The comparison of prediction models for the risk of hypertension incidence among adults aged 40 years or older: the Cardiovascular Disease Association Study (CAVAS)

Abstract

[Background] Hypertension is a major public health problem worldwide and one of the major risk factors for cardiovascular and kidney diseases. The Framingham hypertension risk score (F-HTN) was firstly developed as hypertension prediction models and has been validated in several other populations. However, most validation studies were Western populations, but a few in Asian populations. In Korea, the hypertension risk prediction model was developed in Ansan and Ansung cohort (KoGES A&A) using the only risk factors same to F-HTN. However, it did not considered some other risk factors such as that have been reported to be associated with incidence of hypertension.

[Objectives] The purpose of this study is to present a more suitable model for Koreans by comparing the predicted performance between models, in detail as follows. 1) After developing a CAVAS-HTN model using the same risk factors as F-HTN in CAVAS, compare the predicted performance between F-HTN, A & A-HTN and CAVAS-HTN models. 2) And add new predictors to the F-HTN model to compare the predictive performance of the aCAVAS-HTN model developed and the CAVAS-HTN model.

[Method] Total 5,967 participants (2,115 men and 3,852 women) who were recruited between 2005 and 2011 and who did not have hypertension at baseline were included to develop and validate the hypertension prediction model. It was randomly split into 6: 4 ratios for derivation and internal validation set. In the derivation set, the risk scores for predicting incident hypertension were calculated using Weibull regression coefficients. We compared the F-HTN, A&A-HTN, CAVAS -HTN, and aCAVAS-HTN models for discrimination using C-statistics, calibration using Hosmer-Lemeshow X^2 tests (H-L X^2), and reclassification ability using the net reclassification improvement (NRI).

[Result] During follow-up (22,285 person-years), 718 hypertension cases were newly diagnosed. All four models showed good discrimination (C-statistic≥0.7; 0.757 for the F-HTN, 0.757 for the A&A-HTN, 0.762 for the CAVAS -HTN, and 0.760 for the aCAVAS-HTN), but the agreement was poor in F-HTN and A&A-HTN (H-L X^2=103.42, P < 0.0001; H-L X^2=251.0, P < 0.0001, respectively). Comparing reclassification between all four models, the NRI were slightly negative, but not significantly (F-HTN vs CAVAS-HTN was -0.00, 95% confidence interval (CI): -0.049, 0.047; A&A vs CAVAS-HTN was -0.05, 95% CI: -0.284, 0.190; CAVAS-HTN vs aCAVAS-HTN was -0.04, 95% CI: -0.067, -0.009).

[Conclusion] There may be no difference in the prediction performance among F-HTN, A&A-HTN CAVAS-HTN, and aCAVAS-HTN models. However, calibration of CAVAS-HTN and aCAVAS-HTN models showed better than F-HTN and A&A-HTN models. There seems to be no difference in prediction performance between models. Therefore, we consider that validation studies will be needed in other population.