Decoding motor intentions from electromyographic (EMG) signals holds transformative potential for neurorehabilitation and assistive technologies, yet existing approaches remain fundamentally constrained by limbspecific architectures, poor generalization across individuals and recording sessions, and an unsustainable dependence on large volumes of labeled data. Here we present a unified cross-limb framework that, for the first time, systematically addresses both upper- and lower-limb motor decoding within a single, consistent learning paradigm. The framework operates through two complementary levels of generalization: at the data level, a principled feature selection process identifies domain-invariant representations transferable across heterogeneous EMG datasets; at the architecture level, a Reptile-based meta-learning mechanism enables rapid few-shot adaptation through self-supervised confidence-based pseudo-labeling, eliminating the need for ground-truth annotations during deployment. A bias-aware sampling strategy further stabilizes iterative adaptation by enforcing balanced class distributions under low-data conditions, mitigating confirmation bias in self-supervised convergence. Evaluated across nine public benchmarks spanning 170 subjects, the framework achieves state-of-the-art decoding accuracy under inter-session, intersubject, and inter-dataset conditions for both limb types, while retaining subject-specific adaptability without repeated recalibration, offering a scalable and clinically viable pathway toward comprehensive motor-intention decoding in real-world rehabilitation settings.
Lee et al. (Thu,) studied this question.