Recognition and Label-Free Adaptation Across Recording Sessions in Surface-EMG Gesture Decoding

Surface-EMG gesture recognition systems often suffer from significant performance degradation when electrodes are removed and reapplied, a phenomenon known as inter-session variability. This paper investigates whether a previously developed montage-agnostic encoder, originally designed for cross-user and cross-montage transfer, can maintain robust gesture recognition across different recording sessions without requiring time-consuming recalibration.

The authors evaluate a montage-agnostic encoder on the NinaPro DB6 dataset, which includes ten subjects recorded across two separate sessions. The model uses causal rolling-time normalization, a shared per-channel convolutional tokenizer, and cross-channel transformer attention to map EMG signals into a uniform embedding space. The researchers test the encoder's performance in a cross-session regime—where the model is trained on one session and tested on another without any retraining—and compare it against a standard per-user LDA pipeline and other published source-only baselines. Additionally, they explore five label-free test-time adaptation methods to see if the model can recover performance using only unlabeled data from the new session.

The encoder demonstrates superior robustness to session changes, retaining a macro-F1 score of 0.688 compared to 0.540 for the per-user LDA pipeline. While the encoder experiences a performance drop when moving between sessions, it consistently outperforms the traditional LDA baseline for nine out of ten subjects. Among the tested label-free adaptation methods, feature-statistic alignment—which rescales the target session's embeddings to match the source session's statistics—consistently improves performance. Conversely, batch-normalization re-estimation, a common domain-adaptation technique, was found to be detrimental, causing the model to collapse to a single class.