction. The
signal is known as Electromyography (EMG), and it is crucial for clinical diagnosis, treatment, rehabilitation,
and automatic control of smart prosthetics and orthotics [1]. Therefore, it is important to study, analyze and
classify the EMG signals then compared between them. Analysis of EMG signals using appropriate methods
and algorithms can give efficient ways for understanding of their characteristics, and the EMG devices
implementations can be made for various EMG signal related applications. The surgical resection of all or
segments of an upper-lower limb, i.e., arm, hand, finger, leg, foot, or toe, is known as an amputation [2].
Amputations of the upper limbs range from losing only a finger part to losing an entire arm and part of the
shoulder, furthermore, a transradial or below elbow amputation impacts hand function by losing part of the
forearm below the elbow joint Morgan et al. [3]. EMG signals can be taken from the anterior and posterior
muscles at Forearm. The anterior and posterior muscles of the forearm can generate EMG signals
McCausland et al. [4]. Francesco Redi's reporting of the EMG dates back to 1666 [5], whereas Marey made

Journal homepage: http://ijai.iaescore.com

1736  ISSN: 2252-8938

the first recording of EMG electrical activity in 1890. Early in the 1980s, Cram and Steger developed a
clinical technique for using an EMG measuring device to scan variety muscles [6]. Then, the scientific
research in this field improved our knowledge of the characteristics of surface EMG recording.
Depolarization produces an electromagnetic field called EMG, which is quantified as very low voltage.
Motor neuron action potentials are propagated to the tendons after being transmitted to muscles by end-
plates. Surface electrodes are used to detect these bioelectrical signals [7]. Disease and injury to the lower
motoneuron or the muscle may result in muscular atrophy and partial paralysis of variable degrees [8]. There
are two EMG types, (1) Clinical EMG, also known as "diagnostic EMG", and (2) "Nerve Conduction EMG,"
which is frequently performed by physiatrists and neurologists, while kinesiological EMG is used in
movement analysis [9]. The MYO armband is a device that has eight sensors to measure the EMG signals of
forearm muscles at a frequency of 200 Hz (sample rate) and are normalized to values in the range of 1 and -1
as a result of muscle activation, sending them to computer via Bluetooth [10]. The MYO armband is also
equipped with several other sensors, such as accelerometer, gyroscope, and magnetometer which collect
spatial data about the gesture and can facilitate gesture recognition [11]. MYO armband can be used to allow
an amputee to control a prosthetic, robotic hand, a humanoid robot, interaction for the local surgeon by
laparoscopic, in addition to other applications such as games, Maps, Myo music, and Keynotes. Feature
extraction methods can be used to infer the efficacy information in EMG signal to study its characteristic and
behavior in (time, frequency, or time-frequency domains). The time domain (TD) features are simple and
efficient in EMG pattern recognition [12]. TD features are extracted from the variations of signal amplitude
with time such as root mean square (RMS), mean absolute value (MAV) Zecca et al. [13], standard deviation
(STD), integrated absolute value (IAV), maximum amplitude (MAX), minimum amplitude (MIN), and many
other functions that are used in the time domain such as zero crossings (ZC), carrier operated relay (COR),
Wide-area multilateration (WAM), and man-portable vector (MPV), the EMG signals can be employed to
detection the dynamic activity of a muscle (contract and relax), whereas RMS or moving average (MA)
values are used to gauge the muscle's level of activity or gauge how much force it generates Liu et al. [14].
The optimization the hyper-parameters of machine learning algorithm with several techniques for difference
applications and the performances compared to develop, furthermore present the relationship among dataset,
performance and hyperparameters [15]–[17]. There are six machine learning classifiers, i.e., support vector
machine (SVM), random forest (RF) decision tree (DT), logistic regression (LR), naive bayes (NB) and
ridge classifier (RC) are comparison and different hyperparameters methods tuning are employed
Elgeldawi et al. [18]. Development the automatic machine learning to classification EMG signal that
extracted with needle electrode based on hyper-parameters optimization for healthy and disease muscles
Kefalas et al. [19]. The classification of EMG pick up from forearm using MYO armband and the training
dataset's 5-fold cross-validation methods are used to choose the hyperparamet ers for classifiers [20].
In current study, the primary challenge is to employ the wireless MYO armband instead of cabled
electrodes to collect myoelectric signal data from transradial amputation and a healthy forearm. This which
increase the freedom and comfort of movement for the forearm-hand of the subject. Furthermore, the low
cost of wearable MYO armband devices relative to other EMG equipment, thus, it may be an alternative to
expensive devices.
The contribution of this paper is to propose tuning the parameters, functions and optimization hyper-
parameters of the machine learning algorithms. Also, to recognize ten gestures (Clenched fist, fingers
abduction, wrist extension, wrist flexion, and rest), which helps in reducing the difficulty of classification
myoelectric signal data. Then the extracting during contraction muscles of transradial amputation resulting
from weak or atrophic muscles and give high-rate performance to move hand prosthesis.

