Analyzing 55,004 participants from the National Alzheimer's Coordinating Center (NACC), race-stratified Random Survival Forest (RSF) models achieved 5-year dementia prediction AUCs of 0.88–0.91—substantially outperforming traditional Cox proportional hazards models (AUC 0.69–0.81)—across Black, White, Asian, and American Indian cohorts. Incorporating resting heart rate (RHR) alongside the 2024 Lancet Commission's 14 modifiable risk factors modestly but consistently improved prediction. Critically, predictor importance rankings differed meaningfully between racial groups, suggesting dementia risk pathways are not uniform across populations.
This is a preprint posted on medRxiv and has not yet been peer-reviewed; conclusions should be treated as preliminary. That said, the scale and methodological ambition here are notable. Most dementia risk tools were developed predominantly in White European cohorts, creating well-documented generalizability failures when applied to Black Americans—who face roughly twice the dementia incidence rate. The RSF approach captures nonlinear interactions that Cox models cannot, which may explain the large performance gap. The addition of RHR as a biomarker is biologically plausible: elevated resting heart rate correlates with autonomic dysfunction, cardiovascular disease burden, and neuroinflammatory pathways implicated in Alzheimer's and vascular dementia. Practically, RHR is effortlessly obtainable via consumer wearables, making it a low-cost enrichment variable for clinical screening. The race-stratified modeling philosophy represents a meaningful step toward equitable precision medicine in neurology, though prospective validation in independent cohorts and interrogation of socioeconomic confounders remain essential before clinical deployment.
