Explainable survival analytics for dementia risk management: Comparing Cox regression and DeepSurv using the national health and aging trends study

Abstract

This study develops an explainable survival-analysis framework for predicting time to first observed dementia and evaluates whether DeepSurv, a neural-network-based survival model, offers predictive and decision-support advantages over conventional Cox proportional hazards (Cox PH) regression. We analyzed longitudinal data from Rounds 1–14 of the National Health and Aging Trends Study (NHATS), including 5,527 older adults and 1,454 observed dementia events. We examined 25 demographic, cognitive, clinical, psychological, behavioral, social, and technology-related predictors. We evaluated model performance using Harrell and Antolini concordance indices, time-dependent AUC, integrated Brier score, paired bootstrap validation, and identical-fold five-fold cross-validation. In identical-fold five-fold cross-validation, Cox PH had a higher mean C-index and lower variability than DeepSurv (0.7532 ± 0.0121 vs. 0.7361 ± 0.0440). On the independent test set, Cox PH and DeepSurv showed comparable discrimination (Harrell C-index = 0.7879 and 0.7833, respectively), and the paired bootstrap difference was small (DeepSurv − Cox PH = −0.0048, 95% CI [−0.0112, 0.0021]). Cox PH showed slightly higher time-dependent AUCs across time horizons, whereas DeepSurv had a marginally lower integrated Brier score (0.1279 vs. 0.1290). Age, self-rated memory, digital communication, online daily activity, and social participation emerged as influential predictors across both modeling approaches. For the highest-risk test participant, a combined favorable scenario was associated with a 35.8% reduction in the DeepSurv risk score; this was a model-based scenario, not a causal effect. The findings underscore that healthcare organizations should balance predictive performance, interpretability, and implementation complexity when evaluating AI-based decision-support models for dementia risk management.