One-shot random forest model calibration for hand gesture decoding.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Existing machine learning models for myoelectric control demand extensive user-specific electromyographic (EMG) data for effective calibration.
• This data requirement presents a significant burden for new users of myoelectric devices.
• There is a need for efficient and low-data calibration methods for myoelectric control systems.

Purpose of the Study:

• To develop a novel approach for calibrating a pre-trained machine learning model with minimal data from new myoelectric users.
• To enable efficient and accurate user-specific adaptation of myoelectric control models.

Main Methods:

• A Random Forest (RF) model was initially trained on EMG data from 20 individuals performing various hand grips.
• For new users, the pre-trained decision trees were pruned using limited validation data.
• New decision trees trained solely on the new user's data were appended to the pruned model.

Main Results:

• Real-time experiments with 18 participants over two days showed the proposed approach significantly outperformed benchmark user-specific RF and linear discriminant analysis models in accuracy.
• The RF model calibrated on day one maintained significantly higher accuracy on day two compared to benchmarks, demonstrating robustness.
• The calibration procedure is source-free, requiring no access to the original training data after initial model pre-training.

Conclusions:

• The proposed source-free model calibration method effectively reduces the data burden for new myoelectric users.
• This approach enhances the accuracy and robustness of myoelectric control systems.
• The study advocates for the use of efficient, explainable, and simple models in myoelectric control applications.