A Multi-Stage Machine Learning Pipeline for Alzheimer's Disease Classification Using Optimized Ensemble Models

Early and accurate identification of Alzheimer’s disease (AD) is critical, particularly in light of the rapidly aging global population and the substantial public health burden associated with the condition. In this study, a multi-stage and integrated machine learning pipeline was developed for AD classification. The pipeline consists of data scaling, model selection, voting ensemble creation, hyperparameter optimization, ADASYN resampling to eliminate class imbalance and Sequential Feature Selection (SFS) steps. Initially, the dataset was normalized through a combined application of MinMaxScaler and StandardScaler. Gradient Boosting, Random Forest, Bernoulli Naive Bayes, and RBF-kernel SVC were then designated as the primary candidate models. Hyperparameter optimization was performed with Optuna's Tree-structured Parzen Estimator (TPESampler) algorithm, and early discontinuation of low-performance trials was achieved with the Hyperband pruner mechanism. Subsequently, class imbalance was addressed using ADASYN-based resampling, and the most informative predictors were identified through the Sequential Feature Selection (SFS) method. The proposed pipeline was evaluated using stratified 5-fold cross-validation at each stage and an accuracy of 97.97% was achieved with the soft-voting ensemble comprising Gradient Boosting, Random Forest and Radial Basis Function (RBF) kernel Support Vector Machine models. In particular, the soft-voting ensemble models enhanced performance stability, providing more consistent decision boundaries compared with the individual base classifiers. As a result, the proposed pipeline delivers a systematic framework that ensures high accuracy, interpretability, and reproducibility in Alzheimer’s disease classification. The findings underscore the value of integrating multi-stage strategies, including data scaling, hyperparameter optimization, resampling, and feature selection, to enhance model robustness and diagnostic reliability.