Abstract
Electromyographic (EMG) signals serve as critical indicators of neuromuscular dynamics, with growing applications in clinical diagnostics and wearable neuroprosthetics. Despite their widespread use, there is a critical gap in the systematic evaluation of diverse learning paradigms under varying data conditions. This study proposes a comprehensive comparative framework to evaluate supervised, self-supervised, and unsupervised deep learning models for EMG signal classification using data from both neurologically intact and post-stroke individuals. Using the U-Limb dataset, we implement representative architectures for each paradigm: (i) a long short-term memory (LSTM)–based classifier, (ii) a SimCLR-based representation learning framework followed by a random forest classifier, and (iii) an autoencoder optimized with structural similarity index measure (SSIM) and latent space regularization losses. Model performance is evaluated in both in-distribution (ID) and out-of-distribution (OOD) scenarios to examine robustness to inter-subject variability. Results show that while the supervised model achieved near-perfect ID accuracy (> 0.98), it exhibited unstable OOD performance (0.4–0.8). In contrast, the self-supervised model demonstrated superior generalization, achieving OOD accuracy above 0.99 when trained with 18 subject pairs. The unsupervised autoencoder, however, performed near random chance, indicating limited feature learning capability. The self-supervised approach shows strong potential for robust EMG classification in label-scarce and distribution-shifted scenarios.