An optimized machine learning framework integrating feature selection and Bayesian–Tree Structured Parzen estimator hyperparameter optimization for alzheimer’s disease prediction
Alzheimer's disease (AD) is an increasing global health challenge due to its rising prevalence and the limited effectiveness of existing diagnostic approaches. Neuropathological changes often begin more than a decade before clinical symptoms emerge, yet most diagnostic methods detect the disease only after significant cognitive decline. Machine learning offers promising solutions for early prediction; however, predictive performance is often limited by suboptimal feature selection and hyperparameter optimization. This study proposes an integrated predictive framework that combines three feature selection methods—ReliefF, Boruta, and Recursive Feature Elimination with Cross-Validation (RFE(RFECV))—with probabilistic hyperparameter optimization using the Tree-Structured Parzen Estimator (TPE) and Bayesian Optimization using Gaussian Processes (BayesGP) for Random Forest classification. A clinical dataset comprising 2,149 subjects and 34 features was preprocessed using missing value imputation, ordinal encoding, Min–Max normalization, and SMOTE applied exclusively to the training data. Model performance was evaluated on an independent test set using bootstrap confidence intervals, McNemar's test, and calibration analysis. The Boruta + TPE + Random Forest configuration achieved the best overall performance with 95.04% accuracy, 93.13% F1-score, and an AUC of 98.41%, while the RFE(RFECV) + TPE model achieved the highest AUC (98.48%). These findings demonstrate that integrating robust feature selection with advanced hyperparameter optimization substantially improves predictive accuracy, model reliability, and interpretability, providing an effective framework for early Alzheimer's disease risk stratification and clinical decision support.