A Statistically Validated Machine Learning Framework for Early Alzheimer's Disease Detection Using Structured Clinical Data

Abstract

Early detection of Alzheimer's disease (AD) is vital in resource-limited settings that rely on structured clinical data. This study presents a rigorous, interpretable benchmarking framework using a dataset of 2,149 subjects (64.6\% cognitively normal; 35.4\% AD), split into training (80\%) and testing (20\%) sets. Five models—Logistic Regression, Random Forest, a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), and Optimized Gradient Boosting—were evaluated under identical conditions. Optimized Gradient Boosting achieved the highest performance on the test set ($n = 430$) with 95.10\% accuracy, 92.10\% sensitivity, 96.80\% specificity, an F1-score of 0.93, and the fewest false negatives ($n = 12$). Random Forest also performed strongly (93.95\% accuracy), while linear and deep learning models were less effective. SHAP analysis aligned model predictions with key clinical biomarkers, including functional assessments, activities of daily living (ADL), and MMSE scores, demonstrating that ensemble tree-based models excel in structured clinical settings.