Hand Gesture Intention Detection Using sEMG and Transfer Learning in Stroke Survivors

Surface electromyography (sEMG) has shown potential for intuitive control of wearable assistive technologies in individuals with motor impairments. However, accurately classifying sEMG signals in stroke patients, particularly those with severe impairments and no visible voluntary hand movement, is difficult due to altered neuromuscular patterns. While robust machine learning models require large and diverse datasets, high variability among stroke patients means that even extensive data may fail to capture individual differences. To address this challenge, this study applies a transfer learning approach that leverages richer sEMG signals from healthy individuals to pretrain a model, which is then adapted to the altered patterns of stroke patients. This process enables effective classification even with limited patient-specific data. Ten stroke patients wore a Myo armband, a wearable sEMG acquisition device, and performed hand-closing, rest, and hand-opening gestures across up to 26 arm postures. Healthy sEMG data were obtained from the Ninapro DB5 dataset. Multiple transfer learning configurations were evaluated by varying the number of frozen layers. The best-performing configuration achieved an overall accuracy of 93.6% across all participants, with 84.32% for severe, 97.42% for moderate, and 99.07% for low impairments. In comparison, training the model solely on stroke data without transfer learning resulted in a considerably lower accuracy of 46%. These findings highlight the value of incorporating healthy sEMG data for pattern recognition in stroke patients and demonstrate the potential of transfer learning to support inclusive and adaptive wearable sEMG-based control systems for individuals with varying levels of post-stroke motor impairment.