Deep learning prediction of the knee adduction moment in females with osteoarthritis based on frontal plane data collection

The knee adduction moment (KAM) is of great importance in treating medial knee osteoarthritis; however, its clinical applications are limited by the cost of gait laboratory analysis. This study evaluates the possibility of predicting the first peak of the KAM in females with osteoarthritis using a restricted frontal-plane kinematic representation. We investigate whether the knee adduction angle provides additional predictive information beyond anatomical landmark trajectories. Visual3D was used for KAM calculations, and thirty-two females with osteoarthritis were studied during normal walking. Frontal-plane trajectories of seven lower-limb anatomical landmarks on each limb were used as predictors, with and without incorporating the knee adduction angle. Two predictive pipelines were proposed: a fully connected neural network trained on time-series predictors and a convolutional neural network trained on image-like representations of the same data. Model development comprised 29 participants and used repeated participant-level cross-validation. Three participants (six limbs) were held out as an independent test set. The results were r²=0.78 and r²=0.79 excluding the knee adduction angle, and r²=0.87 and r²=0.83 with its inclusion for the first and second pipelines, respectively. Additional analyses using corresponding 3D inputs showed no statistically significant differences in MAE between matched frontal-plane and 3D-input configurations after correction for multiple comparisons. This study demonstrates that the inclusion of the knee adduction angle as an additional predictor alongside frontal-plane landmark positions notably enhances the results. These findings motivate further investigation of markerless approaches using video-derived kinematic information.