This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

11institutetext: School of Computer Science and Engineering, The University of Aizu, Aizuwakamatsu, Fukushima, Japan.
11email: [email protected], *11email: [email protected]

EMG-Based Hand Gesture Recognition through Diverse Domain Feature Enhancement and Machine Learning-Based Approach

Abu Saleh Musa Miah 11 0000-0002-1238-0464    Najmul Hassan 11 0009-0000-6499-1825    Md. Maniruzzaman 11 0000-0001-6151-8071    Nobuyoshi Asai 11    Jungpil Shin 11** 0000-0002-7476-2468
Abstract

Surface electromyography (EMG) serves as a pivotal tool in hand gesture recognition and human-computer interaction, offering a non-invasive means of signal acquisition. This study presents a novel methodology for classifying hand gestures using EMG signals. To address the challenges associated with feature extraction where, we explored 23 distinct morphological, time domain and frequency domain feature extraction techniques. However, the substantial size of the features may increase the computational complexity issues that can hinder machine learning algorithm performance. To mitigate this, we employ an efficient feature selection approach, specifically an extra tree classifier. The selected potential feature fed into the various machine learning-based classification algorithms where our model achieved 97.43% accuracy with the KNN algorithm and selected feature. By leveraging a comprehensive feature extraction and selection strategy, our methodology enhances the accuracy and usability of EMG-based hand gesture recognition systems. The higher performance accuracy proves the effectiveness of the proposed model over the existing system.

Keywords:
EMG signal, machine learning approach, hand gesture recognition, feature selection, sensor-based hand gesture recognition.

1 Introduction

The electrical activity of the skeletal muscles generates electromyography (EMG) signals, which can be captured in two ways: invasive and noninvasive methods [1, 2]. These kinds of signals is one of the most crucial ways to implement human-machine communication based on hand gestures or human activity information [3, 4]. EMG-based hand gesture recognition is crucial for various applications, including human-computer interaction, prosthetics, rehabilitation, and biomedical research. It enables individuals to control devices and interfaces using natural hand movements, improving accessibility and quality of life [5]. EMG signals are utilized in systems for movement classification [6, 7], recognition of diagnosis of neuromuscular diseases [8] and motor unit action potential (MUAP) [9]. For instance, EMG signals are used to recognize finger movements for controlling prosthetic devices or computer interfaces, facilitating intuitive interaction for individuals with limb disabilities [10]. Many researchers have been working to develop EMG-based hand gesture recognition using various statistical and mathematical formula-based features. Among them, Hudgins et al.  cites tabares2019deep extracted 4-time domain features and then fed them into the classification approach that produced good performance accuracy compared to the previous study. However, challenges such as changes in fatigue, arm posture, body movement, and natural variability of the human body may also impact the feature effectiveness [11, 12]. More recently, several studies have been done and compared with other feature extraction excluded time and frequency-domain features, but simpler time-domain features remain robust in classification tasks [13, 14]. More recently, Rubio et proposed an EMG-based hand gesture recognition method with various machine learning models, and they reported 95.39% accuracy [1]. The drawback of their model is that it did not achieve satisfactory performance accuracy due to a lack of feature effectiveness. To increase performance accuracy, it is important to extract the effective feature using various statistical and mathematical formulas. To overcome the problem, we proposed morphological, time domain and frequency domain feature-based EMG-based hand gesture recognition with effective feature selection and classification modules.

The proposed methodology contributes to the field of hand gesture classification using EMG signals in several key ways:

  • •

    Comprehensive Feature Extraction: The paper introduces the novel feature extraction approach by employing morphological, time domain, and frequency domain feature extraction techniques. This comprehensive approach addresses the challenge of capturing diverse aspects of the EMG signal, enhancing the richness of information available for classification.

  • •

    Effective Feature Selection: To mitigate the computational complexity and reduce the cost associated with large datasets, the methodology incorporates an Extra tree classifier (ETC) algorithm for feature selection. This step helps identify the most relevant features for classification, improving the efficiency of subsequent analysis.