2. MATERIAL AND METHOD

A significant improvement in the dexterity of control in upper-limb prosthetics has been
demonstrated by the intelligent classifier of EMG signal with the control method. In this study, the proposed
methodology comprises several components as block diagram illustrated in Figure 1. The methodology
procedures employ the MYO armband to pick up the myoelectric signal from transradial amputee and intact
forearm with different hand movements intention at the wrist joint. The myoelectric signal preprocessing to
reduce the noise associated with the myoelectric signal record, then windowing-overlapping data and features
extracting to keep the most discerning hand movement-related information possible. The classification of
data features collection using various machine learning algorithms with optimization to predict the best
model of limb movements, and send as commands to control the movement that drives the prosthesis. These
methodology stations are important in finding the total performance of the prosthetic system movement
patterns. Therefore, the details components of the methodology are described.

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1735-1743

Int J Artif Intell ISSN: 2252-8938  1737

Figure 1. Proposed methodology

2.1. Myoelectric data acquisitions

The current investigation involves the MYO armband as shown in Figure 2 to capture the
myoelectric signals Tavakoli et al. [21]. In this method, we use an external HM-10 Bluetooth module to
connect the MYO armband to a controller board. These sensors use a sampling rate of 200 Hz and an 8-bit
resolution for each of its eight myoelectric sensors to measure the electrical activity of the forearm muscles.
The objectives of preprocessing are to minimize the noise in the collected myoelectric signal and make
feature extraction easier. The raw signal has some additional noise, which might produce invalid features and
cause classification problems, Therefore, the notch filter technique that collects low and high pass filters was
implemented for removing the small area of bandwidth interference from the myoelectric signal [22]. The
observed signals are first normalized to prepare for training, with each member of each matrix signal
M=(M1, M2,...., M8) in the range [-1, 1]. An absolute value function is used to rectify the original signal in
each eight-channel.

Figure 2. MYO armband for data acquisitions

The volunteers were requested to maintain a consistent behavior and record the whole myoelectric
data signals as part of the study methodology. The myoelectric database often needs the participant to hold
each motion for a brief period of time before repeating it multiple times. The movements were alternated
with a resting stance to reduce fatigue. During frequent use, it is customary to keep a long posture. As a
result, in order to design a bionic prosthesis that is more suitable for frequent use by amputees, each gesture
necessitates the subject to sustain a relatively long duration, resulting in noise in the acquired signal owing to
tiredness and recuperation. In this paper, the myoelectric signal was extracted from five male subjects
participating in the experiments with transradial amputation (below the elbow) and intact forearm limb. Data
sets of ten gestures (Clenched fist, fingers abduction, wrist extension, wrist flexion, and rest) were recorded
for the amputated and healthy forearm of the same volunteer for this study, as illustrated in Figure 3(a) and

Comparative analysis of machine learning algorithms on myoelectric signal … (Dhirgaam A. Kadhim)

1738  ISSN: 2252-8938

Figure 3(b). Each of the ten gestures was repeated three times and a myoelectric signal was recorded for a
period of five seconds per gesture with a frequency of 200 Hz.

