Population-specific dementia risk prediction using deep transfer learning in diverse populations

Abstract Existing dementia risk models often perform less well in underrepresented ethnic populations, potentially contributing to inequities in dementia risk assessment. We developed and validated population-specific dementia risk models using a mixture-of-experts deep transfer learning framework in 490,031 UK Biobank participants from diverse ethnic groups (White, Black and Asian) using 282 candidate predictors. Deep transfer learning improved predictive performance in underrepresented groups, with area under the curve (AUC) of 0.903 (95% CI 0.827–0.961) for Black and 0.835 (0.763–0.894) for Asian participants. We identified 30 shared and 24 population-specific predictors and derived interpretable risk scores (0–100 scale) with estimated 5-, 10-, and 15-year dementia risk. External validation in the All of Us cohort showed consistent risk discrimination across populations, although predictor matching and follow-up duration were limited. An open-access web application enables population-specific risk estimation. These findings illustrate the potential of deep transfer learning to improve dementia risk prediction and support more equitable dementia risk assessment across diverse populations.