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

Area of Science:

• Biomechanics
• Computational modeling
• Sports science

Background:

• Physics-based models are standard for analyzing human motion mechanics.
• Artificial neural networks (ANNs) are increasingly used to support biomechanical analysis.

Purpose of the Study:

• To compare the accuracy of recurrent ANNs against physics-based models in predicting running ground reaction forces (GRF) and lower limb joint moments.
• To evaluate the ANN's ability to predict kinetic data from kinematics acquired using different experimental methods.

Main Methods:

• Trained recurrent ANNs using biomechanics data from an inertial motion capture system and force plate.
• Utilized publicly available kinematic data from optical motion capture systems for ANN evaluation.
• Calculated GRF and joint moments using linear and angular momentum theorems for the physics-based approach.

Main Results:

• Recurrent ANNs significantly outperformed physics-based models (P < .05) in predicting GRF (anteroposterior, vertical, mediolateral) and lower limb joint moments (knee/ankle flexion, hip abduction/rotation) at trained and higher running velocities.
• The trained ANN demonstrated robust prediction capabilities across different kinematic data sources and experimental techniques.

Conclusions:

• Recurrent ANNs offer a highly accurate alternative to traditional physics-based models for predicting running kinetics.
• Trained ANNs can generalize predictions to kinematic data from diverse sources, enhancing their practical applicability in biomechanical research and performance analysis.