The Reliability of Probability Analysis in the Prediction of Coronary Artery Disease in Two Hospitals

Abstract

To assess the accuracy of the Bayesian computer program CADENZA for the prediction of coronary artery disease, the authors examined the probabilities generated by the application of this program to the clinical and noninvasive test results of 303 patients in a private referral center and 199 patients in a veterans' hospital. These probabilities were compared with those produced by applying a six-variable discriminant function derived by logistic regression at the private referral center. Two statistical approaches were employed in evaluating the relative performances of the Bayesian program and the discriminant function. The first of these involved the sorting of patients in both test groups into ascending deciles of probability and comparing expected probability with observed angiographic disease prevalence in each decile. The second involved the calculation and comparison of a standardized reliability measure. The latter was significantly lower for the discriminant function both at the private hospital (0.200 for the discriminant function versus -17.5 +/- 1.96 for the Bayesian program) and at the veterans' hospital (-0.8 +/- 1.96 for the discriminant function versus -11.3 for Bayesian program). This suggests that the discriminant function is significantly superior to the Bayesian algorithm CADENZA for predicting coronary artery disease probabilities in subjects who have relatively high pretest disease probabilities.