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

Abstract

Functional electrical stimulation (FES) involves artificial activation of muscles with surface or implanted electrodes to restore motor function in paralyzed individuals. Currently, FES-based prostheses produce only a limited range of movements due to the difficulty associated with identifying patterns of muscle activity needed to evoke more complex behaviour. Here we test three probability-based models (Bayesian density estimation, polynomial curve fitting and dynamic neural network) that use the trajectory of the hand to predict the electromyographic (EMG) activities of 12 arm muscles during complex two- and three-dimensional movements. Across most conditions, the neural network model yielded the best predictions of muscle activity. For three-dimensional movements, the predicted patterns of muscle activity using the neural network accounted for 40% of the variance in the actual EMG signals and were associated with an average root-mean-squared error of 6%. These results suggest that such probabilistic models could be used effectively to predict patterns of muscle stimulation needed to produce complex movements with an FES-based neuroprosthetic.