  • •

    Classification Approach: The methodology applies various machine learning algorithms for classification with the original features and selected features. By exploring different algorithms, the study identifies models that yield high classification accuracy, contributing to the advancement of classification techniques.

  • •

    Extensive Evaluation: The experimental results demonstrate the effectiveness of the proposed approach, with classification accuracy exceeding 97.43% with the selected features using the KNN approach. This performance surpasses the existing state-of-the-art models in terms of performance accuracy and efficiency.

Table 1: Dataset description
Label Gesture Name
0 Intervals between gestures
1 Wrist extension
2 Grasped hand
3 Resting hand
4 Wrist flexion
5 Radial deviation
6 Ulnar deviation

2 Related Work

Many academic and industrial research projects have been done to develop EMG based hand gesture recognition systems. Some researchers used feature extraction techniques, including autoregressive models [15], signal entropy measurements [16], and amplitude of the time and frequency domain-based statistical measurements. Wavelet transforms are used by some researchers for extracting time-frequency domain features from the EMG-based hand gesture dataset to develop a hand gesture recognition system [17]. Several studies have been working on the development of HGR systems, which are based on the feature extraction and machine learning (ML) algorithm [18, 19, 20, 21, 22, 23, 24]. To improve the performance accuracy, Li et al.[25] proposed an ensemble approach that combined the random forest and hosting method and achieved 80% accuracy on an sEMG-based hand gesture dataset. More recently, Khushaba et al. employed a hand gesture recognition system with 12 FD and TD features of the EMG signal, and they showed 90% accuracy using the SVM approach [26]. Praveen et al. introduced a feature extraction approach incorporating root mean square and integrated absolute value techniques with KNN and Bayesian pattern classification algorithms. Their method yielded promising results, achieving a recognition accuracy 92% in sEMG-based gesture recognition tasks [27]. More recently, Azhhiri et al. employed frequency domain (FD) features extracted from different levels of wavelet decomposition, leading to notable enhancements in sEMG-based Hand Gesture Recognition (HGR) when integrated with a multilayer perceptron model [28]. Arteaga et al. augmented performance accuracy and efficiency by gathering sEMG signals pertaining to six distinct hand gestures from a cohort of 20 participants and subsequently employing the k-Nearest Neighbors (kNN) algorithm [29]. More recently, Rubio et proposed an EMG-based hand gesture recognition method with various machine learning models, and they reported 95.39% accuracy [1]. The drawback of their model is that it did not achieve satisfactory performance accuracy due to a lack of feature effectiveness. It is important to extract the effective feature using various statistical and mathematical formulas to increase performance accuracy. To overcome the problem, we proposed morphological, time domain and frequency domain feature-based EMG-based hand gesture recognition with effective feature selection and classification modules.

3 Dataset

We collected a free EMG-based hand gesture dataset, which is available in the UCI library [30]. The dataset consisted of six static hand gestures collected from 36 patients and a total of 72 recordings. Two MYO thalamic bracelet devices, which have eight channels, were used to collect the dataset. The set of static hand gestures encompassed wrist extension, grasped hand, resting hand, wrist flexion, radial deviation and ulnar deviation which are shown in Table 1. The dataset of each gesture was recorded for 3 and 3 seconds for the duration and interval between gestures, respectively. [1].

Figure 1: Workflow architecture of the proposed methodology.
Refer to caption

4 Proposed Methodology

This study proposes an innovative methodology for hand gesture classification using surface electromyography (EMG) signals. Figure 1 demonstrated the proposed methodology where we included several key components: preprocessing, feature extraction, feature selection and classification module. Firstly, we employ 23 different morphological, time-domain and frequency-domain feature extraction techniques to tackle feature extraction challenges from the EMG signals. We extracted these features to capture diverse aspects of the signal’s characteristics, thereby enhancing the discriminative power of the feature set. In addition, to reduce the computational complexity, we employed an effective feature selection approach to mitigate these challenges, which leads to increased performance accuracy and model efficiency as well. By applying a combination of feature extraction, selection, and classification techniques, our approach aims to optimize the overall classification process. We leverage various machine learning algorithms to classify EMG hand gestures here. These algorithms are chosen based on their suitability for handling high-dimensional data and their capacity to generalize well across different subjects. By prioritizing this information, we aim to identify patterns associated with six distinct gestures, irrespective of the individual acting. This generalization among different subjects addresses a significant challenge in EMG-based hand gesture recognition. The proposed methodology is evaluated using EMG recordings from 36 patients performing six hand gestures.

