External validation of four dementia prediction models for use in the general community-dwelling population: a comparative analysis from the Rotterdam Study

To systematically review the literature for dementia prediction models for use in the general population and externally validate their performance in a head-to-head comparison. We selected four prediction models for validation: CAIDE, BDSI, ANU-ADRI and DRS. From the Rotterdam Study, 6667 non-demented individuals aged 55 years and older were assessed between 1997 and 2001. Subsequently, participants were followed for dementia until 1 January, 2015. For each individual, we computed the risk of dementia using the reported scores from each prediction model. We used the C-statistic and calibration plots to assess the performance of each model to predict 10-year risk of all-cause dementia. For comparisons, we also evaluated discriminative accuracy using only the age component of these risk scores for each model separately. During 75,581 person-years of follow-up, 867 participants developed dementia. C-statistics for 10-year dementia risk prediction were 0.55 (95% CI 0.53–0.58) for CAIDE, 0.78 (0.76–0.81) for BDSI, 0.75 (0.74–0.77) for ANU-ADRI, and 0.81 (0.78–0.83) for DRS. Calibration plots showed that predicted risks were too extreme with underestimation at low risk and overestimation at high risk. Importantly, in all models age alone already showed nearly identical discriminative accuracy as the full model (C-statistics: 0.55 (0.53–0.58) for CAIDE, 0.81 (0.78–0.83) for BDSI, 0.77 (0.75–0.79) for ANU-ADRI, and 0.81 (0.78–0.83) for DRS). In this study, we found high variability in discriminative ability for predicting dementia in an elderly, community-dwelling population. All models showed similar discriminative ability when compared to prediction based on age alone. These findings highlight the urgent need for updated or new models to predict dementia risk in the general population.