Machine learning full 3-D lower-body kinematics and kinetics on patients with osteoarthritis from electromyography

Osteoarthritis (OA) is a degenerative disease that causes severe pain and reduces the range of motion of the joint, decreasing the quality of life for millions of individuals in the United States. Electromyography (EMG) sensors have been widely studied in biomechanics, showing applications in prosthetics, robotics, and control. While complex musculoskeletal models have been well established, the attempt of directly correlating EMG with kinematics and kinetics is still quite limited. Particularly, little work has been conducted on OA patients. In this work, we propose a method for estimating lower body joint angles (JAs) and ground reaction forces (GRFs) from surface-EMG sensors during a step-down task for individuals diagnosed with OA. The JAs and GRFs were measured by a Vicon motion capture system and force plates, respectively. EMG, JAs, and GRFs were used to train an echo state network (ESN) which afforded JAs with relative errors of 3.78% and 3.71% and the GRFs with relative errors of 3.619% and 4.596% for training and testing datasets, respectively. This study suggests the high fidelity of the ESN in automatically predicting full lower body kinematics and kinetics from the EMG signals. The results of this work promote the development of an EMG-controlled lower limb rehabilitation robot system for patients with OA.