Leakage-Aware Explainable Alzheimer’s Disease Detection Using Ensemble-Based Tree Models and Shapley Additive Explanations

Alzheimer’s disease and dementia are leading causes of cognitive decline among older adults, creating a need for accurate, transparent, and clinically interpretable decision-support tools for routine screening. This study proposes a leakage-aware and explainable machine learning framework for dementia classification using ensemble tree-based models and Shapley Additive Explanations (SHAP). The longitudinal OASIS dataset, comprising demographic variables, cognitive assessments, and MRI-derived volumetric biomarkers, was evaluated using stratified 10-fold cross-validation with discrimination and calibration metrics. Among the evaluated classifiers, Random Forest achieved the highest accuracy (0.818 ± 0.106), while Extra Trees achieved the highest ROC–AUC (0.874 ± 0.109). Feature-group analysis demonstrated that cognitive variables provided the strongest standalone predictive performance, whereas combining demographic, cognitive, and MRI-derived features yielded the highest overall discrimination. Wilcoxon signed-rank testing revealed no statistically significant differences among the leading ensemble classifiers, suggesting that feature composition exerted a greater influence on predictive performance than model selection. SHAP-based global and patient-level explanations identified the Mini-Mental State Examination (MMSE), normalized whole brain volume, and education level as major contributors to dementia prediction. Additional MMSE ablation and leakage-aware analyses highlighted the substantial influence of cognitive assessment variables on model performance and underscored the importance of transparent evaluation practices. These findings demonstrate that integrating ensemble learning, calibration assessment, feature-group analysis, and explainable artificial intelligence can provide trustworthy and clinically interpretable decision-support systems for dementia screening using routinely collected clinical and neuroimaging data.