Explainable Machine Learning Using Sensor-Derived Biomechanical Features to Classify Elevated VALR-Related Loading Across Midsole Hardness Conditions in School-Aged Boys

(1) Background: Changes in midsole hardness may affect lower-limb impact loading during forefoot strike (FFS) running in children, yet the biomechanical basis for discriminating elevated VALR-related loading remains unclear. (2) Methods: Fourteen school-aged boys performed FFS running tests in experimental shoes with four midsole hardness levels (37, 42, 47, and 52 Shore C). Lower-limb kinematics and surface electromyography (sEMG) data were collected during the dominant leg stance phase. After preprocessing, VALR was calculated from 336 valid trials, and 28 stance-phase biomechanical features were extracted, yielding a final machine-learning dataset of 324 trials after excluding incomplete feature data. VALR was used to compare loading changes and define trial-level elevated-loading labels based on the median VALR value. Classification models were evaluated under participant-level GroupKFold validation, and XGBoost was retained for exploratory SHAP analysis. (3) Results: VALR showed an upward trend with increasing hardness, but no statistically supported change point was identified. XGBoost achieved an accuracy of 75.93%, precision of 74.14%, recall of 79.63%, F1-value of 0.768, and pooled out-of-fold AUC of 0.738. SHAP analysis indicated that distal and non-sagittal kinematic features contributed most to model classification. (4) Conclusions: Elevated VALR-related loading during children’s FFS running may be characterized by a multi-feature model-based pattern rather than a fixed midsole hardness threshold.