Prototype-Guided Test-Time Adaptation for Real-Time EMG-Based Gesture Recognition

Surface electromyography (sEMG) signals facilitate intuitive human-machine interaction and are widely employed in applications such as prosthetics and rehabilitation. However, sEMG signal distributions are highly sensitive to variations in electrode placement, user physiology, and session conditions, which significantly hinder model generalization in real-world settings. To address this, we propose EMG-based structure-aware test-time adaptive refinement (EMG-STAR), a test-time adaptation (TTA) framework tailored for online deployment without access to target data during training. EMG-STAR improves model stability by fusing historical and batch-wise normalization statistics from both source and target domains, effectively alleviating domain shifts. Furthermore, it refines predictions using a prototype-assisted feature clustering method that enforces consistency among semantically similar target samples. To further enhance reliability, the framework jointly leverages prediction confidence and distributional alignment to filter out noisy pseudo-labels during adaptation. Comprehensive evaluations conducted on two public datasets, Hyser, and NinaPro-DB2, demonstrate that EMG-STAR consistently outperforms ten existing TTA methods. Interpretability analysis using Grad-CAM reveals that EMG-STAR predominantly attends to physiologically meaningful muscle activation regions, suggesting that the model captures neuromuscular structures rather than transient signal noise.