Figure 3. Experimental setup to record ten hand gestures: (a) intact forearm limb and (b) transradial
amputation limb

These ten gestures: i.e., clenched fist, fingers abduction, wrist extension, wrist flexion, and rest are
simulated with 1,000 elements per gesture as shown in Figure 4. Since the extracted myoelectric data was
found have different length sizes for each gesture in the range 1,171-1,339 samples, the silence removal (SR)
feature was used to delete extra data in order to obtain the same length size for all gesture data samples. One
of the main criteria for improving the accuracy classification of the myoelectric data signal is the window-
length based segmentation technique, therefore, the data sets signal is overlapped for feature extraction in this
work. There are four features of the myoelectric signal after the segmentation window to equal size were
extracted in the time domain and to reduce the time computational cost, the features selected are RMS, zero
crossings (ZC), integrated ElectroMyography (IEMG), and waveform length (WL) for each window length
and gesture [23]–[25].

Figure 4. Myoelectric signals for ten gestures from both intact hand and transradial amputation

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1735-1743

Int J Artif Intell ISSN: 2252-8938  1739

2.2. Gestures classifiers

In this work, two categories of the classifier are implemented: the first category consists of support
vector machines (SVM) [26], [27], k-nearest neighbors (KNN) [28], decision tree (DT) [29], ensemble
boosted tree (EBT) [30] and tri-layered neural network (TLNN) [31], Ye et al. [32]. These supervised
machine learning classifiers trained the features of myoelectric data. The second category involved optimized
classifiers i.e., optimal support vector machine (OSVM), optimal k-nearest neighbors (OKNN), optical
diffraction tomography (ODT) and ensemble optimal electronic benefits transfer (OEBT) by optimization of
the internal parameters of classifiers using the hyper-parameters model. Evaluation of the classification to
obtain a robust classifier model is adopted by implementing cross-validation value i.e. K-Fold=10.

3. EXPERIMENTAL RESULTS AND DISCUSSION

This paper implemented three different protocol of experiments procedure: the first is pre-setting the
parameters of the fine gaussian SVM, TLNN, fine KNN, fine DT and EBT classifiers before training with
different parameters and functions: i.e., kernel functions (Gaussian, linear, quadratic), The structure of TLNN
classifier included 32 neurons in input layer, three fully-connected layers have with three activation-functions
rectified linear unit (ReLU), Tanh, Sigmoid connected with each fully-connected layer and softmax function
is connected to the last fully-connected layer. Finally, the output layer that corresponds to the five gestures
labels. The non-optimized EBT classifier trained with three ensemble methods (AdaBoost, Bag, RUSBoost),
while the for fine KNN classifier used three metrics (Chebyshev, Euclidean, Cosine) to find the distance to
points. In fine DT classifier the split criterion measure (maximum deviance reduction, Gini’s diversity index,
twoing rule) are specified to deciding when to split nodes.
The second is setting the hyperparameters after training the machine learning classifiers, SVM,
KNN, EBT, and DT. It will obtaine automated machine-learning classifiers, which are: OSVM, OKNN, ODT
and OEBT. There are various parameters and functions for optimization used as present in Table 1. The third
experiment protocol represents the evaluation classifiers performance.

