Comparison of Ground Reaction Forces and Net Joint Moment Predictions: Skeletal Model Versus Artificial Neural Network-Based Approach.

Abstract

Artificial neural networks (ANNs) are becoming a regular tool to support biomechanical methods, while physics-based models are widespread to understand the mechanics of body in motion. Thus, this study aimed to demonstrate the accuracy of recurrent ANN models compared with a physics-based approach in the task of predicting ground reaction forces and net lower limb joint moments during running. An inertial motion capture system and a force plate were used to collect running biomechanics data for training the ANN. Kinematic data from optical motion capture systems, sourced from publicly available databases, were used to evaluate the prediction performance and accuracy of the ANN. The linear and angular momentum theorems were applied to compute ground reaction forces and joint moments in the physics-based approach. The main finding indicates that the recurrent ANN tends to outperform the physics-based approach significantly (P < .05) at similar and higher running velocities for which the ANN was trained, specifically in the anteroposterior, vertical, and mediolateral ground reaction forces, as well as for the knee and ankle flexion moments, and hip abduction and rotation moments. Furthermore, this study demonstrates that the trained recurrent ANN can be used to predict running kinetic data from kinematics obtained with different experimental techniques and sources.