Abstract

Several authors have proposed using signal detection theory for assessing the value of different clinical tests (Swets and Picket, 1982; Daubs, 1983). Two years ago at the Second Noninvasive Meeting, Masoff (1986) introduced this area of concern to our group. He argued that a criterion-free index is needed for assessing the value of a diagnostic test. He proposed using the area under the ROC curve as such an index. My paper is devoted to three aspects of this topic. 1. The dependence of d' on criterion. 2. The advantages and disadvantages of using the ROC area. An important element missing from the ROC area is the notion of a cost-benefit analysis of the test results. Cost-benefit considerations are much more important for diagnostic tests than for basic research, especially when the ROC curve deviates strongly from that expected for equal variance Gaussian distributions. 3. Optimal combination of the results of several tests. How, for example, should one combine information about the intraocular pressure, disk cupping and disk pallor in order to obtain a single index for glaucoma susceptibility. This is a difficult problem because any given disease may exhibit very different symptoms and the test results may have to be combined in a highly nonlinear fashion. The knowledge and intuition of a glaucoma expert should be incorporated into our algorithm for combining test results. Surprisingly, the ROC area has a relatively straightforward extension to multiple tests.

© 1988 Optical Society of America