Decision-Change Informed Rejection Improves Robustness in Pattern Recognition-Based Myoelectric Control.

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Machine Learning for Prosthetics

Background:

• Post-processing techniques enhance pattern-recognition-based myoelectric control decision streams.
• Existing methods are often evaluated individually on stationary data, limiting understanding of dynamic performance trade-offs.
• Evaluating smoothing vs. latency in dynamic scenarios is crucial for real-world applications.

Purpose of the Study:

• To survey and compare eight established post-processing schemes for myoelectric control.
• To introduce and evaluate two novel, temporally aware post-processing methods.
• To assess the performance of these techniques during continuous and dynamic class transitions.

Main Methods:

• Comparison of eight post-processing techniques: majority vote, Bayesian fusion, onset locking, outlier detection, confidence-based rejection, confidence scaling, prior adjustment, and adaptive windowing.
• Development and implementation of two new temporally aware schemes: decision-change informed rejection (DCIR) and a related approach.
• Evaluation using both conventional and deep classifiers on dynamic datasets with continuous class transitions.

Main Results:

• The proposed decision-change informed rejection (DCIR) approach demonstrated superior performance compared to existing schemes.
• DCIR reduced error rates and decision stream volatility during both steady-state conditions and dynamic transitions.
• The effectiveness of DCIR was consistent across both conventional and deep classifier types.

Conclusions:

• Leveraging temporal context in post-processing significantly enhances the robustness of myoelectric control systems.
• Temporally aware methods, like DCIR, offer improved performance by better rejecting uncertain decisions during dynamic use.
• The findings suggest a new direction for developing more reliable and responsive myoelectric control interfaces.