Montage-Agnostic Encoder for Cross-User Gesture Recognition

Myoelectric prostheses, which promise control over many intended gestures, have largely remained confined to laboratory settings. A major hurdle is the poor transferability of recognition models from one user to another, typically necessitating extensive, labeled calibration from each end-user for practical performance. This research introduces a novel montage-agnostic encoder designed to overcome this limitation. The encoder processes each electrode with shared weights and identifies its location by physical coordinates rather than an index. This design allows a single architecture to accommodate any number of channels without requiring montage-specific parameters, simplifying deployment across diverse users and hardware configurations. When trained across multiple users, this encoder significantly outperforms traditional per-user Hudgins and linear-discriminant classifiers. For instance, it achieved a 0.234 macro-F1 improvement on DB1 and 0.108 on DB2 for held-out subjects. Ablation studies confirmed that each of the encoder's three core components contributes substantially to its performance. The study also found that while a sufficient training pool size is needed for convergence, the performance margin against baselines is more closely tied to the signal fidelity of the per-user baseline rather than just the number of training subjects. Interestingly, self-supervised pretraining offered no additional benefits once a supervised model was adequately trained.