Value of the History and Physical in Identifying Patients at Increased Risk for Coronary Artery Disease

Physicians frequently evaluate patients with symptoms that may represent angina. The initial assessment usually begins with a history, physical examination, electrocardiogram, and chest radiograph. On the basis of this initial assessment, the physician must decide whether to begin empiric therapy or to consider further evaluation with noninvasive testing, cardiac catheterization, or both. Additional testing is often justified on the grounds that much of the information collected in the initial assessment is soft data and not sufficiently precise to permit the accurate identification of patients at high or low risk. Further testing, although often justified, exposes the patient to additional risk and cost. Strategies for evaluating patients with suspected ischemic heart disease that maximize the quality of care while minimizing the use of unnecessary tests depend on the accurate identification of patients who need further evaluation. The accurate identification of high- and low-risk patients based on the physician's initial assessment would permit the development of cost-efficient strategies for evaluating patients with suspected ischemic heart disease. Stored in the Duke Database for Cardiovascular Disease is the accumulated experience at Duke of all patients with suspected coronary artery disease who were referred for cardiac catheterization [1-7]. At the time of cardiac catheterization, findings from the history, physical examination, electrocardiogram, chest radiograph, noninvasive tests, and catheterization are recorded. Patients are then prospectively followed at regular intervals. We have previously developed statistical models that use a subset of this informationthe history, physical examination, electrocardiogram, and chest radiographto estimate the anatomic severity of catheterization findings and to estimate long-term survival. Outpatients with chest pain who are evaluated in a physician's office might differ substantially from patients subsequently referred for cardiac catheterization [8]. Thus, we were not certain that models developed in the catheterization cohort would perform well when applied to outpatients. We describe the performance of models based on information from the physician's initial assessment when prospectively applied to a cohort of outpatients. We wished to determine whether a physician's office evaluation of a patient with nonacute chest pain could identify high-and low-risk patients and to evaluate the potential importance of this information in the delivery of cost-effective quality care. Methods Patients Our study sample included 1030 consecutive, symptomatic patients who had not had previous cardiac catheterization and who were referred for outpatient noninvasive testing at the Duke University Medical Center between 28 March 1983 and 31 January 1985. All patients had complete baseline evaluations that were done prospectively before testing. The sample included 602 patients referred by Duke cardiologists or fellows and 428 patients referred by other physicians at Duke or in the surrounding community. Our study sample comprised a consecutive series of patients with suspected coronary artery disease for whom the physician felt noninvasive testing was warranted. Baseline evaluations were done by a cardiology fellow or physician assistant who completed a standardized form containing all descriptors. The evaluation was facilitated by two other forms: a self-administered questionnaire completed by each patient and a referral form completed by the Duke staff cardiologist (for patients referred by cardiology staff) that together provided all descriptors. Chest pain histories were classified at the time of the patient interview by the examiner. Definitions and further descriptors have been previously described [6, 9-11]. The methods of data management and follow-up have been reported previously [6, 9-11]. In brief, baseline information was entered prospectively into the Duke Database for Cardiovascular Disease. Because missing information interrupts the clinical report process, descriptors were complete on all patients. Follow-up information was obtained at 1 and 3 years using a mailed, self-administered patient questionnaire. Patients not returning the questionnaire were contacted via telephone by trained interviewers. For patients who died, we obtained death certificates as well as physician and hospital records (including autopsy information when available), and we conducted telephone interviews with the next of kin to discuss the circumstances of the patient's death. All deaths were classified by an independent events committee (blinded to baseline information). Analysis We examined three diagnostic outcomes and one prognostic outcome. The diagnostic outcomes (available only in the 168 patients subsequently referred for cardiac catheterization within 90 days) were the presence of significant coronary artery disease ( 75% luminal diameter narrowing of at least one major coronary artery); the presence of severe coronary artery disease (the presence of significant obstruction of all three major coronary arteries or of the left main coronary artery); and the presence of significant left main coronary artery obstruction. Survival at 3 years was the prognostic end point. In the survival model, patients who were referred for angioplasty or coronary artery bypass graft surgery or who were dying of noncardiovascular causes were censored (withdrawn alive) the first time one of these events occurred. The development of the predictive models evaluated in our study has been described previously [1-7], and model details are included in the Appendix. In brief, the models were developed in consecutive series of patients referred for cardiac catheterization between 1969 and 1983; none of these patients were included in the present study. The strategy used to develop the models required the division of patients into training and test samples to minimize spurious associations. Model development in each case was done entirely in the training sample. Logistic multiple regression [12] was used for diagnostic outcomes, and the Cox proportional-hazards regression model [13, 14] was used for survival. All candidate predictor variables were examined graphically to ensure that their relation with the outcome was modeled appropriately. When nonlinearities were present that would violate model assumptions, appropriate recoding or transformation of the variables was carried out so that model assumptions were satisfied in each case. To decrease the risk for spurious relations and overfitting the models, a series of clinical indexes were developed to reflect important areas of pathophysiology [4]. Forward stepwise variable selection was used to aid in identifying important baseline predictors. Selected interactions among predictor variables were also examined. When a final model had been developed, it was tested and validated in the independent test sample. Baseline variables important for estimating each of the diagnostic and prognostic outcomes are listed in Table 1. Baseline descriptors collected for each patient were entered into each model to generate a patient-specific estimate of the probability of each outcome. Model predictions of the likelihood of significant coronary artery disease, severe coronary artery disease, left main coronary artery disease, and survival at 3 years were generated for each outpatient in this study at the time of his or her initial evaluation based solely on information collected before noninvasive testing. Table 1. Characteristics Used To Estimate Outcomes* Assessing the quality of predictions requires the use of statistics unfamiliar to most clinicians. Two components of predictive quality were examined. Reliability, the concordance between predicted and observed outcomes, was assessed by grouping all patients into quantiles of predicted risk and graphically comparing the observed prevalence of the outcome as a function of the mean predicted risk for each quantile group. Discrimination, the ability to separate patients with and without the outcome of interest, was assessed in two ways. First, the distribution of predictions for patients with and for patients without each outcome was graphically compared. Second, a concordance probability or c-index was computed [5]. The c-index is calculated by pairing each patient who has the outcome with each patient who does not have the outcome and determining the proportion of patient pairs in which the patient with the outcome had a higher estimated probability. A c-index of 0.80, for example, can be interpreted as follows: Eighty percent of the time a patient with the outcome was given a higher predicted probability of the outcome than the patient without the outcome. The c-index ranges from 1 to 0, with 1 corresponding to perfect discrimination, 0.5 to random performance of a predictor, and 0 to perfectly incorrect discrimination. For a binary outcome, the c-index equals the area under the receiver-operating characteristic (ROC) curve [15]. To further show the discrimination of the survival model, the sample was divided into subgroups of equal size based on the risk for dying within 3 years, and Kaplan-Meier [16] empirical survival curves were calculated. Placing the Results in Perspective The two approaches to describing the discriminatory ability of the models (the distribution of predictions for patients with and without the outcome and the c-index) do not effectively communicate a perspective on the importance of information. A traditional approach to showing the discriminatory ability of two tests is to compare the ROC curves of each test. Receiver-operating characteristic curves show the tradeoff between sensitivity (among patients with the outcome, the proportion with a positive test) and specificity (among patients without the outcome, the proportion with a negative test), as the threshold value above which the test is conside