Fig. 2From: An information-oriented paradigm in evaluating accuracy and agreement in radiologyReceiver operating characteristic (ROC) and global information ratio (GIR) analyses of data reported in Table 1. The blue points in the figures denotes the 6 possible cutoffs, i.e., all the category below the ith one, where \(i\in\;\left[0,5\right]\) are considered negative diagnoses. a The ROC analysis plots the cutoff points in the (1—SP) × SE space, and it connects them by using the ROC curve. The area under this curve is a cutoff-independent measure of the effectiveness of the diagnostic approach: the higher the area, the better the approach. The ROC area under the curve (AUC) ranges in the [0, 1] interval. The figure represents the ROC curve and its AUC as a black line and a dark gray region, respectively. In the depicted scenario, the AUC is about 0.793. b The GIR analysis of the same data depicts the cutoff points in the (1—SP) × IR space. The point themselves are connected by the information ratio curve (IRC) which is represented as a black line. As the ROC AUC, the IRC AUC (the dark gray region in the figure) is a measure of the effectiveness of the diagnostic approach, but, since it is computed by using IR, it is prevalence-independent. Unfortunately, it does not range in the interval [0, 1], and to normalize it, it must be divided by the IRC AUC of the best theoretical diagnostic approach, i.e., those whose sensitivity is always 1: the limit information curve (LIC). The LIC AUC (the light and dark gray regions in the figure) always equals \(2-{\pi }^{2}/6\) [4]; thus, the ratio between IRC and LIC AUCs, i.e., the global information ratio (GIR), equals IRC AUC divided by \(2-{\pi }^{2}/6\). In the scenario depicted by panel b, IRC AUC and GIR are about 0.116 and 0.326, respectivelyBack to article page