Recalibration of myoelectric control with active learning

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Machine Learning

Background:

• Myoelectric control robustness is crucial for reducing prosthesis abandonment, as current decoders degrade over time.
• Existing recalibration methods face scalability and computational challenges.
• Active learning offers a scalable, human-in-the-loop solution for improving myoelectric control adaptation.

Purpose of the Study:

• To investigate active learning as a framework for enhancing the long-term robustness of myoelectric control.
• To evaluate different active learning sampling strategies for optimizing decoder adaptation.
• To reduce the need for frequent recalibration in myoelectric prosthetic devices.

Main Methods:

• Employed active learning and linear discriminant analysis for an iterative learning process.
• Simulated a real-time scenario using least confidence, smallest margin, and entropy reduction sampling strategies.
• Investigated single-mode and batch-mode sample selection, including ranked batch-mode active learning.

Main Results:

• Active learning achieved superior decoder performance compared to random sampling with minimal data (3.2 min).
• Demonstrated significant enhancement in decoder adaptation and optimized training data selection.
• Identified smallest margin and least confidence as the most effective uncertainty sampling strategies.

Conclusions:

• Introduced a novel active learning framework for long-term adaptation in myoelectric control.
• Validated the framework's effectiveness in a simulated closed-loop environment.
• Proposed a pipeline for future real-time deployment of adaptive myoelectric control systems.