Background: Current dementia risk prediction models are predominantly developed and validated in White populations, limiting their generalizability to diverse racial groups. Models trained in minority samples are often underpowered due to small sample sizes, while those trained on pooled multi-racial data may be biased toward the majority group, failing to account for racial heterogeneity.

Objective: To address these issues and investigate race-specific dementia risk factors, we propose a deep transfer learning framework to predict dementia risk in racial minority populations using UK Biobank (UKB) data.

Methods: We collected data on 310 potential dementia risk predictors (including genetic and environmental factors) from 490,553 UKB participants across three racial groups: White (W), Black (B) and Asian (A), including 6649 dementia cases (W: 6433; B: 114; A: 102) and 483,904 controls (W: 466,183; B: 7944; A: 9777). The deep neural network was pre-trained on White only and then transferred the knowledge to fine-tune on Black/Asian to predict dementia risk in the minority groups. We also identified race-specific dementia risk factors using SHAP values and developed a risk score for the minority groups.

Findings: Our deep transfer learning-based models achieved higher prediction accuracy (B: AUC=0.836; A: AUC=0.751) compared to minority-only models (B: AUC= 0.692; A: AUC=0.523). Social support and diet emerged as race-specific dementia risk factors for Black and Asian populations.

Public health implications: Our deep transfer learning framework constructs a dementia risk prediction model for racial minority populations, enhancing early detection and prevention of dementia in underrepresented populations.