Longitudinal Alzheimer’s Disease Progression Analysis via a Hybrid Feature-Selected Stacking Ensemble Framework with Explainable AI

This study introduces a leak-free, highly interpretable machine learning framework for tracking Alzheimer’s disease progression using longitudinal MRI data from the OASIS-2 cohort ($N = 373$ visits). The primary contribution lies in formulating dynamic rate-of-change trajectory features ($\Delta\text{MMSE}$, $\Delta\text{nWBV}$) to capture temporal neurodegenerative changes, while embedding a 3-stage hybrid feature selection pipeline and Borderline-SMOTE oversampling strictly within a subject-aware 5-Fold Stratified GroupKFold cross-validation scheme to eliminate intra-subject data leakage. Evaluated across heterogeneous tree-based base estimators and a meta-learned stacking ensemble, the framework achieves high diagnostic performance—peaking at 80.92% precision for XGBoost, 76.68% accuracy for Random Forest, and 0.838 ROC-AUC for Extra Trees—complemented by dual-level XAI (SHAP and LIME) to validate predictions against clinically established neuroimaging biomarkers.