sEMG-Based Hand Movement Regression by Prediction of Joint Angles With Recurrent Neural Networks

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Accurate hand gesture recognition is crucial for advanced prosthetic limb control.
• Surface electromyography (sEMG) signals offer a non-invasive method for capturing muscle activity related to hand movements.
• Existing methods for translating sEMG to hand joint angles often lack the precision required for naturalistic control.

Purpose of the Study:

• To investigate and propose strategies for reliable prediction of hand joint angles from sEMG data.
• To enhance the performance of biosignal-based hand gesture recognition systems.
• To improve the acceptance and functionality of prosthetic hands for amputees.

Main Methods:

• Utilized a small-footprint recurrent neural network (RNN) to estimate hand joint positions from forearm sEMG data.
• Developed and evaluated post-processing methods to smooth predicted joint angle trajectories.
• Introduced a regularization term within the network's objective function to improve prediction stability.

Main Results:

• Both post-processing strategies and the regularization term demonstrated a positive impact on prediction accuracy.
• A maximal relative improvement of 6.13% in prediction accuracy was achieved.
• The regularization strategy offered improvements without introducing additional delay and allowed for flexible adjustment.

Conclusions:

• The proposed strategies, particularly regularization, significantly enhance the reliability of sEMG-based hand joint angle prediction.
• These advancements contribute to more intuitive and effective control of prosthetic hands.
• The regularization method presents a delay-free and adaptable solution for improving biosignal-based gesture recognition in medical applications.