Artificial neural networks for HD-sEMG-based hand position estimation: addressing inter- and intra-subject...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Human-Computer Interaction

Background:

• High-Density surface Electromyography (HD-sEMG) offers potential for controlling rehabilitation and assistive devices.
• Current HD-sEMG control systems lack robustness due to electrode shifts, skin condition changes, and user variability.
• This limits the reliable deployment of muscle-computer interfaces in real-world scenarios.

Purpose of the Study:

• To evaluate the robustness of the Recursive Prosthetic Control Network (RPC-Net) with a High-Density Electrode Array (HDE-Array) system.
• To assess system performance under realistic usage conditions, including electrode repositioning and cross-subject generalization.
• To determine if incorporating variability during training improves system performance.

Main Methods:

• The RPC-Net/HDE-Array system was tested under static training conditions versus conditions simulating real-life usage.
• Performance was evaluated with and without accounting for electrode repositioning during training.
• Cross-subject generalization was assessed by training on data from multiple subjects.

Main Results:

• The RPC-Net/HDE-Array system showed high sensitivity to electrode repositioning and skin variability when trained statically.
• Robustness significantly improved when training data included realistic variability, such as electrode shifts.
• Performance enhanced with more subjects in the training pool, excluding the tested subject, indicating subject-specific pattern dependency.

Conclusions:

• The RPC-Net/HDE-Array system can achieve robust performance across different sessions and users when trained under realistic conditions.
• Training with variability is crucial for mitigating performance degradation.
• This research is a significant advancement toward the practical application of muscle-computer interfaces.