New sEMG model disentangles signal components for better gesture recognition

Surface electromyogram (sEMG) signals are widely used in human-machine interfaces for gesture recognition and user identification, but existing models struggle to generalize across individuals due to subject-specific neuromuscular differences. Researchers from a multi-institutional team (Yangyang Yuan, Jionghui Liu, Xinyu Jiang, et al.) have developed a new disentanglement model that explicitly separates task-specific and subject-specific components from raw sEMG data. This approach allows the model to learn gesture-invariant patterns across users while capturing individual neuromuscular signatures for identification. Experimental results show that the disentangled components significantly improve both gesture classification and user identification accuracy across different subjects and recording sessions, consistently outperforming conventional methods under the same conditions.

Further analysis reveals that task-specific components capture consistent muscle activation patterns for the same gestures across individuals, enabling robust cross-subject gesture recognition. In contrast, subject-specific components reflect unique neuromuscular characteristics, making them highly effective for user authentication. Notably, subject-specific components exhibit lower similarity across days than task-specific ones, leading to a greater drop in user identification accuracy compared to gesture recognition accuracy over time. This finding highlights the need for periodic recalibration in authentication systems. The model not only boosts performance but also provides deeper physiological insights, and the open-source code (available on arXiv) enables further research and practical deployment in rehabilitation, prosthetics, and secure user authentication.