4.1 Preprocessing

During preprocessing, irrelevant EMG signal segments denoted as zero, representing resting intervals between gestures, were eliminated to focus solely on hand movement data. Instead of feature extraction, raw data from the device’s eight channels were directly used. This approach aimed to exploit the inherent discriminatory information in raw EMG signals, maintaining data richness and granularity to enhance classification accuracy [1, 2].

Table 2: Morphological features description
Feature Name Definition Formula
Absolute Amplitude
Denoted as xx, is calculated by taking the absolute
value of the input signal xx.
x=|x|x=\lvert x\rvert
Minimum Amplitude
Denoted as xminx_{\text{min}}, is the smallest value present in the input signal xx.
xmin=min⁡(x)x_{\text{min}}=\min(x)
Maximum Amplitude
Denoted as xmaxx_{\text{max}}, is the largest value present in the input signal xx.
xmax=max⁡(x)x_{\text{max}}=\max(x)
Peak to Peak (PTP)
This range, denoted as PTP, is a feature extraction method for EMG signals. It is calculated as the difference between the maximum amplitude (xmaxx_{\text{max}}) and the minimum amplitude (xminx_{\text{min}})
PTP=xmax−xmin\text{PTP}=x_{\text{max}}-x_{\text{min}}

4.2 Feature Extraction

Effective feature extraction is used to highlight the meaningful structure and pattern of the signal, which is one of the most challenging tasks for hand gesture recognition with EMG signals. We extracted morphological features, time domain features, and frequency domain features we described below:

4.2.1 Morphological Feature

We calculated the morphological features from the EMG signals based on the following formula that is described in Table 2.

4.2.2 Time Domain Feature

We have employed different statistical approaches to obtain the time domain features from the EMG signal. The details description of this feature is described in Table 3.

