Abstract
BACKGROUND: Emergency-department (ED) visits capture diagnostic data that could flag patients at heightened risk for Alzheimer's disease (AD) long before cognitive symptoms are formally recognized. OBJECTIVE: To examine how age and multimorbidity interact to predict AD and to test whether explainable machine learning enhances risk stratification using national ED data. METHODS: We analyzed 554,985 ED visits (2010-2014 National Emergency Department Sample) from adults ≥ 60 y. ICD-9-CM codes identified AD and 17 chronic conditions. Logistic regression estimated odds ratios (ORs) for single and combined comorbidities across five age bands. Predictive performance of logistic regression, decision tree, random forest and XGBoost was compared; Shapley Additive exPlanations (SHAP) interpreted model output. RESULTS: Urinary-tract infection (UTI; OR = 2.74), depression (1.93), hypothyroidism (1.68) and anemia (1.57) independently increased AD odds. Possessing all four "red-flag" conditions tripled risk (OR = 3.31), and each additional red-flag raised risk by 74%. Age showed a steep gradient: relative to 60-65 y, ORs climbed from 2.58 (66-70 y) to 25.8 (86-90 y; all p < 0.001). XGBoost performed best (AUC = 0.782; recall = 0.821), and SHAP confirmed age and red-flag multimorbidity as dominant predictors. CONCLUSIONS: Routine ED codes reveal an age-dependent, dose-response relationship between specific multimorbidity clusters and AD. An interpretable XGBoost model accurately identifies high-risk patients, outlining a practical pathway for real-time cognitive-risk alerts in acute-care settings.