Classification model for LESS-defined bilateral jump-landing patterns based on XGBoost–SHAP and analysis under cognitive dual-task conditions

The Landing Error Scoring System (LESS) is widely used to assess jump-landing mechanics and injury-related movement patterns but requires expert raters. This study developed an interpretable XGBoost–SHAP model for classifying LESS-defined high-risk bilateral jump-landing patterns. Three-dimensional (3D) motion capture, video, and force platforms were used to assess landing trials in 600 young men; 50 additional men completed single- and cognitive dual-task landings for external validation and condition comparison. Ten kinematic features were derived from OpenSim and three binary features were coded from synchronized video recordings; physiotherapist-rated LESS scores provided reference labels. Four classifiers were compared using repeated stratified nested cross-validation. XGBoost achieved the greatest mean cross-validated AUC (0.991 ± 0.001) and, on the held-out test set, an accuracy of 0.983, recall of 0.976, precision of 0.988, F1-score of 0.982, and AUC of 0.997. SHAP identified foot contact symmetry (FCS), trunk flexion displacement (TFD), knee flexion at initial contact (KF-IC), ankle plantarflexion at initial contact (APF-IC), and trunk lateral flexion at initial contact (TLF-IC) as the five most influential features. Under dual-task conditions, asymmetric FCS increased from 14 to 36%. These findings support interpretable classification of LESS-defined landing patterns and highlight cognitive-task interference in movement screening. However, video-derived inputs remain observer-dependent.