A Montage-Agnostic Encoder for Calibration-Light Cross-User Gesture Recognition from Surface Electromyography

Surface electromyography (sEMG) gesture recognition typically requires extensive, time-consuming per-user calibration to account for anatomical differences and varying electrode placements. This paper investigates whether a deep learning encoder can be trained across a large pool of users to provide a robust, calibration-light system that generalizes to new users and different hardware configurations without requiring montage-specific adjustments.

The authors propose a montage-agnostic encoder that treats each electrode as an independent token. Key architectural innovations include:

• Causal Rolling-Time Normalization: Adapts to new users at runtime using only past signal statistics, eliminating the need for user-specific calibration data.
• Electrode-Coordinate Encoding: Maps electrodes to their physical forearm positions rather than fixed indices, allowing the model to ingest varying channel counts (8–16 channels) using shared weights.
• Cross-Channel Attention: A transformer-based mechanism that learns spatial relationships between electrodes to improve feature extraction.

The model was evaluated using a leave-one-subject-out protocol across four public datasets (NinaPro DB1, DB2, DB5, and EMG-EPN612), comparing its performance against a standard per-user Linear Discriminant Analysis (LDA) baseline at varying calibration budgets (0, 1, and 3 repetitions).

The proposed encoder demonstrates superior cross-user performance compared to the traditional LDA baseline. On the NinaPro DB1 dataset, the encoder achieved a macro-F1 score of 0.827 at three calibration repetitions, outperforming the LDA baseline by 0.234. Ablation studies confirmed that the cross-channel attention mechanism, rolling-time normalization, and electrode-coordinate encoding are all critical, with each component contributing significantly to the model's overall capacity. Interestingly, the study found that the size of the training pool acts primarily as a stability floor; once a sufficient number of subjects is reached, performance plateaus, suggesting that signal fidelity is the primary driver of performance differences across datasets.