Comparing Reinforcement Learning Agents and Supervised Learning Neural Networks for EMG-Based Decoding of Continuous...

Area of Science:

• Biomechanics
• Machine Learning
• Neuroscience

Background:

• Electromyography (EMG)-based continuous joint kinematics decoding utilizes model-based (e.g., musculoskeletal modeling) and model-free (e.g., supervised learning neural networks - SLNN) approaches.
• Existing methods face challenges in robustness and adaptability to varying movement dynamics.

Purpose of the Study:

• To present a new kinematics decoding framework utilizing reinforcement learning (RL).
• To combine machine learning and model-based approaches for enhanced EMG-based kinematics decoding.
• To compare the performance and robustness of the RL framework against the SLNN approach.

Main Methods:

• Collected electromyography (EMG) and kinematic data from 5 able-bodied subjects performing simultaneous metacarpophalangeal (MCP) and wrist joint flexion/extension at slow and fast tempos.
• Trained a reinforcement learning (RL) agent and a supervised learning neural network (SLNN) for each tempo.
• Tested all trained agents and SLNNs using both fast and slow kinematic data.

Main Results:

• The RL-based kinematics decoder exhibited greater robustness to changes in movement speeds between training and testing datasets.
• The RL decoder demonstrated superior performance metrics compared to the SLNN approach.
• Performance was evaluated using Pearson's correlation coefficient (r) and normalized root mean square error (NRMSE) for joint angle estimation.

Conclusions:

• The proposed RL-based framework offers a more robust and effective method for EMG-based continuous joint kinematics decoding.
• This approach holds promise for improving the accuracy and adaptability of prosthetic and assistive device control.
• Combining machine learning with model-based principles via RL enhances decoding capabilities beyond traditional SLNN methods.