Explainable Ensemble Learning for Early Diagnosis of Alzheimer’s Disease Using SHAP and FSHAP on ADNI Tabular Data

Alzheimer’s disease (AD) is a serious neurodegenerative disorder worldwide with devastating consequences for cognitive function and quality of life. Early and accurate diagnosis of AD is important for effective clinical intervention and management. Existing artificial intelligence (AI) diagnostic models for AD are generally black-box systems, which diminishes their transparency and clinical acceptability. This issue is a major obstacle to integrating AI technologies into healthcare systems, where interpretability and clarity in decision-making are critical factors for gaining the trust of medical professionals. Based on these challenges, we propose an explainable ensemble framework for the early diagnosis of AD using structured clinical data from the AD Neuroimaging Initiative (ADNI). Our model integrates state-of-the-art ensemble classifiers-random forest, LightGBM, stacking, and voting ensembles-with SHAP (Shapley additive explanations) and fuzzy SHAP (FSHAP) to provide transparent feature attributions. We preprocess ADNI’s 180,799 records, which include 15 clinical and biomarker features, by handling missing values (through deletion or imputation), applying synthetic minority oversampling technique (SMOTE) for class balancing, encoding categorical variables, and correcting outliers. Among the evaluated models, the voting ensemble achieved the highest test accuracy ($98.6 \%$) along with robust precision and recall scores (approximately $98 \%$). SHAP analysis identified cognitive and neuropsychiatric scores, such as Mini-Mental State Examination (MMSE) Total, Functional Activities Questionnaire (FAQ), Neuropsychiatric Inventory Questionnaire (NPIQ), and Clinical Dementia Rating (CDR). FSHAP further quantified the confidence levels associated with each feature’s contribution, demonstrating that LightGBM models assign stronger and more definitive importance to certain features (e.g., Global CDR, FAQ), whereas ensemble averaging produces more conservative attributions under uncertainty.