Despite notable improvements in risk quantification and management, cardiovascular disease (CVD) remains one of the leading causes of mortality and morbidity. A substantial proportion of CVD events is experienced by individuals below treatment thresholds established based on standard risk factors. This motivates researchers to look for new risk factors or markers that could further improve risk prediction.
The area under the receiver operating characteristic curve (AUC) has been the most commonly used measure of performance for predictive models. It quantifies discrimination, defined as the ability of the model to separate subjects who will experience the event (“events”) from those who will not (“nonevents”), or in the time-to-event context, to rank them according to their event times based on the predicted probabilities of event calculated at baseline. Notwithstanding the fact that the AUC remains a key measure of model performance, it is becoming more apparent that it may not be the most informative quantifier of the added usefulness of a new risk factor. Once the AUC reaches a reasonable level (say, 0.70), extremely large effect sizes are needed to raise it even by a small amount. Besides, the interpretation of this raise is not very intuitive.
Pencina MJ, D’Agostino RB. Evaluation of the Framingham Risk Score in the European Prospective Investigation of Cancer–Norfolk Cohort—Invited Commentary. Arch Intern Med. 2008;168(11):1216–1218. doi:10.1001/archinte.168.11.1216