Optimizing Random Forest for Alzheimer's Disease Classification: A Transparent AI Approach

Early detection of Alzheimer's disease (AD) faces challenges like limited dataset diversity, poor model interpretability, overfitting, and noisy medical data, impeding reliable diagnostics. This study proposes a novel binning algorithm to preprocess data by creating optimal bins based on target correlation, reducing noise and computational complexity in time and storage. A Random Forest (RF) classifier is optimized to enhance accuracy, leveraging ensemble learning and SHapley Additive exPlanations (SHAP) for transparency. Using RandomizedSearchCV, 13 hyperparameters are tuned on a dataset of \mathbf{2, 1 4 9} patient records with 35 features, achieving optimal settings like 250 estimators and a max depth of 5. The model yields a 95.58 % test accuracy and 95.49% cross-validation accuracy, with only 19 misclassifications, and precision, recall, and F 1 -scores above 94 %. SHAP analysis highlights cognitive and functional assessments, such as Functional Assessment and Activities of Daily Living, as key predictors, aligning with clinical insights. The binning algorithm streamlines feature processing, while hyperparameter tuning boosts efficiency and performance, outperforming existing methods. This approach establishes a transparent, accurate, and computationally efficient framework for AD classification, advancing early diagnosis.