Investigating Backpropagation Neural Networks for Joint Torque Prediction During Walking: A Gait Analysis Study

Assessing the overall health and condition of individuals across various age groups is enhanced by prior examination of gait analysis. Nonetheless, the conventional method of collecting gait data typically requires the physical presence of participants. In this study, a machine learning technique called the backpropagation neural network (BPNN) is employed to estimate lower-limb joint torques based on anthropometric and joint movements. The joint movements of 40 healthy participants (25 males, 15 females) aged 7–65 are collected using an existing motion capture setup. The input data consists of age, sex, total body mass, total height, thigh length, calf length, ankle-foot length, thigh mass, calf mass, ankle mass, hip joint, knee joint, and ankle joint angles of all participants. The output data contains the joint torques of the simplified leg model computed from the inverse dynamics formulation using the Euler-Lagrangian principle. The BPNN is trained using two optimization algorithms, Levenberg-Marquardt (LM) and resilient propagation (RP), and a comparative analysis reveals the notable capabilities of the LM-BPNN model in accurately estimating joint torques. The effectiveness of the BPNN models is further demonstrated through the estimation of joint torques for a female subject (60 years) from the testing dataset. The LM-BPNN model is found to be effective by 63.28%, 29.07%, and 4.23% for hip, knee, and ankle joint torque prediction, respectively.