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

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Real-time surface electromyography (sEMG) decomposition is crucial for neural interfacing.
• Performance degrades significantly with non-stationary factors like noise, new motor unit (MU) recruitment, and MU property variations.

Purpose of the Study:

• To develop a deep learning-based (DL-based) adaptive decomposition method for moderate non-stationary sEMG scenarios.
• To enable dynamic online adaptation of DL models for improved decomposition accuracy.

Main Methods:

• Lightened DL architecture using Tree-structured Parzen Estimator-based search for efficient online adaptation.
• Multi-factor data augmentation to improve model generalization capabilities.
• Online adaptation strategy for dynamically updating DL decomposition models.

Main Results:

• Outperformed blind source separation methods in noisy conditions (5 dB increase) with higher F1-scores on simulated and experimental data.
• Successfully decoded newly recruited MUs (11 simulated, 6.93 experimental) and improved performance on MUs with property variations (F1-score 0.647 vs. 0.538).
• Identified more physiologically plausible MU spike trains and met real-time decomposition constraints (≤ 0.25 s) on edge devices.

Conclusions:

• Lightweight DL-based adaptation is effective for non-stationary sEMG decomposition.
• The proposed method enhances neural interfacing capabilities by improving MU decoding accuracy and real-time performance.
• This approach paves the way for advanced MU-based neural interfaces in challenging environments.