Table 3: Time domain features description
Feature Name Definition Formula
Arithmetic Mean
A common feature of the time series data is the mean.
μ=1N​∑i=1Nxi\mu=\frac{1}{N}\sum_{i=1}^{N}x_{i}
Median
Let nn be the number of samples and μ\mu be the mean of the nth numbers. , where “n” is the number of items in the set and “th” just means the (n)th number..
m=(n+1)2⋅μm=\frac{(n+1)}{2}\cdot\mu
Sample Variance (s​v​a​r​ssvars)
the deviation summed up over the samples and then divided by the sample size.
σ2=1(n−1)​∑i=1n(xi−μ)2\sigma^{2}=\frac{1}{(n-1)}\sum_{i=1}^{n}(x_{i}-\mu)^{2}
Population Variance (s​v​a​r​s​psvarsp)
It is the same as σ2\sigma^{2}, with the main difference being that the population variance is calculated based on the total population, whereas sample variance is calculated based on the sample size nn
s​i​g​m​a2=1N​∑i=1N(xi−μ)2sigma^{2}=\frac{1}{N}\sum_{i=1}^{N}(x_{i}-\mu)^{2}
Mean Absolute Deviation
measure of the average absolute difference between each data point and the mean of the dataset.
MAD=1n​∑i=1n|xi−x¯|\text{MAD}=\frac{1}{n}\sum_{i=1}^{n}|x_{i}-\bar{x}|
Sample Standard Deviation (s​t​d​sstds)
Denoted as σ\sigma, is calculated as the square root of the sample variance.
σ=1(n−1)​∑i=1n(xi−μ)2\sigma=\sqrt{\frac{1}{(n-1)}\sum_{i=1}^{n}(x_{i}-\mu)^{2}}
Population Standard Deviation (s​s​t​d​psstdp)
Calculated from population variance
σ=1N​∑i=1N(xi−μ)2\sigma=\sqrt{\frac{1}{N}\sum_{i=1}^{N}(x_{i}-\mu)^{2}}
Percentile
Let pp-value be 50 and nn be the total number of samples.
Percentile=p100×(n+1)\text{Percentile}=\frac{p}{100\times(n+1)}
Quantile
calculated with the first quartile. It also known as the lower quartile
Q​1=((n+1)4)​th TermQ1=\left(\frac{(n+1)}{4}\right)\text{th Term}
Interquartile Range (i​q​siqs)
Calculated as the difference between the upper quartile and
the lower quartile. Let Q​2=Q​3−Q​1Q2=Q3-Q1
Q1=(14)​[(N+1)]​th term\text{Q1}=\left(\frac{1}{4}\right)\left[(N+1)\right]\text{th term},
Q3=(34)​[(N+1)]​th term\text{Q3}=\left(\frac{3}{4}\right)\left[(N+1)\right]\text{th term}
n=number of data pointsn=\text{number of data points}
I​Q​R=Q​3−Q​1IQR=Q3-Q1
Skewness
Symmetry of the signal measured with this. Let 3rd and 2nd order central moments of the respective EMG signals m3m_{3} and m2m_{2} then we can write:
m3=E​[(xi−μ)3]=1n​∑i=1n(xi−μ)3m_{3}=E\left[(x_{i}-\mu)^{3}\right]=\frac{1}{n}\sum_{i=1}^{n}(x_{i}-\mu)^{3}
m2=E​[(xi−μ)2]=1n​∑i=1n(xi−μ)2m_{2}=E\left[(x_{i}-\mu)^{2}\right]=\frac{1}{n}\sum_{i=1}^{n}(x_{i}-\mu)^{2}
G1=m3m23/2G_{1}=\frac{m_{3}}{m_{2}^{3/2}}
Kurtosis
Tailness of distribution relative to the normal distribution measurement. Let 4th and 2nd order central moments of the respective EMG signals m4m_{4} and m2m_{2} then we can write:
m4=E​[(xi−μ)4]=1n​∑i=1n(xi−μ)4m_{4}=E\left[(x_{i}-\mu)^{4}\right]=\frac{1}{n}\sum_{i=1}^{n}(x_{i}-\mu)^{4}
m2=E​[(xi−μ)2]=1n​∑i=1n(xi−μ)2m_{2}=E\left[(x_{i}-\mu)^{2}\right]=\frac{1}{n}\sum_{i=1}^{n}(x_{i}-\mu)^{2}
G2=m4m22G_{2}=\frac{m_{4}}{m_{2}^{2}}

4.2.3 Frequency Domain Feature

We also extracted frequency-domain features, which are shown in Table 4.

Table 4: Frequency domain features description
Feature Name Definition Formula
Energy
It is a major frequency domain feature extraction method.
En=∑i=1Nxi2\text{En}=\sum_{i=1}^{N}x_{i}^{2}
Power P=1N​∑i=1Nxi2P=\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}
Root Mean Square (RMS)
Denoted as xmaxx_{\text{max}}, is the largest value present in the input signal xx.
RMS=1N​∑i=1Nxi2\text{RMS}=\sqrt{\frac{1}{N}\sum_{i=1}^{N}x_{i}^{2}}
Hjort Parameter Activity (HPA) HPA=1N​∑i=1N(xi−μ)2\text{HPA}=\frac{1}{N}\sum_{i=1}^{N}(x_{i}-\mu)^{2}

4.3 Feature Selection With Extra Tree Classifier

After extracting features from EMG signals, we employed the Extra Trees classifier (ETC) for feature selection due to its robustness and efficiency. Extra Trees introduces significant randomization during node splitting, generating fully randomized trees in some cases. This approach enhances diversity and reduces variance, making it well-suited for feature selection tasks. ETC first divide the features into subsets by splitting to ensure a diverse feature combination. This splitting is performed by criteria like Gini impurity or entropy, with the selected feature and threshold yielding the highest impurity reduction per split. Then, it calculates the importance of each individual feature after constructing all trees; feature importance is computed by averaging impurity reduction values across the ensemble. Finally, Importances are typically normalized or converted to percentages for ranking, aiding in identifying the most relevant features [2].

