Wearable Sensor-Based Prediction of Knee Joint Moments in Competitive Rugby Players During Multidirectional Movements

This study aimed to develop an interpretable machine learning framework using plantar-pressure sensors to estimate knee moments during rugby-specific movements. A total of 31 North American competitive rugby players were recruited to perform four common rugby movements on an infilled artificial turf surface in a motion capture laboratory. Synchronized data acquisition was performed for plantar pressure, motion capture, and ground reaction forces. Knee moments were calculated via inverse dynamics. Genetic Algorithm-Deep Forest Regression (GA-DFR) was used to select plantar-pressure sensors and predict resultant moments within participant-level nested cross-validation. Four alternative models predicted the resultant using the same layouts and participant splits. Complementary deep forest regression (DFR) models predicted the three moment components using these layouts. SHapley Additive exPlanations (SHAP) interpreted held-out predictions. Layouts contained 7-18 sensors. Pooled resultant R² was 0.795, with root mean square error of 0.0512 N·m/N. DFR performed better than linear regression and the convolutional neural network, with no significant differences from the deep neural network or long short-term memory network. Pooled R² was 0.806, 0.662, and 0.675 for flexion-extension, adduction-abduction, and internal-external rotation moments, respectively. Resultant temporal SHAP contributions exceeded their stance-wide mean during 9%-55% of stance, peaking at 24%. Reduced plantar-pressure inputs supported estimation of resultant and direction-specific knee-moment waveforms in unseen rugby players.