SRM: A Source-Reprojection Module for Cross-Day sEMG Gesture Recognition

Surface electromyography (sEMG) gesture recognition degrades across recording days under domain shift, increasing calibration burden for myoelectric interfaces. Many cross-day adaptation pipelines retrain the deployed recognizer or require labeled target-session data, which can be impractical in assistive-device settings where classifier versions may need to remain locked for traceability and regulatory compliance. We study unsupervised cross-day adaptation under two constraints: the task classifier remains frozen and holdout-day labels are not used when training the adaptor. We propose the Source-Reprojection Module (SRM), a plug-in front end that combines conditional adversarial feature learning with a residual signal-space projector guided by the frozen classifier's gradients, identity regularization, and latent-space distribution matching, using labeled source days and unlabeled adaptation days only. On a multi-day protocol with four healthy participants (at least five calendar-day sessions per participant, split 3:1:1 into source, adaptation, and holdout domains) and three random seeds per participant (12 runs), mean holdout accuracy increases from 70.9% for the frozen classifier alone to 72.8% with SRM (+1.98±0.91 percentage points averaged across subjects). SRM outperforms the frozen baseline in 10 of 12 subject-seed runs. The gain is modest and the cohort is small, so the result supports proof-of-mechanism under the stated protocol rather than population-level clinical generalization.