Table 1. Optimized hyper-parameters details of the SVM, KNN, EBT AND DT classifiers
Optimized classifier Selected parameters for optimization Optimized hyperparameters
OKNN Number of neighbors [1-8,000]
Distance metric Chebyshev Cosine Ecuildean
Distance weight Equal Inverse Squared inverse
standardize data TRUE FALSE
OSVM Kernel function Gaussian Cubic Linear
Box constraint level [0.001-1]
Kernel scale Auto
Multiclass method standardize data
OEBT Ensemble method AdaBoost Bag RUSBoost
Learner type Decision Tree
Maximum number of split [1-8,000]
Number of learners [5-1,000]
Learning rate [0.01-1]
Number of predictors to sample [1-10]
Optimizer type Bayesian optimization Grid search Random search
ODT Maximum number of splits [1-8,000]
Split criterion Maximum deviance Gini’s diversity Twoing rule
reduction index

Initially, the evaluation performance of the fine gaussian SVM, TLNN, fine KNN, fine DT and EBT
classifiers function are examined by measure the accuracy as shown in Table 2, respectively. From results the
performance of the Gaussian Kernel Function of Fine Gaussian SVM classifier is better than the others
function. Its accuracy is 84.1% for transradial amputation and 91.4% for intact forearm limb.
The results conducted that TLNN based Tanh activation function high accuracy than other function
for amputee and intact volunteers with an accuracy 84.2% and 94.6%, respectively. Overall, the results
indicated that the EBT classifier outperformed other classifiers with an average accuracy 95.4% for intact and
84.1% for amputee by using ensemble method Bootstrap Aggregation (Bagging) followed by classifier KNN
with an accuracy equals to 95.4% and 88.1%. On other hand, The DT classifier performance was the worst
with an accuracy rate of 90.3% for intact and SVM classifier based Quadratic Kernel Function for amputee
with an accuracy equal to 70.9%.
The results of classification accuracy, MCE and AUROC (area under the curve of convergence
receiver operating characteristic (ROC)), that examined by the proposed classifiers for intact and amputee

Comparative analysis of machine learning algorithms on myoelectric signal … (Dhirgaam A. Kadhim)

1740  ISSN: 2252-8938

myoelectric signal are listed in Table 3. From Table 3 it is obvious that OEBT performed better than other
classifiers (OKNN, OSVM, ODT) for amputee with an accuracy of 91.9%, followed by OKNN with 91.4%
and with high rate AUROC, and MCE performances of 0.99 and 0.081, respectively. While, the ODT
classifier, whose accuracy is 89.41% has the worst performance. Based on the results, the optimized
classifiers can satisfactorily perform better than those specified parameter classifiers.

Table 2. Comparison performance accuracy results-based functions and parameters classifiers models
Classifier Classifier Functions and Parameters
Kernel functions Gaussian Linear Quadratic
Fine Gaussian SVM Volunteer Intact Amputee Intact Amputee Intact Amputee
Accuracy (%) 91.4 84.1 87 71.9 82.3 70.9
Activation functions ReLU Tanh Sigmoid
TLNN Volunteer Intact Amputee Intact Amputee Intact Amputee
Accuracy (%) 90.7 83.4 94.6 84.2 91.9 83.6
Distance metric Chebyshev Euclidean Cosine
Fine KNN Volunteer Intact Amputee Intact Amputee Intact Amputee
Accuracy (%) 94.6 84.4 95.4 88.1 91.9 83.3
Split criterion Maximum deviance reduction Gini’s diversity index Twoing rule
Fine DT Volunteer Intact Amputee Intact Amputee Intact Amputee
Accuracy (%) 90.3 81.5 90.7 81.7 91.5 83.9
Ensemble methods AdaBoost Bag RUSBoost
EBT Volunteer Intact Amputee Intact Amputee Intact Amputee
Accuracy (%) 92.4 82.6 95.4 84.1 93.1 79.1

Table 3. Comparative performance of different optimized classifiers-based accuracy, MCE AND AUROC
Classification accuracy (%) MCE AUROC
Optimized Classifier
Intact Amputee Intact Amputee Intact Amputee
OKNN 97.1 91.4 0.029 0.086 0.99 0.99
OSVM 93.4 89.9 0.066 0.101 0.98 0.98
OEBT 97.8 91.9 0.022 0.081 0.99 0.99
ODT 92.1 89.4 0.079 0.106 0.97 0.96

