Predicting joint moments and angles from EMG signals

Electromyographic (EMG) signals are a direct reflection of muscle activity. Efforts have been made to extract additional information from them regarding the kinematic results of muscle use, with limited success. The purpose of this study was to conduct a wheelchair propulsion experiment using a limited number of subjects to gather information for developing a neural network model to correlate EMG signals to kinematic features of the wrist, elbow, and shoulder joints. The trained model was used to predict joint dynamics from previously unseen EMG data. EMG signals were collected from four muscle groups while five able-bodied subjects propelled a wheelchair. The procedure was videotaped with the kinematics and dynamics of the three joints obtained by digitizing and transforming the images into numerical data. A neural network model, based on the back-propagation algorithm, was built and trained on the collected data. The net's predicted values were close to the observed values, particularly for the joint moments (within 7% of actual values). The results demonstrate the potential of neural networks for predicting the movement patterns of wheelchair users from a small number of subjects.