Comparative Evaluation of the Time, Frequency, and Time‒Frequency Domain Features of EMG Signals for Neural Network-Based Classification of Upper Limb Kinematics

Prosthetic devices have significantly evolved from passive mechanical limbs to intelligent systems capable of mimicking natural movement. Modern prosthetics aim to restore lost functionality by integrating biosignals such as electromyography (EMG), which captures electrical activity generated by muscle contractions. The EMG serves as a vital interface between human intention and prosthetic action, enabling real-time control through signal interpretation. The effectiveness of EMG-based control systems depends largely on accurate feature extraction and robust machine learning classifiers. This research presents a comparative analysis of EMG signal features across three domains—time, frequency, and time-frequency—to determine the optimal approaches for prosthetic control. EMG data, acquired via both surface (sEMG) and intramuscular (iEMG) techniques, were collected from eight healthy male participants performing five distinct hand postures and four arm positions. Neural network classifiers, particularly narrow neural networks, were applied to assess classification accuracy under two conditions: fixed arm positions (FAPs) and fixed hand postures (FHPs). The results showed that time‒frequency domain features consistently outperformed those from the time and frequency domains. In the FAP scenario, the narrow neural network achieves a maximum accuracy of 99.09% at hand rest. In the FHP scenario, the same model reached 97.9% accuracy at a 135° arm angle. The observed performance hierarchy was time-frequency > frequency > time for the FAP and time-frequency > time > frequency for the FHP. These findings emphasize the potential of time-frequency domain features and neural network classifiers in enhancing the accuracy and efficiency of EMG-based prosthetic systems, contributing to more accessible and responsive assistive technologies.