Flagging high-risk comorbidities in Alzheimer's disease emergency department visits: A machine learning analysis of mortality outcomes

Abstract

Background

Alzheimer's disease (AD) patients experience elevated mortality during emergency department (ED) encounters, yet the associated risk factors remain insufficiently characterized.

Objective

To identify predictors of mortality among older adults with AD during ED visits and examine differences in comorbidity patterns between those who died and those who survived.

Methods

We analyzed 20,532,351 ED visits for adults aged ≥60 from the 2012–2014 Nationwide Emergency Department Sample (NEDS). Visits were stratified by AD status, ZIP-code income quartile, and mortality outcome (defined as in-ED or in-hospital death following ED presentation). We used logistic regression and machine learning models (random forest, gradient boosting, XGBoost) to predict mortality in a 1:1 matched case-control dataset. SHapley Additive exPlanations (SHAP) were applied to interpret model outputs.

Results

AD patients accounted for 1.76–2.13% of all ED visits, with mortality rates of 2.55–2.68% compared to 1.10–1.76% for non-AD patients. Socioeconomic status and ED charges were not associated with increased mortality. Odds ratio analysis identified rare terminal events as top predictors (e.g., respiratory arrest, OR = 55.5), while SHAP analysis highlighted more prevalent and clinically actionable conditions such as acute respiratory failure and septicemia as major drivers of mortality. All models performed comparably (AUC ≈ 0.85).

Conclusions

AD patients face significantly higher mortality during ED encounters. Integrating explainable machine learning with large-scale administrative data may help flag lethal comorbidities in real time and improve outcomes through better ED triage and care prioritization.