4.4 Classification

We implemented various ML models in the study, including Random Forest, Extra Tree Classifier, SVM, Gaussian NB, Gradient Boosting, HGboosting, Decision Tree and KNN [31, 32, 33, 34, 35].
K-nearest neighbour is a popular ML-based algorithm employed for classification tasks that involve storing labelled training data. Then, it calculated the similarity matrix between the label and the unlabeled object, aiming to identify and group the k closest elements. These number groups are constructed with the closest element and then used to classify the object. If k is greater than 2, then the label is measured by the voting system; otherwise, the label is measured by the closest element’s label [31].

Random Forest (RF) consists of several numbers of the Decision Trees (DT), each unique in its construction. This uniqueness is achieved by randomly selecting training data that is also considered a feature in each tree. The majority voting-based approach is used to predict the final prediction of this ensemble approach [31].

Gaussian Naive Bayes (GBN) rely on a probabilistic model obtained from Bayes’ theorem. The working procedure of this algorithm is that it considers all features attribute to be independent and then, based on the feature probability, it calculate the probability of the corresponding classes, and it uses Gaussian distribution for this [31].

Decision Trees (DT) creates a tree-like structure to make decisions based on input features. The fundamental idea behind a DT is to learn a hierarchy of yes/no (if/else) questions about the specific feature, and each branch represent the possible answers (yes or no) [31].
SVM finds the optimal hyperplane to separate data points into different classes, and it also maximizes the margin between the closest different classes. SVM can perform both linear and non-linear tasks by using kernel tricks.
Gradient Boost (GB) sequentially combines multiple weak learners, typically DT, to improve the accuracy of performance prediction. It performs by fitting new weak learners to the errors of the previous ones, gradually reducing prediction errors.
HGBoost utilizes histograms to efficiently compute gradients during training. It organizes data into histograms, which reduces the computational complexity of computing gradients, especially for large datasets.

5 Result and Discussion

In this section, we experimented with our proposed model with an existing benchmark EMG-based hand gesture recognition dataset and produced the performance matrix.

5.1 The Metrics Used to Evaluate Models

We calculated different performance matrices, including accuracy, recall, precision, and specificity, demonstrated in Table 5.

Table 5: The commonly used offline performance metrics
Evaluation Metric Formula Comments
Accuracy TP+TNTP+TN+FP+FN\frac{T_{P}+T_{N}}{T_{P}+T_{N}+F_{P}+F_{N}} -
Recall TPTP+FN\frac{T_{P}}{T_{P}+F_{N}} -
Precision TPTP+FP\frac{T_{P}}{T_{P}+F_{P}} -
F1 score 2×P​r​e​c​i​s​i​o​n×R​e​c​a​l​lP​r​e​c​i​s​i​o​n+R​e​c​a​l​l2\times\frac{Precision\times Recall}{Precision+Recall}
Specificity TNTN+FP\frac{T_{N}}{T_{N}+F_{P}} -
Table 6: Performance Accuracy with all the features and selected features
Feature
Number
Accuracy
Selected
Feature
Performance Sensitivity Specificity
Random Forest 21 90.00 10 97.00 99.81 99.41
SVM 21 96.00 10 96.00
Gaussian NB 21 39.00 10 36.00 n/a n/a
Gradient Boost 21 36.00 10 40.00 n/a n/a
HGboost 21 40.00 10 45.00 n/a n/a
Decision Tree 21 97.00 10 97.00 99.88 99.87
KNN 21 97.00 10 97.43 99.88 99.86

5.2 Performance Result with Various Methodology

Table 6 presents performance metrics for machine learning models using all features and a subset of selected features. Each row corresponds to a model, showing feature count, accuracy, selected feature count, performance, sensitivity, and specificity. Extra Tree led with 97.30% accuracy using ten features, followed closely by Decision Tree at 97.00%. Random Forest scored 90.00% with all features and 97.00% with ten selected features. SVM maintained 96.00% accuracy for both settings. Gaussian NB, Gradient Boost, and HGboost ranged from 36.00% to 40.00% with 10 selected features. KNN notably achieved 99.88% sensitivity and 99.86% specificity with ten selected features.

