Plug-and-play myoelectric control via a self-calibrating random forest common model.

Area of Science:

• Biomedical Engineering
• Rehabilitation Engineering
• Machine Learning

Background:

• Electromyographic (EMG) signal variability degrades myoelectric control performance over time.
• Traditional models require frequent recalibration, limiting long-term usability.
• EMG signal characteristics can change rapidly, necessitating adaptive solutions.

Purpose of the Study:

• To develop an automatic and unsupervised self-calibrating model for myoelectric control.
• To address the performance degradation of EMG-based control systems in long-term use.
• To create a robust and adaptable myoelectric control system requiring minimal calibration.

Main Methods:

• Developed a computationally efficient random forest (RF) model.
• Implemented a one-shot calibration for new users.
• Incorporated an unsupervised self-calibration mechanism using a data buffer and pseudo-labels for continuous model adaptation.
• Validated the model through extensive offline and real-time experiments, including long-term (5 weeks) and closed-loop studies with 66 participants.

Main Results:

• The self-calibrating RF model demonstrated improved performance in long-term myoelectric control.
• Bidirectional user-model co-adaptation was observed in closed-loop experiments.
• Users adapted to the dynamic model, reducing muscle effort (EMG amplitudes) for hand gestures.
• The model showed gradual performance enhancement over extended usage periods.

Conclusions:

• The proposed RF-based approach offers an explainable, computationally efficient, and data-minimal solution for myoelectric control.
• The self-calibrating mechanism enhances the robustness and reliability of EMG-based control systems.
• This method provides a viable alternative for improving long-term performance in myoelectric prosthetics and human-computer interfaces.