Estimation of Neuromuscular Activities Using Gait Analysis and Deep Learning for Rehabilitation Purposes

Abstract

Gait analysis using modern motion tracking techniques including measurement of kinematic variables is an important modality in rehabilitation research and applications. Functional electrical stimulation (FES) for patients with paralysis and cerebral diseases is one of the most important applications of rehabilitation science. Efficient muscle stimulation requires a pre-knowledge about limb motility and muscles synergy. In this paper, we are working to track the angle changes of the thigh and shin during walking phases based on accelerometer and gyroscope sensors, and estimating the thigh-shin angle and its derivative using HuGaDB dataset. Those three features are used with a feedforward neural network (FNN) to determine the activity of the rectus femoris muscle by pre-training of neural network with gait analysis as input and electromyography (EMG) signal as the output of the same patient. The results illustrate the ability of FNN to reproduce EMG for each gait cycle of the same patient with average precision equal to 96% as training and 92.5% as testing. The proposed method presents a good tool for FES systems, especially for EMG encoding stages.