Objectives: The receiver operating characteristic curves (ROC curves) are often used to compare continuous diagnostic tests or determine the optimal threshold of a test; however, they do not consider the costs of misclassifications or the disease prevalence. The ROC graph was extended to allow for these aspects.
Study design and setting: Two new lines are added to the ROC graph: a sensitivity line and a specificity line. Their slopes depend on the disease prevalence and on the ratio of the net benefit of treating a diseased subject to the net cost of treating a nondiseased one. First, these lines help researchers determine the range of specificities within which test comparisons of partial areas under the curves is clinically relevant. Second, the ROC curve point the farthest from the specificity line is shown to be the optimal threshold in terms of expected utility.
Results: This method was applied: (1) to determine the optimal threshold of ratio specific immunoglobulin G (IgG)/total IgG for the diagnosis of congenital toxoplasmosis and (2) to select, among two markers, the most accurate for the diagnosis of left ventricular hypertrophy in hypertensive subjects.
Conclusion: The two additional lines transform the statistically valid ROC graph into a clinically relevant tool for test selection and threshold determination.
Keywords: Decision analysis; Diagnostic test; Optimal threshold; ROC curve; Test selection; Utility.
Copyright © 2015 Elsevier Inc. All rights reserved.