Age specification matters? A multi-cohort evaluation of age specification and follow-up duration in the performance of midlife dementia risk scores.

Background: The heterogeneous performance of dementia and Alzheimer's disease risk scores makes it difficult to identify tools suitable for supporting risk reduction and prevention strategies. Although multiple factors may contribute, we hypothesized that differences in age specification and study design are the primary contributors to this variability. Methods: Longitudinal data were drawn from the ARIC, UK Biobank, and Whitehall II cohorts. Participants free of dementia at baseline were included. Baseline variables were used to calculate each risk score. Risk prediction was evaluated using Cox proportional hazards models with three age specifications: UKBDRS (continuous), CogDrisk-ML (categorical: 40-49, 50-54, 55-59, 60-64 years), and CAIDE (categorical age: <47, 47-53, >53 years). Predictive accuracy was assessed using Harrell's C-statistics and 95% confidence intervals (CIs). Results: Across all cohorts and risk scores, excluding age from risk scores resulted in similar C-statistics. In age- and sex-only models, continuous age produced the highest Harrell's C-statistics (95% CI) [ARIC: 0.71 (0.68, 0.73); UK Biobank: 0.78 (0.77, 0.79), and Whitehall II: 0.75 (0.62, 0.88)], followed by categorical age specifications. Compared with continuous age, categorical specifications reduced C-statistics, with the largest reduction for CAIDE. Shorter follow-up yielded higher but less precise C-statistics. Conclusions: Age specification strongly influences dementia risk model performance, especially with shorter follow-up, with continuous age yielding the highest accuracy. Our findings emphasize that differences in dementia risk prediction performance are driven primarily by age specification and cohort age structure rather than intrinsic superiority of individual risk tools.