Evaluation of probabilistic methods to predict muscle activity: implications for neuroprosthetics

Area of Science:

• Biomedical Engineering
• Neuroscience
• Rehabilitation Technology

Background:

• Functional electrical stimulation (FES) aims to restore motor function in paralyzed individuals by artificially activating muscles.
• Current FES prostheses have limited movement range due to challenges in identifying muscle activity patterns for complex behaviors.

Purpose of the Study:

• To evaluate three probability-based models for predicting electromyographic (EMG) activity during complex arm movements.
• To determine the efficacy of these models in enabling more sophisticated control of FES-based neuroprosthetics.

Main Methods:

• Tested Bayesian density estimation, polynomial curve fitting, and a dynamic neural network.
• Models predicted EMG activities of 12 arm muscles using hand trajectory data during 2D and 3D movements.

Main Results:

• The dynamic neural network model demonstrated superior prediction accuracy across most conditions.
• For 3D movements, the neural network model explained 40% of the variance in EMG signals.
• Predicted muscle activity showed an average root-mean-squared error of 6% compared to actual EMG signals.

Conclusions:

• Probabilistic models, particularly dynamic neural networks, can effectively predict muscle stimulation patterns for complex movements.
• These findings suggest a pathway for enhancing the capabilities of FES-based neuroprosthetics for improved motor function restoration.