Table 7: Precision recall and F1 score with K-nearest neighbor algorithm
KNN Random Forest Decision Tree
Gesture Name Precision Recall
F1-
Score
Precision Recall
F1-
Score
Precision Recall
F1-
Score
Resting hand 99.00 99.00 99.00 96.00 99.00 97.00 97.00 97.00 97.00
Grasped hand 97.00 97.00 97.00 96.00 97.00 97.00 97.00 97.00 97.00
Wrist flexion 97.00 97.00 97.00 96.00 95.00 96.00 97.00 97.00 97.00
Wrist extension 97.00 97.00 97.00 97.00 97.00 97.00 97.00 97.00 97.00
Ulnar deviation 97.00 97.00 97.00 97.00 96.00 96.00 97.00 97.00 97.00
Radial Variation 97.00 97.00 97.00 97.00 96.00 97.00 99.00 99.00 99.00

Table 7 displays labels’ performance accuracy with the KNN approach, where we included precision, recall, and F1 scores for each individual label. We observed that all gestures produced almost similar performance matrix information that proved the superiority of the proposed model.

5.3 State of the art comparison

Table 8 showcases a state-of-the-art comparison of the proposed model where our proposed model achieves good performance accuracy compared to the existing systems. Rubio et al. proposed models utilizing Random Forest and Convolutional Neural Network (CNN) algorithms, achieving accuracies of 95.39% and 94.70%, respectively, but lacked feature extraction or selection techniques. In contrast, our method, employing geometric and statistical feature extraction with Extra Trees Classifier (ETC) for selection, outperformed prior models with 97.43% accuracy. This highlights our approach’s efficacy in enhancing classification performance. Our model’s strength lies in its comprehensive feature extraction, capturing nuanced gesture data representation. ETC’s feature selection further boosts efficiency, while the K-nearest neighbour (KNN) algorithm ensures simplicity and robustness, making it ideal for real-world applications like human-computer interaction and healthcare technologies. The KNN achieved high-performance accuracy compared to the existing CNN algorithm due to the comprehensive feature extraction and efficient feature selection process, which provided a rich and relevant representation of the EMG signals, allowing KNN to classify hand gestures effectively. In contrast, the CNN model’s performance was hindered by its higher computational complexity and potential overfitting, especially with a limited dataset, making it less effective than KNN in this specific application.

Table 8: State of the art comparison of the proposed model
Author
Name
Feature
Extraction
Feature
Selection
Machine
Learning Algorithm
Performance
accuracy
Rubio et al.[1] No No Random Forest 95.39
Rubio et al.[1] No No CNN 94.70
Proposed method
Morphological, Time and
Frequency domain features
ETC KNN 97.43

6 Conclusion

In the study, we introduce a robust methodology for classifying hand gestures using various statistical and mathematical features. We address the challenges associated with feature effectiveness and computational complexity by leveraging 23 feature extraction techniques and employing effective feature selection via the ETC approach. Our model achieved the goal by generating high-performance accuracy compared to the existing system. This signifies the potential feature extraction technique that leads to the accuracy and efficiency of EMG-based hand gesture recognition systems. Overall, our study contributes to advancing the field of EMG-based hand gesture recognition by providing a comprehensive methodology that optimizes feature extraction, selection, and classification processes. The challenges of this study include computational complexity during feature extraction and variability of EMG signals among individuals. Future work will focus on optimizing the model for real-time applications, validating robustness across diverse datasets, exploring hybrid models for improved accuracy and efficiency, and enhancing adaptability for diverse hand gestures and user populations.

