A convolutional autoencoder approach for high-density surface electromyogram decomposition

This is the authors' abstract. We don't add key points for this paper.

Deep learning methods have emerged as an alternative for decomposing high-density surface electromyogram (HD sEMG) signals. Most of these applications have relied on supervised approaches, which often require a priori training data from classical linear blind source separation algorithms. Here, we propose an unsupervised convolutional autoencoder for HD sEMG decomposition that incorporates sparsity and source-decorrelation regularization at the bottleneck, thereby enabling multi-layer architectures beyond the single-layer dense networks employed in a recent work. The method was evaluated on physiologically grounded synthetic HD sEMG generated under three contraction intensities and multiple noise levels, and on experimental HD sEMG recorded from the first dorsal interosseous muscle during low-intensity isometric contractions. The proposed algorithms were validated against a traditional linear unsupervised algorithm and a recent autoencoder-based algorithm. In synthetic data, the proposed deep learning approaches outperformed previously reported methods, yielding higher pulse-to-noise ratio (PNR) values, particularly in challenging conditions, such as increasing noise and contraction intensity. Decomposition of the experimental data showed that our autoencoder-based method yielded sources with higher pulse-to-noise ratios than the linear method, while maintaining comparable source counts and rate of agreement (RoA). These findings indicate that the proposed convolutional algorithm is more accurate than the linear approach for estimating motor unit spike trains. This advancement paves the way for the application of unsupervised deep learning methods in HD sEMG decomposition, particularly in settings where post-decomposition refinement by an expert should be minimized, such as in prosthetic control and human–machine interfaces.