Improving Myoelectric Pattern Recognition Robustness to Electrode Shift by Autoencoder.

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Signal Processing

Background:

• Electrode shift during prosthetic socket use degrades myoelectric pattern recognition accuracy.
• This limitation hinders real-world application of advanced myoelectric control systems.
• Robust feature extraction is crucial to overcome electrode displacement issues.

Purpose of the Study:

• To propose an unsupervised feature extraction method, sparse autoencoder (SAE), to address electrode shift in high-density electromyography (HD-EMG).
• To evaluate the effectiveness of SAE in maintaining classification accuracy despite electrode displacement.
• To compare SAE performance against traditional time-domain and autoregressive (TDAR) features.

Main Methods:

• Implemented a sparse autoencoder (SAE) for unsupervised feature extraction from HD-EMG signals.
• Evaluated the SAE algorithm on nine intact-limbed subjects and one amputee.
• Compared classification accuracy and sensitivity to electrode shift (±1 cm) with TDAR features.

Main Results:

• SAE demonstrated lower classification error compared to TDAR features when no shift occurred.
• SAE significantly reduced sensitivity to electrode shifts of ±1 cm.
• SAE showed robustness to shift directions perpendicular to muscle fibers.

Conclusions:

• SAE is an effective method for extracting robust spatial structures and correlations from HD-EMG.
• The proposed SAE method significantly improves the resilience of myoelectric control to electrode shifts.
• This advancement holds promise for real-world applications of pattern recognition-based myoelectric control systems.