Utilizing Reinforcement Learning to Overcome the Challenge of Muscle-Specific EMG Placements for Musculoskeletal...

Area of Science:

• Biomechanics
• Rehabilitation Engineering
• Machine Learning in Healthcare

Background:

• Electromyography (EMG)-driven musculoskeletal (MSK) models are crucial in biomechanics and rehabilitation.
• Obtaining muscle-specific (MSp) EMG recordings from surface electrodes (sEMGs) is often challenging.
• This limits the widespread application of MSK models.

Purpose of the Study:

• To develop a novel framework for driving MSK models using non-muscle-specific (NMSp) EMG recordings.
• To overcome the limitations of acquiring MSp EMG data.
• To enhance the applicability of MSK models in clinical and research settings.

Main Methods:

• A reinforcement learning (RL) based approach was developed to estimate MSp excitations from NMSp EMG signals.
• The framework was validated offline and online with ten able-bodied individuals and one with a transradial amputation.
• Participants performed virtual posture matching tasks to evaluate the framework's performance.

Main Results:

• The RL-estimated neural excitations achieved higher or equivalent motion estimation accuracy compared to MSp EMG recordings.
• The RL approach demonstrated similar co-contraction levels and flexion/extension timing as MSp EMG.
• Both offline and online evaluations confirmed the framework's effectiveness.

Conclusions:

• Reinforcement learning (RL) provides a viable alternative to MSp EMG placement for driving MSK models.
• This novel RL-based method effectively maps NMSp EMG recordings to MSp neural excitation.
• The approach broadens the potential applications of MSK models, especially when MSp EMG acquisition is difficult.