Poor performance of historical prediction models in patients investigated for chest pain: a prospective single centre, head-to-head comparison in a large cohort of patients

BACKGROUND: An optimal investigation strategy for patients with suspected angina pectoris (AP) remains elusive. Present guidelines use the Duke Clinical Score (DCS) or the Diamond-Forrester (DF) model to compute the likelihood of coronary artery disease (CAD). This prospective study of patients referred to a chest pain clinic compares the relative values of these two historical models and of pain characteristics only to predict the presence of CAD. PATIENTS AND METHODS: Overall, 1376 patients reviewed in a chest pain clinic were assigned to five CAD likelihood groups ( 90%) using DCS and to three CAD likelihood groups ( 85%) using the DF model. Patients were diagnosed with CAD when they had either obstructive (>70%) coronary stenoses or a positive functional test. RESULTS: In all, 652 (47%) patients had nonanginal CP, 412 (30%) patients had atypical AP and 312 (23%) had typical AP. Four hundred seventeen (30%) patients were not investigated for CAD because of nonanginal symptoms and/or low CAD probability. The actual CAD prevalence was 21% versus a DCS predicted one of 51% and a DF model predicted one of 38% (P<0.001). Both models had modest predictive abilities with areas under the curve of of 0.695 and 0.693 and did not show useful clinical superiority over a prediction model using pain characteristics only (area under the curve: 0.65). CONCLUSION: CAD prevalence in patients referred for suspected AP is significantly lower than expected by using historical prediction models. The use of risk factors profile and demographics in addition to symptoms characteristics does not improve diagnostic accuracy.