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

OBJECTIVE

Electromyography (EMG)-driven musculoskeletal (MSK) models are widely used in biomechanics and rehabilitation, but they require muscle-specific (MSp) EMG recordings, which can be difficult to obtain from surface electrodes (sEMGs). This study aims to develop a novel framework to drive a MSK model using non-muscle-specific (NMSp) EMG recordings.

METHODS

The key innovation is a reinforcement learning (RL)-based solution that can train a policy to estimate MSp excitations for MSK model control inputs without precise sEMG placement. To validate the method, ten able-bodied individuals and one individual with a transradial amputation participated. NMSp EMGs were recorded simultaneously with intended wrist and hand motion. These data were used to validate how accurately the NMSp EMG-driven MSK model estimates joint motion via our new framework offline. In a second visit, an online evaluation of this framework was conducted as the participants performed a virtual posture matching task.

RESULTS

Offline and online performance showed that the RL estimated neural excitations yielded higher or equivalent motion estimation accuracy, compared to MSp EMG recordings, when driving the same MSK model. RL estimated versus actual neural excitations recorded from MSp EMG were similar in terms of co-contraction level and flexion/extension timing.

CONCLUSION

Using RL can be an alternative to MSp EMG placement to obtain muscle signals needed to drive MSK models.

SIGNIFICANCE: Our RL-based approach is a novel, effective solution to map NMSp EMG recordings to MSp neural excitation. This method may broaden future applications of MSK models when recording MSp EMG is difficult.