Figure 4 shows a less idealized diagnostic scenario, in which there is a small degree of horses between positive Ground Truth and Ground Truth positive patients. We look at such a typical high performance test and appreciate the deterioration of apparent test power under increasing uncertainty conditions. Panel A shows the distribution of test results against soil truth. Panel B shows the expected decrease in all test performance parameters as a monotonous function of increasing comparison uncertainty. Note the generally worse apparent test performance of Figure 4 at all levels of comparative classification compared to Figure 3, where ground Truth negative and Ground Truth positive patients do not overlap in diagnostic test results. Figure 7, Panel A shows the distribution of test results for the super-unanimous subset of 290 patients (119 sepsis, 171 SIMAS), defined as patients classified as sepsis or SIRS by the three external expert panelists, as well as by researchers at clinical sites where patients were recruited. As has already been said, these patients represent the cohort layer with the lowest probability of expected error in the comparison. Panel B presents calculated performance estimates (AUC, sensitivity/AAE, specificity/APA, APP, APV), as increasing uncertainty is introduced into the comparison device. False positive (FP) and false negative (FN) rates for the Super Unanimous subset are considered null and void, as shown by the dotted left vertical line of panel B. For the consensus subgroup, the incorrect classification rates observed were 4.9% FP and 4.7% FN, which corresponds to an injection of about 4.84% of random misclassifications into the super-unanimity subset.
For the forced subset, the observed misclassification rates were 6.1% FP and 9.0% FN, which corresponds to an injection of about 7.46% of random misclassifications into the super-unanimous subset. In Panel B, the triangles indicate in the reference value the calculated values of the performance parameters (ASC, sensitivity/AAE, specificity/APA, APP, APV) after injection of the quantities of declared uncertainties (random noise of misclassification).