According to an evaluation of the results in Tables 2 and 3 can be it can be deduced that the best
classifiers are OEBT and OKNN for transradial amputation and intact forearm cases. Therefore, both
classifiers: OEBT and OKNN are adopted in the rest of the results. The comparison of the minimum
classification error (MCE) convergence graph of selected optimization methods based on the OKNN and
OEBT as the best classifiers for transradial amputation are shown in Figure 5(a) and Figure 5(b) and for
intact forearm limbs are displayed in Figure 5(c) and Figure 5(d).
The MCE plot consists of two convergence graphs, the first represents the estimated MCE that is
indicated in light blue points color calculated by the optimization process taking into account all of the
hyperparameter value sets that have been tried up to this point, including the present iteration. While the
second convergence graph represents the observed MCE that is indicated in dark blue points color calculated
thus far through the optimization procedure. From Figure 5(b) it is observed that reaching the best point
hyperparameters at fourth iteration that less than iteration number in Figure 5(a), while in Figure 5(d) at the
eighth iteration. It is seen that attaining the best point hyperparameters that less than iteration number in
Figure 5(c).
Figure 6 shows the confusion matrix for both groups (transradial amputation and intact forearm
limb). The recognition rate of five different hand gestures reaches 90.3% for clenched fist, 86.1% for finger
abduction, 98% for rest, 98.2% for wrist extension and 87.2% for wrist flexion gesture in transradial
amputation case as illustrated in Figure 6(a) based OEBT classifier. While after classification using the
OKNN classifier the result of the confusion matrix for transradial amputation, the recognition rate for each
gesture gives 89.3% for clenched fist,86.3% for finger abduction, 97.8% for rest, 98% for wrist extension and
85.9% for wrist flexion as shown in Figure 6(b). As for the outcomesof recognition rate related to intact
volunteers are shown in Figure 6(c) based OEBT classifier for five various hand movement, i.e., 100% for
clenched fist, 95.5% for finger abduction, 99.4% for rest, 97.1% for wrist extension and 97.1% for wrist
flexion gesture. Whereas the recognition accuracy using OKNN classifier for clenched fist is 100%, for
finger abduction is 93.9%, for rest is 99.6%, for wrist extension is 96.6% and 95% for wrist flexion
movement as shown in Figure 6(d).

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1735-1743

Int J Artif Intell ISSN: 2252-8938  1741

(a) (b)

(a) (b)

Figure 5. Minimum classification error graph: (a) OEBT classifier for transradial amputation limb, (b)
OKNN classifier for transradial amputation limb, (c) OEBT classifier for intact forearm limb, and
(d) OKNN classifier for intact forearm limb

(a) (b)

(c) (d)

Figure 6. Confusion matrix for five gestures: (a) OEBT classifier for transradial amputation limb,
(b) OKNN classifier for transradial amputation limb, (c) OEBT classifier for intact forearm limb, and
(d) OKNN classifier for intact forearm limb

Comparative analysis of machine learning algorithms on myoelectric signal … (Dhirgaam A. Kadhim)

1742  ISSN: 2252-8938

4. CONCLUSION
In the cases of most transradial amputees, it is difficult to extract features and performed classify
myoelectric signals due to weak muscle activities, causing complexity to control wearable hand prostheses.
The myoelectric data was collected using the MYO armband device, where the employed MYO armband
increased the system mobility and flexibility while effectively processing the myoelectric signals. The
proposed preprocessing and overlap-segmentation techniques improved the results recognition of five hand
gestures using myoelectric data signals and in extracting four time-domain features RMS, ZC, IEMG, and
WL for each segment and the classifier. In this paper, the SVM, KNN, DT, TLNN and EBT classification
algorithms were proposed, developed, examined and optimized using the hyperparameters optimization
technique, i.e., OSVM, OKNN, ODT and OEBT on myoelectric data of five gestures for both transradial
amputees and intact forearm limb. The results showed that the classification utilizing OEBT and OKNN that
were more accurate and robustness than those from other classifiers and had a higher rate of recognition for
different hand movements for volunteers either with transradial amputation or healthy forearm limb,
respectively. A recommendation for future study in this field can be combined the MYO armband device
Neuro-headsets device that extracted Electroencephalography signal (EEG) to design couple brain-computer
interface (BCI) system in recognizing patterns for different forearm anatomical movements to obtain high
rate accuracy. Furthermore, connected this system to low-cost field programmable gate arrays (FPGA)
platform to improve machine learning classifier performance.

