Limitations of a conventional logistic regression model based on left ventricular ejection fraction in predicting coronary events after myocardial infarction

Abstract

The clinical utility of conventional logistic regression models based on left ventricular ejection fraction (LVEF) for the prediction of cardiac events (death or recurrent infarction) was assessed in 646 postinfarction patients undergoing radionuclide ventriculography at rest and during exercise. The discriminant power of 2 different models (LVEF at rest alone vs LVEF at rest plus LVEF at peak exercise) was quantified in terms of the area under receiver-operating characteristic curves based on knowledge of patient outcome in the year after testing and the logistic probability of that outcome. Although LVEF at rest provided a significant amount of prognostic information (receiver-operating characteristic curve area = 62 ±4%, p < 0.001), several limitations were observed: (1) powerful predictors of risk were uncommon (32% of patients with an LVEF at rest < 0.20 had a cardiac event, but only 3% of the population had such extreme values); (2) the accuracy of predictions for high risk patients was less than for low risk patients (28 vs 98%, p < 0.001); (3) addition of exercise LVEF to the model did not improve the accuracy of prediction (receiver-operating characteristic curve area = 68 ±4%, p = 0.11); and (4) predictions for individual patients were very imprecise (the 95% confidence interval of percent risk for an LVEF at rest of 0.20 [11 to 36%] overlapped that for an LVEF at rest of 0.60 [0 to 14%]). These limitations restrict the utility of statistical prediction models when applied to individual patient decisions, and indicate that observations should not be interpreted in isolation of other clinical findings just because they have been identified as “important” by some statistical procedure.