ViEMGT: A Novel Transformer Architecture for Enhanced Hand Gesture Recognition from HD-sEMG in Individuals with Traumatic Brain Injury

This paper proposes introduces the Vision-EMG Transformer (ViEMGT), a novel deep learning architecture designed for hand gesture recognition using high-density surface electromyography (HD-sEMG) in individuals with traumatic brain injury (TBI). ViEMGT reformulates multichannel EMG signals into structured spatio-temporal patches, enabling attention mechanisms to capture intricate neuromuscular patterns. On a dataset of 22 gesture classes from 12 individuals with TBI using 56 electrodes, the model achieved a 95.2% accuracy and a macro-averaged F1-score of 95.1%, surpassing CNN-LSTM, 1D-CNN, LSTM, and SVM baselines by up to 12% in accuracy and F1-score. The approach demonstrated robust performance across gesture types, with F1-scores of 0.955 for grip patterns and 0.930 for finger movements. Attention-based modeling enhanced interpretability and provided valuable insights into neuromuscular control. Data augmentation improved robustness, supporting ViEMGT’s application in real-world rehabilitation, prosthetic systems, and adaptive human-computer interfaces.