The clinical utility curve: a proposal to improve the translation of information provided by prediction models to clinicians.

BACKGROUND: Prediction models are essential to the development of prediction rules that guide decision-making, and comparison of prediction models with and without an additional diagnostic or prognostic risk factor allows assessment of the value of the additional factor in risk prediction. However, the many different measures described to translate the information provided by a prediction model do not readily assist clinicians' decision-making. RESULTS: The clinical utility (CU) curve is proposed as an alternative method of communication of information from a prediction model to the clinician. The CU curve is essentially a derivation of the ROC curve that has sensitivity on the y-axis and the number needed to capture one case (NNCOC) on the x-axis. It provides information about the relationship between sensitivity and false positive rate over the full range of prediction score thresholds, and it also indicates the proportion of the patient population with an absolute risk below the threshold for 100% sensitivity that can therefore be classified as free of disease. CONCLUSIONS: The CU curve is proposed as a means to assist the translation of model information to the clinician, in the hope that it will stimulate debate and, through refinement, assist the development of prediction rules with optimal clinical utility