Self-supervised representation learning with continuous training data improves the feel and performance of...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Traditional myoelectric control training uses static or ramp contractions, which do not reflect real-world dynamic movements.
• This limitation impacts the naturalness and effectiveness of pattern recognition-based prosthetic control.
• Developing dynamic training methods is crucial for enhancing prosthetic functionality.

Purpose of the Study:

• To investigate the benefits of training myoelectric control classifiers with continuous dynamic surface electromyography (sEMG) data.
• To compare the performance of conventional (LDA) and deep learning (LSTM) classifiers under different training conditions.
• To evaluate the impact of self-supervised learning (VICReg) pre-training on LSTM classifier performance.

Main Methods:

• Employed linear discriminant analysis (LDA) and long short-term memory (LSTM) classifiers.
• Trained classifiers using static ramp data, continuous dynamic movement data, and self-supervised pre-trained LSTM (VICReg).
• Assessed classifier performance using a Fitts' Law-inspired target acquisition task with 20 participants.

Main Results:

• LSTMs trained with continuous dynamic data significantly outperformed traditional LDA classifiers.
• Self-supervised VICReg pre-training further improved LSTM online performance and user experience.
• Continuous dynamic training captured transitions between movement classes more effectively than static ramp data.

Conclusions:

• Continuous dynamic sEMG data and temporal models (LSTMs) are superior for training myoelectric control.
• Self-supervised learning offers a promising avenue for enhancing prosthetic control systems.
• Findings support the development of more intuitive and user-friendly prosthetic devices through advanced training methodologies.