Data Management for Transfer Learning Approaches to Elbow EMG-Torque Modeling.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Wearable Sensors

Background:

• Subject-specific electromyogram (EMG)-torque models show poor performance on new subjects or re-tests.
• Improving model generalization is critical for practical applications of EMG-based control.

Purpose of the Study:

• To investigate data acquisition and utilization strategies for enhancing cross-subject EMG-torque model performance.
• To propose and evaluate a novel unsupervised data weighting method for model calibration.

Main Methods:

• Analyzed data from 65 subjects to determine the impact of training set diversity versus size.
• Developed a correlation-based data weighting (COR-W) method to assess domain shift using EMG signals.
• Applied weighted least squares for model calibration using COR-W assigned data weights.

Main Results:

• Training set data diversity (number of subjects) is more critical than data size for model performance.
• The COR-W method achieved a low root mean square error (9.29% MVC) in cross-subject evaluation.
• Significant performance improvements were observed compared to uncalibrated models.

Conclusions:

• Data acquisition diversity and the COR-W utilization strategy effectively improve EMG-torque model generalization.
• The COR-W method offers an unsupervised approach for calibrating EMG-torque models in cross-subject scenarios.