Leakage-Safe Boosting Ensembles With SHAP Explainability for Alzheimer's Disease Risk Prediction: A Rigorous Multi-Classifier Benchmark on Global Structured Clinical Data

Hadeer Mahmoud, Department of Artificial Intelligence, Faculty of Computers and Artificial Intelligence, Modern University for Technology & Information, Cairo, EgyptFollow
Fady Salama, Faculty of Computer and Information Technology, Innovation University, Cairo, EgyptFollow

ORCID

Hadeer Mahmoud: https://orcid.org/0000-0002-8681-1909

Fady Salama: https://orcid.org/0009-0007-5668-6522

Keywords

Alzheimer's disease, Machine learning, AdaBoost, SHAP explainability, Ensemble learning, Leakage-safe pipeline, Clinical risk prediction

Article Type

Original Article

Abstract

Alzheimer's disease (AD) is a major public health challenge, and scalable non-invasive risk-stratification tools are needed. We benchmarked 11 machine-learning classifiers on a publicly available global structured Alzheimer's prediction dataset (N = 74,283; 24 predictors) using leakage-safe sklearn/imblearn pipelines. Four resampling strategies were assessed with stratified cross-validation; preprocessing included imputation, IQR-based outlier capping, feature engineering, one-hot encoding, and standardization. The leading models underwent HalvingRandomSearchCV optimization, followed by soft-voting and stacking ensembles. Performance was evaluated on a held-out 20% test set and interpreted with SHAP, permutation importance, and built-in feature importance. No resampling was selected in Stage A (composite score = 0.575). AdaBoost was the best tuned model (10-fold CV ROC-AUC = 0.794 ± 0.003; F1 = 0.688 ± 0.007; MCC = 0.448 ± 0.011) and achieved test-set Accuracy = 72.97%, Recall = 73.63%, F1 = 69.25%, ROC-AUC = 0.80, PR-AUC = 0.70, and MCC = 0.4554. SHAP identified Age, Family History of Alzheimer's, and APOE-ε 4 Genetic Risk Factor as the dominant predictors. Wilcoxon signed-rank testing with Holm correction supported AdaBoost's superiority over Logistic Regression and both ensemble configurations. These findings support leakage-safe boosting as a reproducible approach to risk discrimination from structured clinical and demographic data; however, because the dataset is synthetic/aggregated, the model is a methodological screening benchmark rather than a clinically validated diagnostic tool.

How to Cite

Mahmoud, Hadeer and Salama, Fady (2026) "Leakage-Safe Boosting Ensembles With SHAP Explainability for Alzheimer's Disease Risk Prediction: A Rigorous Multi-Classifier Benchmark on Global Structured Clinical Data," Sustainable Machine Intelligence Journal: Vol. 14: Iss. 3, Article 2.
DOI: https://doi.org/10.63689/3005-3617.1092