AlzStack: Forecasting early-onset Alzheimer's with an explainable AI system using multiple data balancing techniques

Abstract

Alzheimer's disease (AD) is a degenerative neurological disease that progresses over time, making early detection crucial for effective intervention and better patient prognosis. Traditional diagnostic methods such as cognitive assessments, neuroimaging, and biomarker analysis can be time-consuming, costly, and inconsistent. We introduce AlzStack, a soft voting ensemble model to classify AD from a richly detailed dataset containing 2149 patients across demographic, medical, lifestyle, and cognitive variables. To resolve class imbalance, we implemented a pipeline 5-fold cross-validation, randomized search for hyper parameter tuning and advanced resampling methods such as SMOTE (Synthetic Minority Oversampling Technique), ADASYN, BorderlineSMOTE, and SVMSMOTE. Soft Vote Classifier surpassed both stacking ensembles and hard voting with an AUC value of 94.27 %, accuracy of 93.26 %, precision of 89.17 %, a recall of 92.11 %, and F1-score value of 90.61 %.A secondary experiment with only resampling methods applied to data to all base models served as a baseline for comparison confirming the superior performance of cross-validation AlzStack configuration. To improve interpretability, we utilized a wide range of Explainable Artificial Intelligence (XAI methods) and these approaches yielded global and local explanations about model behavior, emphasizing key features like MMSE scores, functional measures, and behavioral markers. Combining robust predictive performance with explainable decision-making makes AlzStack is a healthcare decision-support algorithm for the early detection of AD.