Multi-Modal Data-Driven Approach for Alzheimer's Disease Prediction with Feature Selection and Evaluation

Effective intervention and disease management of Alzheimer's disease (AD) depend on early detection. In order to diagnose AD, this study proposes a multi-modal machine learning model that integrates biomarker, neuroimaging, cognitive, and demographic data. 200–300 patient samples were analyzed utilizing secondary datasets from ADNI and OASIS in order to identify predictive characteristics using Principal Component Analysis (PCA) and Minimum Redundancy Maximum Relevance (MRMR) feature selection. A stratified 5-fold cross-validation technique was used to compare four machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), XG Boost, and Convolutional Neural Networks (CNNs). The findings revealed that RF yielded the highest accuracy (91.2%) and AUC-ROC (0.95), followed by XGBoost (90.5%, 0.94), reflecting the efficacy of ensemble learning techniques. Feature selection enhanced classification performance substantially, and the most significant features were identified as hippocampal volume, beta-amyloid, tau protein concentration, and MMSE scores. ANOVA analysis supported that the variations in model performance were statistically significant (p = 0.001). The results indicate the promise of AI-based diagnostic tools for early AD detection. Future studies need to emphasize larger datasets, deep learning techniques, and explainable AI methods to increase clinical utility.