Abstract 20302: Implementation of AHA Pooled Cohort Equations for ASCVD Risk Prediction Requires Model Recalibration: External Validation Study in Three Large Dutch Cohorts
Introduction: In 2013 the AHA introduced the pooled cohort equations (PCE) to estimate 10-yr risk of atherosclerotic cardiovascular disease (ASCVD) as a basis for initiation of statin treatment. Before prediction models are implemented in new populations, they should be externally validated and, if needed, recalibrated. We assessed performance of the PCE models and the need for recalibration in three large population based Dutch cohorts.
Methods: For this study we included 27,971 predominantly white EPIC-NL participants aged 40-79, free from ASCVD: 5536 males (MORGEN); 6380 females (MORGEN); 16,055 females (PROSPECT). Using the PCE models we calculated 10-yr predicted risks for ASCVD for each participant, and compared with observed risks. Subsequently we recalibrated the PCE model in each of the cohorts on two levels: 1) adjusting the baseline hazard of the Cox model to the observed baseline risk (‘recal’); 2) also adjusting the coefficients by re-estimating the Cox model (‘refit’). Models were compared by discrimination and by calibration plots.
Results: After 10-yr follow-up, 225, 137 and 486 events had occurred in the MORGEN-males, MORGEN-females and PROSPECT cohorts respectively. C-statistics were 0.66 (0.59-0.73), 0.66 (0.57-0.75) and 0.69 (0.65-0.73) respectively, for both the original and recalibrated models; and 0.72 (0.66-0.78), 0.70 (0.60-0.79) and 0.75 (0.71-0.79) for the refit models. Calibration plots for the PROSPECT cohort (see fig) indicate severe underprediction by the original model, which improved after recalibration and even more after refit of the model. The other cohorts showed similar patterns
Conclusion: The PCE models showed moderate discrimination when applied to three different Dutch cohorts. Calibration of the original models was poor and would prevent use of the models for initiation of statin treatment in the Dutch population. After the necessary recalibration and notably refit of the model, model performance clearly improved.
Author Disclosures: C.A. Naaktgeboren: None. E. Schuit: None. K.G. Moons: None. W.M. Verschuren: None. J.M. Boer: None. Y.T. Van der Schouw: None. L.M. Peelen: None.
- © 2014 by American Heart Association, Inc.