Exploring the Relationship Between EMG Feature Space Characteristics and Control Performance in Machine Learning...

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Human-Computer Interaction

Background:

• Myoelectric machine learning (ML) control performance typically improves with user training.
• The underlying factors driving these performance enhancements in Electromyography (EMG) patterns remain poorly understood.
• Existing hypotheses suggest changes in EMG pattern characteristics, such as separability or repeatability, but empirical evidence is limited.

Purpose of the Study:

• To investigate the relationship between common EMG feature space metrics (separability, variability, repeatability) and the performance of ML myoelectric control.
• To determine if changes in these EMG metrics correlate with improvements in both offline and real-time control performance during a learning task.
• To assess the predictability of real-time control performance based on these EMG metrics.

Main Methods:

• Twenty able-bodied participants underwent 15 training blocks over 5 days to learn ML myoelectric control in a virtual environment.
• Offline and real-time control performance were assessed, alongside changes in three EMG metrics: separability, variability, and repeatability.
• Correlation analyses were performed between EMG metrics and performance, and L2-regularized linear regression was used to predict real-time performance.

Main Results:

• Real-time control performance significantly improved with training.
• No significant changes were observed in offline performance or any of the assessed EMG metrics (separability, variability, repeatability).
• A very low correlation was found between separability and real-time performance; other metrics showed no correlation. Real-time performance was not predictable from the EMG metrics.

Conclusions:

• The three common EMG feature space metrics (separability, variability, repeatability) do not appear to be directly related to real-time performance improvements in ML myoelectric control.
• Performance gains during myoelectric control training may be driven by factors other than the investigated EMG pattern characteristics.
• Further research is needed to identify the specific factors that govern performance improvements in myoelectric control systems.