The area under the ROC curves follow the same statistics as nonparametric, comparative rank tests (Wilcoxon statistics). The significance of an AUC compared to the diagonal is therefore easy to calculate with the usual tests (Mann-Whitney's U test). The AUC (see ROC Tool 1) can also be estimated directly from these statistics: AUC = U / (N1 * N2), U test size of the Wilcoxon statistics, N1 and N2 - group sizes).
Therefore, ROC curves are not only suitable for quantitative characteristics, but also for qualitative characteristics that can be classified (ordinal scale), such as B. Findings of X-rays, scores etc.
Comparing ROC curves (test for difference from AUC, see ROC tool 2) are complex. The first step is whether the ROC curves were collected on the same patient or not (connected vs. unconnected samples). With overlapping ROC curves, it makes sense to use z. B. to compare only in a selected specificity range.
One way out is to use 2x2 contingency tables for either the same specificities (then: comparison of the sensitivities) or sensitivities (then: comparison of specificities). The discordant counts can then be compared using the McNemar test (paired sample, several tests are examined in one study; this is the rule for such evaluations) or the Chi2 test (unpaired sample).