Early Prediction of Alzheimer's Disease and Related Dementias Using Machine Learning and Claims Data

We used commercial healthcare claims (2015-2019) to conduct a retrospective study of Alzheimer's Disease and Related Dementias (ADRD) aiming at developing approaches for early ADRD prediction. ADRD cases were identified using diagnosis codes or ADRD-related medications and matched to controls using propensity scores. Prediction windows of 12, 18, 24, and 30 months before the index date were paired with a 6-month observation window. Models (XGBoost, Random Forest, Logistic Regression) were trained using 5-fold cross-validation and grid-search tuning. Performance metrics included AUC, accuracy, precision, recall, and F1 score. SHAP values were used for interpretability. XGBoost achieved the highest performance across all windows (AUC 0.94-0.95), with stable performance up to 30 months before diagnosis. SHAP analysis revealed outpatient services, anesthesia services, diagnostic imaging, and emergency/inpatient encounters as the strongest predictors. Claims-based machine-learning models can accurately predict ADRD long before diagnosis. XGBoost provides strong performance and risk signals. Future work will focus on external validation in extended populations and integration with EHR-based clinical features to further enhance model generalizability and clinical utility.