Lower Limb Joint Torque Estimation via Bayesian Regularized Backpropagation Neural Networks

Prior gait analysis has played a crucial role in assessing overall health and well-being, particularly in individuals of different age groups. Nevertheless, acquiring gait data has necessitated the in-person presence of participants, incurring potential expenses. We introduce a deep learning technique called Bayesian regularized backpropagation neural network (BR-BPNN) to estimate joint torques in lower-limb motion based on anthropometric data and joint movements. The study collects joint movement data from 40 healthy participants (25 males and 15 females) spanning ages 7 to 65, utilizing an established motion-capture system. The input data encompassed various parameters such as age, gender, total body mass, height, thigh length, calf length, ankle-foot length, thigh mass, calf mass, ankle mass, and joint angles for the hip, knee, and ankle. The output data included joint torques calculated for a simplified leg model using the inverse dynamics method derived from the Euler-Lagrangian principle. We presented the conceptual details of the proposed BR-BPNN model and contrast model named Levenberg-Marquardt-based BPNN (LM-BPNN). Performing a comparative analysis, the results revealed that the BR-BPNN model (R
train
= 0.985, R
test
= 0.955) excelled in accurately estimating joint torques compared to the LM-BPNN model (R
train
= 0.973, R
test
= 0.924). Furthermore, the BR-BPNN model is found to be effective by 87.19%, 77.87%, and 72.73% in estimating hip, knee, and ankle joint torques for a 51-year-old male subject from the testing dataset.