Long-term Alzheimer's disease mortality prediction in adults aged ≥60 years: A prospective cohort study benchmarking survival machine learning algorithms.

INTRODUCTION: Accurate risk stratification for long-term Alzheimer's disease (AD)-specific mortality remains limited. METHODS: We analyzed data from 5,149 adults aged ≥60 years in NHANES III (1988-1994), with 116 baseline variables and mortality follow-up through 2019 via the National Death Index. Ten survival machine learning (ML) models were benchmarked. Predictive performance was assessed using Harrell's concordance index (C-index). RESULTS: Over a median follow-up of 12.1 years for survivors and 17.8 years for decedents, Lasso (C-index = 0.76, 95% CI: 0.72-0.80) and Extreme Gradient Boosting (C-index = 0.76, 95% CI: 0.73-0.79) achieved the highest accuracy. Feature importance analyses revealed novel predictors of AD mortality. Models using fewer than 20 variables retained acceptable performance (C-index > 0.70). CONCLUSION: Survival ML models effectively predict long-term AD-specific mortality using routine clinical data. Their interpretability, scalability, and capacity to identify novel risk factors support integration into geriatric risk assessment frameworks. Highlights: We benchmarked 10 survival machine learning (ML) algorithms using 116 clinical variables to predict long-term Alzheimer's disease (AD)-specific mortality.Feature importance analysis identified novel non-imaging clinical predictors, including arm circumference, self-rated physical activity, and alcohol consumption.This work highlights the underused potential of routine clinical data for AD mortality prediction and underscores the need for interpretable, population-based ML applications.