References

  • [1] Alzate Grisales JA, R Tabares-Soto, S Orozco-Arias, Jiménez Varón CF, Padilla Buriticá JI, et al. Identification of hand movements from electromyographic signals using machine learning. 2020.
  • [2] Abu Saleh Musa Miah, Jungpil Shin, and Md Al Mehedi Hasan. Effective features extraction and selection for hand gesture recognition using semg signal. Multimedia Tools and Applications, pages 1–25, 2024.
  • [3] Angkoon Phinyomark, Rami N. Khushaba, and Erik Scheme. Feature extraction and selection for myoelectric control based on wearable emg sensors. Sensors, 18:1615, 2018.
  • [4] Abu Saleh Musa Miah, Md Abdur Rahim, and Jungpil Shin. Motor-imagery classification using riemannian geometry with median absolute deviation. Electronics, 9(10):1584, 2020.
  • [5] Rami N. Khushaba, Sarath Kodagoda, M. Takruri, and Gamini Dissanayake. Toward improved control of prosthetic fingers using surface electromyogram (emg) signals. Expert Systems with Applications, 39:10731–10738, 2012.
  • [6] Y. Ji, S. Sun, and H.B. Xie. Stationary wavelet-based two-directional two-dimensional principal component analysis for emg signal classification. Measurement Science Review, 17:117–124, 2017.
  • [7] M. Ghofrani Jahromi, H. Parsaei, A. Zamani, and D.W. Stashuk. Cross comparison of motor unit potential features used in emg signal decomposition. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 26:1017–1025, 2018.
  • [8] S. Koçer and A.E. Tümer. Classifying neuromuscular diseases using artificial neural networks with applied autoregressive and cepstral analysis. Neural Computing and Applications, 28:945–952, 2017.
  • [9] A. Benazzouz, R. Guilal, F. Amirouche, and Z.E. Hadj Slimane. Emg feature selection for diagnosis of neuromuscular disorders. In 2019 International Conference on Networking and Advanced Systems (ICNAS), pages 1–5, 2019.
  • [10] N. Phukan, N.M. Kakoty, P. Shivam, and J.Q. Gan. Finger movements recognition using minimally redundant features of wavelet denoised emg. Health and Technology, 2019.
  • [11] S. Orozco-Arias, G. Isaza, R. Guyot, and R. Tabares-Soto. A systematic review of the application of machine learning in the detection and classification of transposable elements. PeerJ, 7:e8311, 2019.
  • [12] Abu Saleh Musa Miah, Md. Al Mehedi Hasan, Yoichi Tomioka, and Jungpil Shin. Hand gesture recognition for multi-culture sign language using graph and general deep learning network. IEEE Open Journal of the Computer Society, 5:144–155, 2024.
  • [13] A. Subasi, E. Yaman, Y. Somaily, H.A. Alynabawi, F. Alobaidi, and S. Altheibani. Automated emg signal classification for diagnosis of neuromuscular disorders using dwt and bagging. Procedia Computer Science, 140:230–237, 2018.
  • [14] A. Subasi and E. Yaman. Emg signal classification using discrete wavelet transform and rotation forest. In A. Badnjevic, R. Škrbi´c, and L. Gurbeta Pokvi´c, editors, CMBEBIH 2019, pages 29–35, Cham, 2020. Springer International Publishing.
  • [15] W. Chen, Z. Wang, H. Xie, and W. Yu. Characterization of surface emg signal based on fuzzy entropy. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 15:266–272, 2007.
  • [16] E. Gokgoz and A. Subasi. Comparison of decision tree algorithms for emg signal classification using dwt. Biomedical Signal Processing and Control, 18:138–144, 2015.
  • [17] S.P. Singh and S. Urooj. Wavelets: Biomedical applications. International Journal of Biomedical Engineering and Technology, 19:1–25, 2015.
  • [18] Zahra Taghizadeh, Saeid Rashidi, and Ahmad Shalbaf. Finger movements classification based on fractional fourier transform coefficients extracted from surface emg signals. Biomedical Signal Processing and Control, 68:102573, 2021.
  • [19] Anand Kumar Mukhopadhyay and Suman Samui. An experimental study on upper limb position invariant emg signal classification based on deep neural network. Biomedical signal processing and control, 55:101669, 2020.
