Evaluating Classifier Derivation and Validation Beyond the Receiver Operating Characteristic Curve

The area under the receiver operating characteristic curve (AUROC) summarizes a classifier's discrimination across all possible thresholds, but it does not establish calibration, clinical utility, or performance at the threshold used in practice. This editorial explains the interpretation and limitations of AUROC using a hypothetical genomic enzyme test for susceptibility to symptomatic COVID-19. It emphasizes that threshold selection must be justified according to the intended use, outcome prevalence, and consequences of false-positive and false-negative results. Threshold derivation and validation should be treated as separate processes, and performance should be evaluated in independent or appropriately internally validated data. Reporting should generally include sensitivity and specificity with 95% confidence intervals and the complete confusion matrix, from a validation dataset. AUROC P-values should receive limited emphasis; effect magnitude, precision, and decision-relevant performance are more informative.