A Novel Method for Interrogating Receiver Operating Characteristic Curves for Assessing Prognostic Tests

Background: Disease prevalence is rarely explicitly considered in the early stages of the development of novel prognostic tests. Rather, researchers use the area under the receiver operating characteristic (AUROC) as the key metric to gauge and report predictive performance ability. Because this statistic does not account for disease prevalence, proposed tests may not appropriately address clinical requirements. This ultimately impedes the translation of prognostic tests into clinical practice.

Methods: A method to express positive- and/or negative predictive value criteria (PPV, NPV) within the ROC space is presented. Equations are derived for so-called equi-PPV (and equi-NPV) lines. Herewith it is possible, for any given prevalence, to plot a series of sensitivity-specificity pairs which meet a specified PPV (or NPV) criterion onto the ROC space.This concept is introduced by firstly reviewing the well-established "mechanics", strengths and limitations of the ROC analysis in the context of developing prognostic models. Then, the use of PPV (and/or) NPV criteria to augment the ROC analysis is elaborated.Additionally, an interactive web tool was also created to enable people to explore the dynamics of lines of equi-predictive value in function of prevalence. The web tool also allows to gauge what ROC curve shapes best meet specific positive and/or negative predictive value criteria (http://d4ta.link/ppvnpv/).

Results: To illustrate the merits and implications of this concept, an example on the prediction of pre-eclampsia risk in low-risk nulliparous pregnancies is elaborated.

Conclusions: In risk stratification, the clinical usefulness of a prognostic test can be expressed in positive- and negative predictive value criteria; the development of novel prognostic tests will be facilitated by the possibility to co-visualise such criteria together with ROC curves. To achieve clinically meaningful risk stratification, the development of separate tests to meet either a pre-specified positive value (rule-in) or a negative predictive value (rule-out) criteria should be considered: the characteristics of successful rule-in and rule-out tests may markedly differ.