Ensemble Learning Method for the Continuous Decoding of Hand Joint Angles.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Human-machine interface (HMI) technology for prosthetics is limited by motion decoding accuracy.
• Simultaneous and proportional control (SPC) is crucial for enhancing the dexterity and functionality of smart prostheses.
• Accurate decoding of joint angles from surface electromyography (sEMG) signals is a key challenge.

Purpose of the Study:

• To develop and evaluate an ensemble learning approach for decoding metacarpophalangeal (MCP) joint angles from sEMG signals.
• To compare the performance of various ensemble models against traditional methods like Gaussian process models.
• To establish a comprehensive pipeline for high-performance hand motion recognition using ensemble learning.

Main Methods:

• Designed and tested seven distinct ensemble learning models for angle decoding from sEMG data.
• Recorded sEMG signals during functional tasks to estimate kinematics of five MCP joint angles.
• Evaluated model performance using Pearson correlation coefficient (CC) and Root Mean Square Error (RMSE).

Main Results:

• A combined CatBoost and LightGBM ensemble model demonstrated superior performance, achieving an average CC of 0.897 and RMSE of 7.09.
• The ensemble learning approach significantly outperformed the Gaussian process model in decoding accuracy across all test scenarios.
• The proposed pipeline offers a robust system for hand motion recognition with reduced parameter and data requirements compared to deep learning.

Conclusions:

• Ensemble learning provides a powerful and efficient method for high-accuracy angle decoding from sEMG signals.
• The developed system has significant potential for enabling simultaneous and proportional control (SPC) in next-generation prosthetic hands.
• This research offers a practical framework for engineers and researchers to implement advanced prosthetic control systems.