Dementia care presents formidable challenges due to the diverse paths of disease progression and outcomes. To address this, predictive models are imperative for identifying patients at risk of near-term mortality and understanding the factors influencing mortality risk across various dementia types. In this study, we employed machine-learning techniques to develop predictive models for dementia patient mortality, utilizing data from 45,275 participants and 163,782 visit records from the U.S. National Alzheimer’s Coordinating Center (NACC). Our multi-factorial XGBoost models, incorporating a concise set of mortality predictors, achieved remarkable performance, with an area under the receiver operating characteristic curve (AUC-ROC) exceeding 0.82 across 1-, 3-, 5-, and 10-year thresholds. Notably, dementia-related predictors, particularly specific neuropsychological tests, dominated the trained models, showing minimal influence from other age-related causes of death like stroke and cardiovascular conditions. Furthermore, stratified analyses uncovered both shared and distinct predictors of mortality among eight dementia types, with unsupervised clustering revealing intriguing associations, such as grouping vascular dementia with depression and Lewy body dementia with frontotemporal lobar dementia.
