Prediction of Running Anteroposterior Ground Reaction Force From the Vertical Component Through a Deep Learning Approach.

Studying ground reaction force (GRF) is essential for in situ running analysis, as they provide critical insights into both performance optimization and the etiology of running-related injuries. However, the use of current measurement systems in ecological or field conditions remains limited: their restricted measurement areas constrain data collection to only a few steps, and their high cost hinders widespread adoption. Recently, a size-adjustable instrumented track has been developed to measure the vertical component of GRF outside the laboratory environment. Building on this advancement, the present study aims to develop a neural network approach for predicting the anteroposterior component from the measured vertical component and to evaluate the impact of prediction errors on joint kinetics. A public data set comprising 9500 running steps was used to train, validate, and test the prediction model. The results demonstrated that a convolutional neural network can accurately predict the anteroposterior component of the GRF (relative root mean square error = 7.44%). While these anteroposterior-GRF prediction errors had only a minor impact on ankle moment calculations, estimates at the knee and hip were more substantially affected. These findings suggest that the proposed model may be effectively applied in conjunction with the size-adjustable instrumented track.