Comparing shallow, deep, and transfer learning in predicting joint moments in running.

Joint moments are commonly calculated in biomechanics research and provide an indirect measure of muscular behaviors and joint loads. However, joint moments cannot be easily quantified clinically or in the field, primarily due to challenges measuring ground reaction forces outside the laboratory. The present study aimed to compare the accuracy of three different machine learning (ML) techniques - functional regression [ ML_fregress ], a deep neural network (DNN) built from scratch [ ML_DNN ], and transfer learning [ ML_TL ], in predicting joint moments during running. Data for this study came from an open-source dataset and two studies on running with and without external loads. Three-dimensional (3D) joint moments of the hip, knee, and ankle, were derived using inverse dynamics. 3D joint angle, velocity, and acceleration of the three joints served as predictors for each of the three ML techniques. Prediction performance was generally the best using ML_DNN, and the worse using ML_fregress. Absolute predictive performance was the best for sagittal plane moments, which ranged from a RMSE of 0.16 Nm/kg at the ankle using ML_DNN, to a RMSE of 0.49Nm/kg at the knee using ML_fregress. ML_DNN resulted in the greatest improvement in relative prediction performance (relRMSE) by 20% compared to ML_fregress for the ankle adduction-abduction moment. DNN with or without transfer learning was superior in predicting joint moments using kinematic inputs compared to functional regression. Synergizing ML with kinematic inputs has the potential to solve the constraints of obtaining high fidelity biomechanics data normally only possible during laboratory studies.