ACKNOWLEDGEMENTS
We express our sincere thanks and appreciation to our respected Dean of the College of Engineering
and Head of the Prosthetics and Orthotics Engineering Department in University of Kerbala for their
continuous support and provision of financial and logistical support for the research activity in the faculty.

REFERENCES
[1] U. Ghafoor, S. Kim, and K. S. Hong, “Selectivity and longevity of peripheral-nerve and machine interfaces: A review,” Frontiers
in Neurorobotics, vol. 11, no. OCT, pp. 104–116, Oct. 2017, doi: 10.3389/fnbot.2017.00059.
[2] I. Atroshi and H. E. Rosberg, “Epidemiology of amputations and severe injuries of the hand,” Hand Clinics, vol. 17, no. 3,
pp. 343–350, 2001, doi: 10.1016/s0749-0712(21)00515-1.
[3] E. N. Morgan, B. K. Potter, J. M. Souza, S. M. Tintle, and G. P. Nanos, “Targeted muscle reinnervation for transradial
amputation: Description of operative technique,” Techniques in Hand and Upper Extremity Surgery, vol. 20, no. 4, pp. 166–171,
2016, doi: 10.1097/BTH.0000000000000141.
[4] C. McCausland, E. Sawyer, B. J. Eovaldi, and M. Varacallo, “Shoulder and upper limb, shoulder muscles,” StatPearls Publishing,
2021.
[5] J. V. Basmajian and C. J. De Luca, “Muscles alive: their functions revealed by electromyography, 5th ed.,” Baltimore, Williams &
Wilkins, Jul. 1985.
[6] J. R. Cram, G. S. Kasman, and J. Holtz, “Introduction to surface electromyography,” Aspen Publishers Inc., 1998.
[7] M. Raghavan, D. Fee, and P. E. Barkhaus, “Clinical neurophysiology: Basis and technical aspects,” Handbook of Clinical
Neurology, vol. 160, pp. 3–22, 2019, doi: 10.1080/23279095.2020.1754073.
[8] W. D. Kirkland and D. Blankenbaker, “Muscle atrophy,” Musculoskeletal, 2018, doi: 10.5694/j.1326-5377.1921.tb48390.x.
[9] L. McManus, G. De Vito, and M. M. Lowery, “Analysis and Biophysics of Surface EMG for Physiotherapists and Kinesiologists:
Toward a Common Language With Rehabilitation Engineers,” Frontiers in Neurology, vol. 11, no. 1, pp. 104–116, Oct. 2020,
doi: 10.3389/fneur.2020.576729.
[10] L. Morales and D. Pozo, “An experimental comparative analysis among different classifiers applied to identify hand movements
based on sEMG,” 2017 IEEE 2nd Ecuador Technical Chapters Meeting, ETCM 2017, vol. 2017-January, pp. 1–6, 2018,
doi: 10.1109/ETCM.2017.8247504.
[11] M. Çoban and G. Gelen, “Wireless teleoperation of an industrial robot by using myo arm band,” 2018 International Conference
on Artificial Intelligence and Data Processing, IDAP 2018, 2019, doi: 10.1109/IDAP.2018.8620789.
[12] 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, doi: 10.1016/j.eswa.2012.01.102.
[13] M. Zecca, S. Micera, M. C. Carrozza, and P. Dario, “Control of multifunctional prosthetic hands by processing the
electromyographic signal,” Critical Reviews in Biomedical Engineering, vol. 30, no. 4–6, pp. 459–485, 2002,
doi: 10.1615/CritRevBiomedEng.v30.i456.80.
[14] J. Liu, Y. Ren, D. Xu, S. H. Kang, and L. Q. Zhang, “EMG-Based Real-Time Linear-Nonlinear Cascade Regression Decoding of
Shoulder, Elbow, and Wrist Movements in Able-Bodied Persons and Stroke Survivors,” IEEE Transactions on Biomedical
Engineering, vol. 67, no. 5, pp. 1272–1281, 2020, doi: 10.1109/TBME.2019.2935182.
[15] L. Yang and A. Shami, “On hyperparameter optimization of machine learning algorithms: Theory and practice,” Neurocomputing,
vol. 415, pp. 295–316, 2020, doi: 10.1016/j.neucom.2020.07.061.
[16] P. Probst, A. L. Boulesteix, and B. Bischl, “Tunability: Importance of hyperparameters of machine learning algorithms,” Journal
of Machine Learning Research, vol. 20, no. 1, pp. 1–9, Feb. 2019, [Online]. Available: http://journal.um-
surabaya.ac.id/index.php/JKM/article/view/2203.
[17] F. Hutter, H. Hoos, and K. Leyton-Brown, “An efficient approach for assessing hyperparameter importance,” 31st International
Conference on Machine Learning, ICML 2014, vol. 2, pp. 1130–1144, 2014.
[18] E. Elgeldawi, A. Sayed, A. R. Galal, and A. M. Zaki, “Hyperparameter tuning for machine learning algorithms used for arabic
sentiment analysis,” Informatics, vol. 8, no. 4, 2021, doi: 10.3390/informatics8040079.

