Deep Learning for Robust Decomposition of High-Density Surface EMG Signals

Area of Science:

• Biomedical Engineering
• Signal Processing
• Neuroscience

Background:

• Blind source separation (BSS) algorithms like gCKC decompose high-density surface electromyography (HD-sEMG) signals into motor unit (MU) action potential trains.
• Existing methods require computationally expensive data whitening and are sub-optimal in noisy environments.

Purpose of the Study:

• To explore a supervised learning approach using a gated recurrent unit (GRU) network for HD-sEMG decomposition.
• To evaluate the GRU network's performance against the gCKC algorithm, particularly in noisy conditions and with unwhitened data.

Main Methods:

• A GRU network was trained using paired HD-sEMG signals and BSS outputs.
• The trained GRU model decomposed simulated and experimental unwhitened HD-sEMG signals.
• Performance was validated against concurrently recorded intramuscular EMG signals.

Main Results:

• The GRU network outperformed gCKC at low signal-to-noise ratios.
• GRU demonstrated superior generalization to new data compared to gCKC.
• GRU achieved high agreement rates (92.5%) with intramuscular sources, comparable to gCKC (94.9%).

Conclusions:

• GRU networks offer a promising alternative for HD-sEMG decomposition, especially in challenging noisy conditions.
• This supervised learning approach reduces computational cost by eliminating the need for data whitening.
• GRU networks show potential for improved accuracy and robustness in analyzing EMG signals.