AEMG: Self-Supervised Learning Framework for Generalizable EMG Representations

A team of researchers has introduced Any Electromyography (AEMG), marking the first extensive, self-supervised framework for learning representations of EMG signals. This innovative approach reinterprets neuromuscular dynamics in a linguistic manner through a Neuromuscular Contraction Tokenizer (NCT), which converts muscle contractions into structural terms and temporal activation patterns into sentences. AEMG boasts the largest vocabulary of cross-device EMG signals, facilitating smooth transfer across various channel topologies and sampling rates. Experimental results indicate that AEMG enhances zero-shot leave-one-out generalization among subjects, devices, and tasks, effectively tackling data heterogeneity and the issue of limited labels. This research is available on arXiv with the identifier 2605.03462.