Stable Responsive EMG Sequence Prediction and Adaptive Reinforcement With Temporal Convolutional Networks.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Electromyographic (EMG) signal prediction for movement intentions typically uses frame-wise pattern recognition.
• This approach struggles with the time-varying nature of EMG signals, leading to unstable predictions.
• Temporal context is crucial for accurate EMG decoding.

Purpose of the Study:

• To demonstrate the superiority of sequential prediction models, specifically temporal convolutional networks (TCNs), for EMG-based movement intention prediction.
• To leverage temporal information within EMG signals for enhanced predictive performance.
• To compare TCNs against traditional frame-wise and other sequential models.

Main Methods:

• Compared temporal convolutional networks (TCNs) with frame-wise and other sequential models.
• Predicted 3 simultaneous hand and wrist degrees-of-freedom.
• Utilized data from 2 amputee and 13 non-amputee subjects in a minimally constrained experiment.
• Validated models on publicly available Ninapro and CapgMyo datasets.

Main Results:

• TCNs provided more accurate and stable predictions than frame-wise models, particularly during transitions between movements.
• Achieved an average response delay of 4.6 ms with simpler feature encoding.
• Performance of TCNs was further enhanced through adaptive reinforcement training.

Conclusions:

• Sequential models incorporating temporal EMG information significantly improve movement prediction.
• These models enable novel interactive training paradigms for prosthetic control.
• Treating EMG decoding as a sequential modeling problem enhances prosthesis control reliability, responsiveness, and movement complexity.