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

This study integrates an Artificial Neural Network (ANN) with musculoskeletal (MSK) modeling in OpenSim to predict lower-limb joint reaction forces (JRFs). The ANN was trained, in a previous study, to predict ground reaction forces (GRFs) using ankle position data during walking. This model was used to predict GRFs of 20 unseen subjects (10 male and 10 females) and predicted GRFs were evaluated against experimentally measured GRFs in three directions: antero-posterior (AP), medio-lateral (ML), and proximo-distal (vertical). OpenSim was used to compute JRFs at the hip, knee, and ankle joints in two scenarios: (1) using measured GRFs, and (2) using ANN-predicted GRFs. The results showed that while the ANN performed well in predicting ML and vertical GRFs with moderate correlations (> 0.65), the accuracy in predicting AP GRFs was lower, with weaker correlations and higher errors. The normalized root mean square error (NRMSE), calculated by dividing the root mean square error (RMSE) by the range of data, ranged from 0.16 ± 0.06 to 0.29 ± 0.06 for predicted versus measured GRFs across both legs. The errors in GRF predictions were carried over into the estimation of JRFs, as indicated by the NRMSE values. The NRMSE for JRFs ranged from 0.16 ± 0.07 to 0.33 ± 0.32, indicating a higher variability in JRF prediction accuracy. These results suggest that enhancing the accuracy of GRF predictions, particularly for the vertical GRF, could significantly improve the overall MSK modeling process. Several factors contributing to prediction errors were identified, including the use of absolute ankle positions, differences in coordinate systems between training and testing data, and the resampling of gait cycle data. Despite these limitations, the study suggests that with further refinement, ANN-predicted GRFs could replace measured GRFs for MSK analysis in clinical and research settings.