Nonlinear Association of Length of Stay With In-Hospital Mortality in Alzheimer’s Disease Hospitalizations: Admission-Only vs Inpatient-Course Prediction Using Explainable Machine Learning

BACKGROUND: Hospitalizations among patients with Alzheimer’s disease (AD) carry substantial mortality risk, but the relationship between length of stay (LOS) and in-hospital death may be non-linear. We evaluated LOS–mortality patterns and compared admission-only versus inpatient-course prediction using explainable machine learning. METHODS: Using the 2017 Nationwide Readmissions Database, we identified AD hospitalizations among adults aged ≥60 years. The outcome was in-hospital mortality. LOS was analyzed in clinically interpretable bins and with restricted cubic splines. Two prespecified models were compared: Model A used admission-only variables, excluding LOS, procedure count, and total charges; Model B added these inpatient-course variables. Performance was evaluated using patient-grouped 5-fold out-of-fold validation and summarized by AUROC and AUPRC. SHAP was used for model interpretation. RESULTS: Among 11,377 AD hospitalizations, 600 in-hospital deaths occurred (5.27%). Mortality was highest for LOS 0-1 day (14.0%), lowest at 4-6 days (3.26%), and increased again with prolonged stays. Model A achieved AUROC/AUPRC of 0.729/0.149, whereas Model B improved performance to 0.794/0.283. Sepsis, acute kidney injury, stroke, older age, and higher diagnostic burden were consistently influential predictors. DISCUSSION: In AD hospitalizations, mortality clusters at LOS extremes. Admission-only models identify meaningful early risk, while inpatient-course variables add prognostic information as complications and care intensity evolve.