This study determines whether a mathematical model can be used to assess noninvasively the extent of coronary artery disease (CAD). The model was based on stepwise multivariate discriminant analysis of data obtained in 99 patients from clinical and nonhemodynamic exercise variables, or from radionuclide determination of left ventricular function at rest or during exercise, or both. The extent of CAD was assessed by a scoring system and by the number of diseased vessels. The variables selected by this method (Q-wave infarction, exercise LV ejection fraction, change in systolic blood pressure from rest to exercise, sex and diabetes mellitus) yielded a predictive accuracy of 82% for the identification of patients with extensive CAD (score greater than or equal to 35). Slightly better results were achieved by a subgroup of 77 patients who had adequate exercise end points (exercise heart rate greater than or equal to 120 beats/min, or angina or ST depression during exercise). In these patients, the predictive accuracy was 84%. The model also identified patients with "light" CAD (score less than or equal to 10) with a predictive accuracy of 82%. Thus, noninvasive assessment of the extent of CAD is possible with a stepwise multivariate discriminant analysis of clinical, electrocardiographic and left ventricular function assessed by radionuclide ventriculography at rest and during exercise. The scoring system was superior to the conventional method of classifying patients according to the number of diseased vessels.