Leveraging Convolutional Sparse Autoencoders for Robust Movement Classification from Low-Density sEMG
This study addresses the practical limitations of clinical myoelectric prosthetic control, which are hindered by substantial inter-subject variability and the impracticality of high-density electrode arrays. The authors propose an end-to-end deep learning framework that operates on ultra-low-density (dual-channel) surface electromyography (sEMG) signals. By employing a convolutional sparse autoencoder to directly extract temporal features and integrating few-shot transfer learning with incremental learning strategies, the method drastically reduces user-specific calibration requirements while enabling dynamic expansion of gesture classes. Evaluated on a six-class gesture task, the approach achieves a multi-user F1-score of 94.3% ± 0.3%. For new users, performance improves from 35.1% ± 3.1% to 92.3% ± 0.9% with only minimal calibration data. Furthermore, when extended to ten gesture classes, the system maintains a robust F1-score of 90.0% ± 0.2%.