Electromyography based hand movement classification and feature extraction using machine learning algorithms

Abstract

The categorization of hand gestures holds significant importance in controlling orthotic and prosthetic devices, enabling human-machine interaction, and facilitating telerehabilitation applications. For many years, methods of motion analysis based on image processing techniques have been employed to detect hand motions. However, recent research has focused on utilizing muscle contraction for detecting hand movements. Specifically, there has been an increase in studies that classify hand movements using surface electromyography (sEMG) data from the muscles of the hand and arm. In our study, we estimated the open (extension of the fingers) and closed (flexion of the fingers) positions of the hand by analyzing EMG data obtained from 4 volunteer participants' Extensor digitorum and Flexor carpi radialis muscles. In order to accurately discriminate EMG signals, various statistical measures such as variance, standard deviation, root mean square, average energy, minimum and maximum features were utilized. The dataset containing these additional features was then subjected to classification algorithms including Support Vector Machines (SVM), K Nearest Neighbour (KNN), Decision Tree (DT), and Gaussian Naive Bayes (GNB) for the purpose of classifying hand positions into open or closed states. Among the tested algorithms, SVM achieved the highest success rate with a maximum accuracy of 73.1%, while KNN yielded the lowest success rate at a minimum accuracy of 55.9%. To further enhance prediction accuracy in future studies, it is suggested that data from a larger set of muscles be collected.

Evaluation of Hand Action Classification Performance Using Machine Learning Based on Signals from Two sEMG Electrodes

“…However, a recent review [ 24 ] reporting the use of sEMG signals to classify hand gestures highlighted the lack of focus on two-sEMG-electrode systems in the field. Additionally, classification accuracy based on two-sEMG-electrode systems was reported [ 25 ] to vary by up to 20%, depending on machine learning algorithms utilized. Thus, it is important to compare different classification algorithms and feature types to facilitate comprehensive machine learning evaluation.…”