Myoelectric temporal patching: leveraging long temporal memory for effective sEMG pattern recognition

Robust myoelectric control is essential for advancing lower limbs active assistive devices and prosthetics. Feature extraction from surface Electromyographic (sEMG) signals serves as a fundamental component influencing the performance of such controllers. While deep learning models, such as Long Short-Term Memory (LSTM) networks, have been increasingly used to extract hidden features from raw sEMG data by leveraging temporal contexts in laboratory studies, the use of such models on prosthetic controllers is often associated with large memory and computational overheads, thus necessitating expensive hardware. To address these challenges, this study proposes a novel handcrafted feature extraction method, termed Myoelectric Temporal Patching (MTP), that works on extracting multi-signal features from patches within the short-segments of sEMG and propagating information across the windows, to capture both short- and long-term temporal dynamics of sEMG signals, without the high computational burden. Two pattern recognition experiments were conducted to validate the proposed method: gait phase recognition using the SIAT-LLMD dataset and locomotion mode recognition using the MyoPace dataset. Results demonstrated that the MTP feature set consistently outperformed (p < 0.0001) traditional and spatial feature sets across three machine learning models—Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM). The proposed method achieved peak accuracies of 85.11% in gait phase recognition and 87.70% in locomotion mode recognition when using SVM. These findings underscore the effectiveness of the proposed temporal features in decoding lower-limb motion intentions, emphasizing the critical role of temporal dynamics of sEMG signals.