Multi-Joint Angles Estimation of Forearm Motion Using a Regression Model.

Area of Science:

• Biomedical Engineering
• Rehabilitation Robotics
• Machine Learning in Healthcare

Background:

• Improving the quality of life for forearm amputees necessitates accurate and robust prosthetic hands.
• Controlling prosthetic hands using surface electromyography (sEMG) signals presents significant challenges.
• Existing control methods often lack the precision required for intuitive prosthetic limb movement.

Purpose of the Study:

• To develop and evaluate a time-domain Convolutional Neural Network (CNN) model for predicting joint angles in a 3-DOF prosthetic hand.
• To assess the model's performance using cross-validation and analyze its learning patterns.
• To investigate the efficacy of transfer learning for daily model updates using limited new data.

Main Methods:

• A time-domain CNN model was designed for regression prediction of three prosthetic hand joint angles (wrist flexion/extension, pronation/supination, hand grip/open).
• Five-fold cross-validation was employed to evaluate prediction accuracy using correlation coefficient (CC).
• Transfer learning was applied to assess model adaptability with new data acquired on different days.

Main Results:

• The CNN model achieved high CC values for all three motions (e.g., wrist flexion/extension: 0.87-0.92).
• Transfer learning significantly improved CC values (e.g., wrist flexion/extension: 0.90-0.97), demonstrating effective daily model updates.
• The proposed CNN model outperformed four conventional regression models, both with and without transfer learning.

Conclusions:

• The developed time-domain CNN model offers a reliable method for predicting prosthetic hand joint angles from sEMG signals.
• Transfer learning enhances the model's adaptability and performance for real-time control across different days.
• This approach shows significant promise for future real-time prosthetic control applications, improving amputee quality of life.