Pre-training, personalization, and self-calibration: all a neural network-based myoelectric decoder needs

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Myoelectric control systems translate electromyographic signals (EMG) into movement intentions for prosthetics and robotics.
• High variability and temporal changes in EMG signals challenge system generalization, personalization, and adaptation.
• Deep neural networks require extensive user-specific data, limiting real-world application due to performance degradation.

Purpose of the Study:

• To develop an adaptive workflow for neural network training in myoelectric control.
• To improve and maintain decoding performance of electromyographic signals over extended periods.
• To address the limitations of current systems in handling EMG signal variability and user-specific patterns.

Main Methods:

• Proposed an innovative three-stage neural network training scheme.
• Implemented an adaptive workflow to progressively enhance network performance.
• Evaluated the system on 28 subjects over a 2-day period.

Main Results:

• Demonstrated significant improvement and maintenance of network performance across subjects and time.
• Validated the necessity and effectiveness of each stage within the proposed training framework.
• Showcased the potential for robust and adaptive myoelectric control systems.

Conclusions:

• The proposed three-stage training scheme effectively enhances adaptive myoelectric control.
• The adaptive workflow is crucial for maintaining performance despite EMG signal variability.
• This approach offers a promising solution for personalized and long-term use of EMG-based interfaces.