The problem of continuous endpoints for clinical decision making: a decision curve analysis approach

Decision curve analysis (DCA) and net benefit are widely used tools for assessing the clinical utility of prediction models. Traditional DCA is applied to models with a binary endpoint, such as the presence or absence of a cancer on biopsy. However, some clinical prediction models have continuous outcomes, such as life expectancy or pain reduction after treatment.

We extend the DCA to prediction models with continuous endpoints by using a threshold for the continuous outcome in place of a threshold probability of a binary outcome.

Through both a simulation and case study on pain reduction, we show that the proposed method effectively ranks different models and detects miscalibration, similar to traditional net benefit, with identical rank ordering of different models. However, net benefit calculation for a continuous endpoint assumes that utility is linear with the measurement scale, which is not always valid. Moreover, the decision-making implications of a continuous endpoint are unclear. For instance, a decision rule that patients need to have a 10-year life expectancy to be eligible for prostate cancer treatment implies that patients should only be treated if their probability of surviving 10 years is greater than approximately 50%. We question whether clinicians implementing a 10-year life expectancy rule are comfortable treating a patient with an approximately 50% probability of not surviving long enough to benefit or whether they are using "expectancy" in the lay sense of "most of the time".

We recommend that researchers convert continuous predictions into probabilities of the outcome exceeding a clinically relevant threshold and that decision makers carefully consider appropriate thresholds in light of relative misclassification costs.