THE UNIVERSITY OF BRITISH COLUMBIA
With the increasing availability of research data to develop prediction models and easier integration of such models at point of care via electronic health records, it is likely that multivariable risk prediction will eventually replace simple risk stratification schemes such as the frequent exacerbator phenotype (which was recently shown to be an unstable classifier). However, the clinical utility of a prediction model varies depending on the treatment threshold. As such, summary indices such as AUC and calibration slope (that are not dependent on treatment threshold) are insufficient in determining whether or not the use of risk prediction is clinically beneficial in a given treatment context.
Decision Curves can provide direct information on the net benefits of a management approach (e.g., the use of one model versus another) that can guide therapeutic choices. In addition to researchers who can use Decision Curves to investigate the clinical utility of competing risk prediction models, guideline developers can apply this methodology to investigate whether the benefits of a treatment outweigh its harms at a plausible range of treatment thresholds. We invite the respiratory research community to fully utilise Decision Curves in exploring the clinical utility of risk prediction models.