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1735-1743

Int J Artif Intell ISSN: 2252-8938  1743

[19] M. Kefalas, M. Koch, V. Geraedts, H. Wang, M. Tannemaat, and T. Back, “Automated Machine Learning for the Classification
of Normal and Abnormal Electromyography Data,” Proceedings - 2020 IEEE International Conference on Big Data, Big Data
2020, pp. 1176–1185, 2020, doi: 10.1109/BigData50022.2020.9377780.
[20] S. A. Raurale, J. McAllister, and J. M. Del Rincon, “EMG Biometric Systems Based on Different Wrist-Hand Movements,” IEEE
Access, vol. 9, pp. 12256–12266, 2021, doi: 10.1109/ACCESS.2021.3050704.
[21] M. Tavakoli, C. Benussi, P. Alhais Lopes, L. B. Osorio, and A. T. de Almeida, “Robust hand gesture recognition with a double
channel surface EMG wearable armband and SVM classifier,” Biomedical Signal Processing and Control, vol. 46, pp. 121–130,
2018, doi: 10.1016/j.bspc.2018.07.010.
[22] M. Ladrova, R. Martinek, J. Nedoma, and M. Fajkus, “Methods of power line interference elimination in EMG signals,” Journal
of Biomimetics, Biomaterials and Biomedical Engineering, vol. 40, pp. 64–70, 2019,
doi: 10.4028/www.scientific.net/JBBBE.40.64.
[23] D. Karabulut, F. Ortes, Y. Z. Arslan, and M. A. Adli, “Comparative evaluation of EMG signal features for myoelectric controlled
human arm prosthetics,” Biocybernetics and Biomedical Engineering, vol. 37, no. 2, pp. 326–335, 2017,
doi: 10.1016/j.bbe.2017.03.001.
[24] A. Phinyomark, R. N. Khushaba, and E. Scheme, “Feature extraction and selection for myoelectric control based on wearable
EMG sensors,” Sensors (Switzerland), vol. 18, no. 5, 2018, doi: 10.3390/s18051615.
[25] Z. Yi et al., “Extracting time-frequency feature of single-channel vastus medialis EMG signals for knee exercise pattern
recognition,” PLoS ONE, vol. 12, no. 7, 2017, doi: 10.1371/journal.pone.0180526.
[26] T. Tuncer, S. Dogan, and A. Subasi, “Surface EMG signal classification using ternary pattern and discrete wavelet transform
based feature extraction for hand movement recognition,” Biomedical Signal Processing and Control, vol. 58, 2020,
doi: 10.1016/j.bspc.2020.101872.
[27] A. Krasoulis, S. Vijayakumar, and K. Nazarpour, “Multi-Grip Classification-Based Prosthesis Control with Two EMG-IMU
Sensors,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 28, no. 2, pp. 508–518, 2020,
doi: 10.1109/TNSRE.2019.2959243.
[28] C. Cerci and H. Temeltas, “Feature extraction of EMG signals, classification with ANN and kNN algorithms,” 26th IEEE Signal
Processing and Communications Applications Conference, SIU 2018, pp. 1–4, 2018, doi: 10.1109/SIU.2018.8404207.
[29] E. Gokgoz and A. Subasi, “Comparison of decision tree algorithms for EMG signal classification using DWT,” Biomedical Signal
Processing and Control, vol. 18, pp. 138–144, 2015, doi: 10.1016/j.bspc.2014.12.005.
[30] E. Yaman and A. Subasi, “Comparison of Bagging and Boosting Ensemble Machine Learning Methods for Automated EMG
Signal Classification,” BioMed Research International, vol. 2019, 2019, doi: 10.1155/2019/9152506.
[31] Q. Zhou, F. Zhang, and C. Yang, “AdaNN: Adaptive Neural Network-Based Equalizer via Online Semi-Supervised Learning,”
Journal of Lightwave Technology, vol. 38, no. 16, pp. 4315–4324, 2020, doi: 10.1109/JLT.2020.2991028.
[32] F. Ye, B. Yang, C. Nam, Y. Xie, F. Chen, and X. Hu, “A Data-Driven Investigation on Surface Electromyography Based Clinical
Assessment in Chronic Stroke,” Frontiers in Neurorobotics, vol. 15, no. 1, pp. 104–116, Jul. 2021,
doi: 10.3389/fnbot.2021.648855.

