A Framework to Replace Measured Ground Reaction Forces with ANN-predicted ones for Joint Load Estimation in OpenSim

Area of Science:

• Biomechanics
• Computational modeling
• Artificial Intelligence

Background:

• Musculoskeletal (MSK) modeling is crucial for analyzing lower-limb joint loading.
• Accurate ground reaction forces (GRFs) are essential inputs for MSK models.
• Predicting GRFs using Artificial Neural Networks (ANNs) offers a potential alternative to experimental measurements.

Purpose of the Study:

• To integrate an ANN with OpenSim MSK modeling to predict lower-limb joint reaction forces (JRFs).
• To evaluate the accuracy of ANN-predicted GRFs and their impact on JRF estimations.
• To identify factors influencing prediction accuracy for future model refinement.

Main Methods:

• Trained an ANN to predict GRFs from ankle position data.
• Predicted GRFs for 20 new subjects and compared them to measured GRFs (antero-posterior, medio-lateral, vertical).
• Utilized OpenSim to compute JRFs using both measured and ANN-predicted GRFs.

Main Results:

• ANNs achieved moderate correlations for medio-lateral and vertical GRFs but lower accuracy for antero-posterior GRFs.
• Errors in predicted GRFs propagated to JRF estimations, with higher variability observed.
• Normalized Root Mean Square Errors (NRMSE) for JRFs ranged from 0.16 ± 0.07 to 0.33 ± 0.32.

Conclusions:

• ANN-predicted GRFs show potential for MSK analysis but require improved accuracy, especially for vertical GRFs.
• Factors like coordinate systems and data resampling affected prediction accuracy.
• Further refinement could enable ANN-predicted GRFs to replace measured GRFs in clinical and research settings.