Introducing the AFOM as an alternative metric to AUC for Imbalanced Data

Diagnostic testing refers to the classification of the presence or absence of specific conditions using an index variable. Commonly utilized tools in this domain include the Receiver Operating Characteristic (ROC) curves and their Area Under the Curve (AUC). These methodologies assess the ability of various thresholds of the index variable to correctly classify a binary outcome. However, AUC can provide an overly optimistic assessment when applied to imbalanced data and may evaluate thresholds that lack practical relevance in real-world scenarios. To address these limitations, our study introduces the area under the FOM curve (AFOM) as a novel metric for evaluating diagnostic performance. The AFOM metric prioritizes the most relevant thresholds, making it less dependent on prevalence rates. We demonstrate that AFOM provides a more consistent measure of diagnostic ability when prevalence is low, favoring those thresholds with less quantity disagreement.