The authors have devised a meta-analytic method to summarize diagnostic test performances, which they describe along with a clinical example wherein they analyze the performances of real-time ultrasonography in eight studies of the detection of proximal deep venous thrombosis of the lower extremity, selected on the basis of specific inclusion criteria. To evaluate the evidence for fitting a summary receiver operating characteristic (ROC) curve, a test of correlation between the estimates of true-positive rate (TPR) and false-positive rate (FPR) is performed. A high positive correlation argues for summarizing the studies with an ROC curve. In the absence of such correlation, a test of homogeneity is applied separately to the estimates of sensitivity (TPR) and specificity (TNR) to evaluate whether differences among studies are due to chance alone. If the estimates are homogeneous, a summary point estimate and confidence intervals (CIs) are calculated. As a final step, subgroup analyses can be performed to assess alternative explanations of variability in TPR and FPR. Within groups defined by the presence or absence of symptoms of venous thrombosis, a negative correlation between TPR and FPR and homogeneity among studies were found. For symptomatic patients, the summary TPR was 0.97 with a 95% CI of (0.94, 0.99) and the summary TNR was 0.97 (0.95, 0.99). For asymptomatic patients, the summary TPR was 0.66 (0.50, 0.81), and the summary TNR was 0.96 (0.90, 0.99). The difference in TPR between symptomatic and asymptomatic patients was statistically significant. There was no evidence that test referral bias or lack of independent interpretation of test results influenced these findings. It is concluded that this meta-analytic method represents a quantitative approach to summarizing the performances of diagnostic tests. It provides criteria for deciding among the approaches of fitting a summary ROC curve, calculating a summary point estimate, or reporting ranges of TPRs and TNRs. A summary point estimate is introduced for use when TPR and FPR are not positively correlated but index test characteristics are homogeneous among studies. This method also allows examination of alternative explanations of differences in test performance characteristics such as differences among subgroups of patients or biases in study design.
