Persistent racial disparities in dementia raise concerns about the validity and generalizability of existing prognostic models across diverse populations. This study assessed whether the 2024 Lancet Commission risk factors, together with resting heart rate (RHR), can improve race-specific dementia risk prediction. The analysis aimed to compare traditional survival modeling with machine-learning survival methods and to examine whether the inclusion of RHR provides additional predictive value across racial groups.
Analyses used de-identified data from the National Alzheimer’s Coordinating Center (NACC) Uniform Data Set (UDS), version 3.0. The sample comprised 55,004 participants drawn from NACC, collected by participating Alzheimer Disease Research Centers. Data access was via an approved NACC request and the study received ethical approval from the Health Sciences Research Ethics Board of Brock University (REB #24072). The source notes that ADRCs obtained informed consent from participants and that all data were de-identified prior to analysis.
The models incorporated the risk factors identified by the 2024 Lancet Commission as well as resting heart rate as a candidate prognostic biomarker. The manuscript reports that the inclusion of RHR produced modest but consistent improvements in predictive performance across racial groups. Specific variable definitions, coding details, or the full list of Lancet risk factors used were not reported in the abstract; those details may be in the full manuscript or supplementary material.
Two survival modeling approaches were developed and compared within race-specific strata. Cox proportional hazards models represented the classical statistical approach to time-to-event prediction. Random Survival Forest (RSF) models represented a machine-learning ensemble approach adapted for censored survival data. Models were trained separately for each racial group to generate race-specific risk predictions.
Models were developed to predict time to dementia at 1-, 3-, and 5-year horizons. Separate models were built for Black, White, Asian, and American Indian participants to capture potential heterogeneity in risk factor effects and disease presentation across these groups.
Across all racial groups and prediction horizons, Random Survival Forest models outperformed Cox proportional hazards models. Reported 5-year area under the receiver operating characteristic curve (AUC) values for RSF were 0.88–0.91, compared with 0.69–0.81 for Cox models. The abstract presents these ranges but does not provide per-race AUC point estimates or confidence intervals in the summary.
Including resting heart rate in the prediction models produced modest and consistent improvements in model performance across the racial groups evaluated. The abstract emphasizes that RHR may be a useful complementary prognostic biomarker. Exact effect sizes or numerical changes in AUC attributable to RHR inclusion were not provided in the abstract and would require consulting the full text or supplementary tables.
The analysis found that predictor importance varied by race, indicating heterogeneity in dementia risk profiles and likely in disease presentation. The manuscript notes these differences in variable importance as evidence for the need to tailor prognostic tools to diverse populations. The abstract does not enumerate which predictors were most important within each racial stratum; those details would be in the full paper or supplementary material.
The authors declared no competing interests. Ethical approval was obtained from the Health Sciences Research Ethics Board (REB #24072) at Brock University. The study used NACC UDS v3.0 data collected by ADRCs; participant consent and local IRB approvals were obtained at data collection sites. Data generated in the analysis are available from the corresponding author upon reasonable request. Funding was declared from the Canadian Institutes of Health Research (PJT-186091).
In this large NACC cohort, machine-learning survival models (RSF) using 2024 Lancet Commission risk factors and resting heart rate showed superior predictive performance compared with Cox proportional hazards models, particularly over a 5-year horizon. The consistent but modest improvement with RHR suggests it may serve as a practical complementary biomarker for prognostic models. Differences in predictor importance across racial groups underscore the importance of equitable, race-specific approaches to dementia risk assessment. The abstract indicates these findings support further development of personalized dementia risk prediction tools for diverse populations.