Prediction of joint moments from kinematics using machine learning in children with congenital talipes equino varus and typically developing peers

Understanding joint loading and the crucial role of joint moments is essential for developing treatment strategies in gait analysis, which often requires the precise estimation of joint moments through an inverse dynamic approach. This process necessitates the use of a force plate synchronized with a motion capture system. However, effectively capturing ground reaction force in typically developing (TD) children and those with congenital talipes equino varus (CTEV) presents challenges, while the availability and high cost of additional force plates pose additional challenges. Therefore the study aimed to develop, train, and identify the most effective machine learning (ML) model to predict joint moments from kinematics for TD children and those with CTEV. In a study at the Gait Lab, 13 children with bilateral CTEV and 17 TD children underwent gait analysis to measure kinematics and kinetics, using a 12-camera Qualisys Motion Capture System and an AMTI force plate. ML models were then trained to predict joint moments from kinematic data as input. The random forest regressor and deep neural networks (DNN) proved most effective in predicting joint moments from kinematics for TD children, yielding better results. The Random Forest regressor achieved an average r of 0.75 and nRMSE of 23.03 % for TD children, and r of 0.74 and 23.82 % for CTEV. DNN achieved an average r of 0.75 and nRMSE of 22.83 % for TD children, and r of 0.76 and nRMSE of 23.9 % for CTEV. The findings suggest that using machine learning to predict joint moments from kinematics shows moderate potential as an alternative to traditional gait analysis methods for both TD children and those with CTEV. Despite its potential, the current prediction accuracy limitations hinder the immediate clinical application of these techniques for decision-making in a pediatric population.