External validation of dementia risk prediction models in the EPIC-Norfolk cohort: a UK-based cohort study

External validation of dementia risk models is essential to assess generalisability and clinical utility. We evaluated 12 prediction models, including 10 dementia-specific and two cardiovascular-based models, in the EPIC-Norfolk cohort (n = 25,423) with up to 30 years of follow-up. Performance was assessed using discrimination, calibration, competing risks, and power analyses, stratified by follow-up and sex. Five models were fully validated and seven partially. The CAIDE, CAIDE-APOE, and DRS showed moderate performance declines, while the FRS and CHA₂DS₂-VASc demonstrated good calibration and strong transportability, supporting the value of vascular risk factors in predicting dementia. The UKBDRS maintained high discrimination (>0.80) despite partial validation. Performance was generally lower in women, especially for CAIDE and DRS. Calibration was acceptable in most fully validated models, though several were underpowered. Simplified models focusing on core vascular and lifestyle predictors may offer scalable, clinically relevant tools for dementia risk stratification.