Fig. 1

Illustration of the decision curve analysis. The example illustrates the decision curve of a model to predict whether patients will have AKI, from a population with an AKI prevalence of 9%. In the decision curve analysis, the classification threshold corresponds to the cutoff above which a patient is classified as “will develop AKI.” Knowing whether the patient will or will not have AKI will trigger different therapeutic interventions. Low classification thresholds are used when the associated therapy is not harmful; hence, patients will not suffer from being classified as false positives. High classification thresholds are used when the associated therapy is toxic or has side effects, and therefore, it is important to not classify patients as at risk for AKI when they are not (thereby limiting the number of false positive classifications). Currently, preventive measures for AKI are optimization of hemodynamics and prevention of nephrotoxicity, amenable to all patients and therefore corresponding to low classification thresholds. The net benefit is a weighted measure between true and false positives depending on the classification threshold [35]. The maximum net benefit is obtained by detecting all patients who will later develop AKI; therefore, this net benefit is the prevalence of AKI in the population (9%). The line corresponding to the trivial assumption that all patients will have AKI can be drawn (classify all as AKI, traditionally called treat-all). Similarly, the minimum net benefit is obtained by considering that no patient will develop AKI and is 0 (classify none as AKI, traditionally called treat-none). To be clinically useful, a model should have a higher net benefit than the two trivial classifications. Here, being slightly above the classify all as AKI curve, the model shows usefulness in the range 0–43%. Above 43%, the model shows negative net benefit, which reflects harm and should be avoided in clinical practice. Here, the model is clinically relevant as it shows benefit for low risk thresholds corresponding to its associated preventive measures