Robust Decomposition of Surface EMG Signals via Lightweight Deep Learning-Based Adaptation

Real-time surface electromyography decomposition has emerged as a promising way for neural interfacing. However, the decomposition performance faces dramatic degradation when multiple non-stationary factors coexist, including noise increases, new MU recruitments, and MU property variations. Here, we propose a deep learning-based (DL-based) adaptive decomposition method for moderate non-stationary scenarios, with an online adaptation strategy dynamically updating DL-based decomposition models. As prerequisite, the DL architecture is lightened through Tree-structured Parzen Estimator-based search to enable online adaptation. Additionally, multi-factor data augmentation was designed to enhance generalization capabilities. The proposed method outperforms blind source separation-based methods with noise increased by 5 dB, with F1-score of $0.715\pm 0.227$ vs. $0.388\pm 0.342$ and $0.633\pm 0.047$ vs. $0.202\pm 0.052$ for simulated and experimental data, respectively. Compared with DL-based methods, the proposed method is able to decode newly recruited MUs, with 11 and $6.93\pm 3.04$ MUs for simulated (total of 50 newly recruited MUs) and experimental data, respectively. Meanwhile, it performs better on initial MUs with property variations, with F1-score of $0.647\pm 0.082$ vs. $0.538\pm 0.156$ for experimental data. Regarding firing rates, the proposed method can identify a greater number of physiologically plausible MU spike trains. When deployed on an edge device, the proposed method met real-time decomposition ( $\leq 0.25$ s) and online adaptation latency constraints. The outcomes demonstrate the effectiveness of lightweight DL-based adaptation for non-stationary sEMG decomposition, thus paving the way for MU-based neural interfacing.