  • [20] Mohammad Esmaeili and Aria Nosratinia. Semi-supervised node classification by graph convolutional networks and extracted side information. arXiv preprint arXiv:2009.13734, 2020.
  • [21] Abu Saleh Musa Miah, Md Al Mehedi Hasan, and Jungpil Shin. Dynamic hand gesture recognition using multi-branch attention based graph and general deep learning model. IEEE Access, 2023.
  • [22] Abu Saleh Musa Miah, Jungpil Shin, Md Al Mehedi Hasan, and Md Abdur Rahim. Bensignnet: Bengali sign language alphabet recognition using concatenated segmentation and convolutional neural network. Applied Sciences, 12(8):3933, 2022.
  • [23] Abu Saleh Musa Miah, Md Al Mehedi Hasan, Jungpil Shin, Yuichi Okuyama, and Yoichi Tomioka. Multistage spatial attention-based neural network for hand gesture recognition. Computers, 12(1):13, 2023.
  • [24] Abu Saleh Musa Miah, Jungpil Shin, Md Al Mehedi Hasan, Md Abdur Rahim, and Yuichi Okuyama. Rotation, translation and scale invariant sign word recognition using deep learning. Computer Systems Science and Engineering, 44(3):2521–2536, 2023.
  • [25] Zhijun Li, Baocheng Wang, Chenguang Yang, Qing Xie, and Chun-Yi Su. Boosting-based emg patterns classification scheme for robustness enhancement. IEEE Journal of Biomedical and Health Informatics, 17(3):545–552, 2013.
  • [26] Ali H Al-Timemy, Guido Bugmann, Javier Escudero, and Nicholas Outram. Classification of finger movements for the dexterous hand prosthesis control with surface electromyography. IEEE journal of biomedical and health informatics, 17(3):608–618, 2013.
  • [27] LS Praveen, SN Nagananda, and Preetham Shankapal. Design and development of real time bionic hand control using emg signal. In 2018 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT), pages 1–4. IEEE, 2018.
  • [28] Reza Bagherian Azhiri, Mohammad Esmaeili, and Mehrdad Nourani. Emg-based feature extraction and classification for prosthetic hand control. arXiv preprint arXiv:2107.00733, 2021.
  • [29] Maria V Arteaga, Jenny C Castiblanco, Ivan F Mondragon, Julian D Colorado, and Catalina Alvarado-Rojas. Emg-driven hand model based on the classification of individual finger movements. Biomedical Signal Processing and Control, 58:101834, 2020.
  • [30] N. Krilova, I. Kastalskiy, V. Kazantsev, V. A. Makarov, and S. Lobov. EMG Data for Gestures. UCI Machine Learning Repository, 2019. DOI: https://doi.org/10.24432/C5ZP5C.
  • [31] Alejandro Mora Rubio, Jesus Alejandro Alzate Grisales, and Tabares-Soto. Identification of hand movements from electromyographic signals using machine learning. 2020.
  • [32] Abu Saleh Musa Miah, Jungpil Shin, Md Al Mehedi Hasan, Md Khademul Islam Molla, Yuichi Okuyama, and Yoichi Tomioka. Movie oriented positive negative emotion classification from eeg signal using wavelet transformation and machine learning approaches. In IEEE 15th international symposium on embedded multicore/many-core systems-on-chip (MCSoC), pages 26–31. IEEE, 2022.
  • [33] Abu Saleh Musa Miah, Jungpil Shin, Md Minhajul Islam, Md Khademul Islam Molla, et al. Natural human emotion recognition based on various mixed reality (mr) games and electroencephalography (eeg) signals. In 2022 IEEE 5th Eurasian conference on educational innovation (ECEI), pages 408–411. IEEE, 2022.
  • [34] Md Humaun Kabir, Nadim Ibne Akhtar, Nishat Tasnim, Abu Saleh Musa Miah, Hyoun-Sup Lee, Si-Woong Jang, and Jungpil Shin. Exploring feature selection and classification techniques to improve the performance of an electroencephalography-based motor imagery brain-computer interface system. Sensors (Basel, Switzerland), 24(15):4989, 2024.
  • [35] Abu Saleh Musa Miah, Mumtahina Afroz Mouly, Chandrika Debnath, Jungpil Shin, and SM Sadakatul Bari. Event-related potential classification based on eeg data using xdwan with mdm and knn. In International Conference on Computing Science, Communication and Security, pages 112–126. Springer, 2021.