Abstract 3124: The Magnitude of Reclassification Indices for Individual Predictors of Global Cardiovascular Risk
Methods based on risk stratification have recently been proposed to evaluate and compare predictive risk models, but their performance in clinically relevant populations has not been demonstrated, including for well-known cardiovascular risk predictors. We fit models including traditional risk factors as well as high-sensitivity C-reactive protein (hsCRP) and parental history of myocardial infarction in prospective data from 24,558 women in the Women’s Health Study who experienced 766 cardiovascular events over an average 10-year follow-up. Fit of models with and without each risk factor was evaluated using both traditional methods, including the C index and goodness-of-fit test, and newly proposed methods for reclassification, which compare predicted risk strata for two models. Reclassification calibration compares observed to predicted risk for each model within reclassified risk strata; the net reclassification improvement (NRI) estimates the net proportion of cases moving to higher vs. lower risk strata; the clinical NRI estimates the NRI among those at intermediate risk; and the integrated discrimination improvement (IDI) compares the mean difference in predicted probabilities between cases and controls for two models. For each model set, the reclassification calibration statistic showed a significant lack of fit in models excluding each of these predictors, indicating that the predicted risk did not fit that observed in reclassified strata. Values of the NRI were 20% for age, 8–11% for hemoglobin A1c in diabetics, smoking, and systolic blood pressure, and 3–5% for lipids, hsCRP, and parental history. The clinical NRI for those initially at intermediate risk was 46% for age and ranged from 9–29% for other risk factors. The IDI was less than 2% for all predictors, including age. These data demonstrate the magnitude of several newly proposed reclassification measures for well-known cardiovascular risk factors. These measures can demonstrate the potential of new models and markers to change risk strata and alter treatment decisions, and should be considered as another means to compare models and assess the accuracy of absolute risk estimates and risk strata.