Application of conditional probability analysis to the clinical diagnosis of coronary artery disease.

Analysis of multiple noninvasive tests offers the promise of more accurate diagnosis of coronary artery disease, but discordant test responses can occur frequently and, when observed, result in diagnostic uncertainty. Accordingly, 43 patients undergoing diagnostic coronary angiography were evaluated by noninvasive testing and the results subjected to analysis using Bayes' theorem of conditional probability. The procedures used included electrocardiographic stress testing for detection of exercise-induced ST segment depression, cardiokymographic stress testing for detection of exercise-induced precordial dyskinesis, myocardial perfusion scintigraphy for detection of exercise-induced relative regional hypoperfusion, and cardiac fluoroscopy for detection of coronary artery calcification. The probability for coronary artery disease was estimated by Bayes' theorem from each patient's age, sex, and symptom classification, and from the observed test responses. This analysis revealed a significant linear correlation between the predicted probability for coronary artery disease and the observed prevalence of angiographic disease over the entire range of probability from 0 to 100% (P less than 0.001 by linear regression). The 12 patients without angiographic disease had a mean posttest likelihood of only 7.0 +/- 2.6% despite the fact that 13 of the 60 historical and test responses were falsely "positive." In contrast, the mean posttest likelihood was 94.1 +/- 2.8% in the 31 patients with angiographic coronary artery disease, although 45 of the 155 historical and test responses were falsely "negative." In 8 of the 12 normal patients, the final posttest likelihood was under 10% and in 26 of the 31 coronary artery disease patients, it was over 90%. These estimates also correlated well with the pooled clinical judgment of five experienced cardiologists (P less than 0.001 by linear regression). The observed change in probability for disease for each of the 15 different test combinations correlated with their information content predicted according to Shannon's theorem (P less than 0.001 by linear regression). These results support the use of probability analysis in the clinical diagnosis of coronary artery disease and provide a formal basis for comparing the relative diagnostic effectiveness and cost-effectiveness of different test combinations.