Comparative Study of Prognostic Value for Coronary Disease Risk Between the U.K. Prospective Diabetes Study and Framingham Models

Response to Protopsaltis et al.

The comparison by Protopsaltis et al. (1) of the Framingham risk equations and the U.K. Prospective Diabetes Study (UKPDS) risk engine as predictors of coronary risk in diabetes is of interest, as previous analyses have shown that the Framingham equations can underestimate absolute coronary heart disease (CHD) risk in diabetic subjects by a factor of 2 or more (2,3). The sensitivity and specificity analyses presented by the authors, however, are difficult to interpret because of methodological concerns. They examined the incidence of coronary angiographically determined CHD but not “hard” CHD (defined as fatal or nonfatal myocardial infarction) as estimated by the UKPDS risk engine (4). Their use of a retrospective survivor cohort introduces bias, as patients with fatal CHD will have been excluded. This may explain the apparent poor performance of HbA_1c as a risk predictor, since myocardial infarction is more often fatal in those with higher HbA_1c (5). The use of a mixed cohort of type 1 and type 2 diabetic subjects is problematic because the UKPDS risk engine, a type 2 diabetes-specific risk calculator, has not been evaluated in subjects with type 1 diabetes. The UKPDS analysis (6) cited by the authors shows that a 1% decrement in HbA_1c was associated with a 37% risk reduction in microvascular disease, not a 10-fold reduction (90%).

A full validation of a risk model requires a prospective study of a cohort to which the model is applicable. In the case of the risk engine, a cohort with type 2 diabetes and suitable demographic characteristics (3) is needed. Covariates should be measured at the beginning of follow-up, and the cohort should be monitored for the end points addressed by the model. In the case of the risk engine, these end points would be fatal and nonfatal myocardial infarction and sudden cardiac death. Publication of the observed rate of CHD in the cohort and of the mean predicted rate according to the model would then allow an evaluation of risk tools against true rates of heart disease. An ideal study would also adjust for assay differences, if appropriate, and for regression dilution and competing risks if necessary.