BIOGRAPHIES OF AUTHORS

Dhirgaam A. Kadhim he is currently lecturing with the Prosthetics and Orthotics

Engineering Department, University of Kerbala, Iraq. He received the B.S. and M.Sc. degrees
from the College of Engineering, University of Baghdad, Iraq. He got the Ph.D. degree from the
Mechatronics and Robotics Control Engineering, Vladimir State University, Russian Federation.
His current research interests include control bio-mechatronics systems, smart prosthesis-
orthosis design, biosignal processing control and artificial intelligent. He can be contacted at
email: [email protected].

Mithaq N. Raheema is Assistance Professor in the Department of Prosthetics and

Orthotics Engineering, University of Kerbala, Iraq. He has received a Ph.D. M.Sc. and B.Sc.
Degree in Electronic Engineering from the Department of Electrical and Electronic Engineering,
University of Technology, Iraq. He has published a significant number of articles in international
conferences and journals. His research interests are in the areas of artificial intelligent, e-
Learning, digital and myoelectric control systems. He can be contacted at email:
[email protected].

Jabbar S. Hussein was born in Bagdad Iraq in 1974. He received the B.S., M.S.,
and Ph. D. degrees in Electronic Engineering from University of Technology in 1998, 2001 and
2019 respectively. From 2000 to 2007 he worked at University of Technology/Electrical
Engineering Department as lecturer. Since 2007, he has been working as a lecturer in University
of Kerbala/Collage of Engineering. He worked also in many companies as a Broadcasting
engineer. During this period many research were published in international and local conferences
and journals in the field of Speech Recognitions, Prosthesis Control, Antenna Design and Mobile
Radiation Protection Systems. He can be contacted at email: [email protected].

Comparative analysis of machine learning algorithms on myoelectric signal … (Dhirgaam A. Kadhim)