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

Area of Science:

• Biomechanics
• Sports Science
• Machine Learning

Background:

• Ground reaction forces (GRF) are crucial for analyzing running biomechanics, performance, and injury mechanisms.
• Current GRF measurement systems are limited by small collection areas and high costs, restricting field use.
• A novel size-adjustable instrumented track allows for vertical GRF measurement outside laboratory settings.

Purpose of the Study:

• To develop a neural network model to predict the anteroposterior (AP) component of GRF using only the vertical component.
• To assess the impact of AP-GRF prediction errors on joint kinetics calculations (ankle, knee, hip).

Main Methods:

• Utilized a public dataset of 9500 running steps for model development and validation.
• Employed a convolutional neural network (CNN) architecture for AP-GRF prediction.
• Evaluated prediction accuracy using relative root mean square error (RMSE) and analyzed effects on joint moments.

Main Results:

• The CNN accurately predicted the AP-GRF component with a relative RMSE of 7.44%.
• Prediction errors minimally impacted ankle moment calculations.
• Knee and hip joint moment estimations were more significantly affected by AP-GRF prediction errors.

Conclusions:

• The developed neural network model shows promise for estimating AP-GRF in conjunction with the size-adjustable instrumented track.
• This approach could expand the feasibility of comprehensive in situ running analysis.
• Further research may refine the model to improve accuracy for knee and hip kinetic analysis.