An explainable AI framework using wearable IMU-based motion sensing for estimating and explaining motor skill improvement in school-based Baduanjin education

Estimation of individual skill improvement within specific school-based physical education units remains methodologically underdeveloped, with existing approaches typically lacking external validation and interpretability. Because the full model incorporates training-load features aggregated across the same temporal window as the outcome, it is epistemologically retrospective; readers should not interpret it as a prospectively deployable prediction tool. This study aimed at developing and externally validating an explainable machine learning ensemble to retrospectively explain and estimate standardized skill gain (Δz) in Baduanjin (eight-section brocade) qigong, a mandatory traditional martial arts skill in the Chinese national PE curriculum, among students from four schools ( n = 727). A stacking meta-learner composed of five candidate algorithms, including XGBoost, LightGBM, Random Forest, LASSO, and MLP, was trained using biomechanical, psycho-cognitive, and semester-aggregated training-load features obtained from the development cohort (Schools A + B, n = 486) and then validated against two geographically independent but regionally situated cohorts (Schools C + D, n = 241). Because the training-load features were aggregated across the full 16-week semester, the model functions as a retrospective semester-level estimation framework rather than a prospective early-warning tool. The model yielded an external validation R² of 0.691 (95% CI: 0.661–0.718), with robust subgroup performance across gender, initial skill quartiles, and sport types. SHAP-based interpretability analysis identified exercise self-efficacy as the highest-ranking modifiable feature in model-derived importance, with its predictive contribution predominantly associated with the training-response component rather than baseline level. A baseline-only sensitivity model excluding semester-aggregated training-load variables retained moderate external validity (R² = 0.573), indicating that prospective application at semester onset is feasible in principle but at substantially reduced precision. These findings suggest that, within this dataset and model specification, psycho-cognitive features contributed more strongly to model estimations than biomechanical features.