Context-Informed Incremental Learning Improves Throughput and Reduces Drift in Regression-Based Myoelectric Control.

Area of Science:

• Biomedical Engineering
• Rehabilitation Robotics
• Human-Computer Interaction

Background:

• Current powered myoelectric prostheses rely on sequential, classification-based control.
• Regression-based control offers improved dexterity but suffers from inconsistent training and robustness issues due to a lack of realistic user behavior capture.

Purpose of the Study:

• To investigate the efficacy of context-informed incremental learning (CIIL) for robust regression-based myoelectric control in an unconstrained, velocity-based environment.
• To develop and compare adaptive models that account for user compliance and behavior against traditional training methods.

Main Methods:

• Two adaptive CIIL models (O-CIIL and T-CIIL) were developed and compared to traditional screen-guided training models.
• Sixteen participants performed an online Fitts' Law target acquisition task to evaluate model performance.
• Novel metrics, action interference and simultaneity gain, were introduced to quantify control stability and undesired simultaneous motions.

Main Results:

• Both adaptive CIIL approaches significantly outperformed non-adaptive models (p < 0.05) across multiple performance metrics.
• The T-CIIL model demonstrated superior performance in mitigating drift and action interference compared to the O-CIIL model.
• Findings indicate that incorporating user behavior is crucial for enhancing the stability and performance of regression-based myoelectric control systems.

Conclusions:

• Context-informed incremental learning (CIIL) is a viable approach for unconstrained, velocity-controlled regression-based myoelectric control.
• Accounting for user behaviors and compliance in training protocols is essential for overcoming robustness issues in myoelectric prostheses.
• The developed adaptive models and novel metrics offer a pathway to more intuitive and stable prosthetic control.