Compare Net Benefit at Clinically Meaningful Thresholds

Name the action being considered

Decision curve analysis evaluates a prediction-guided action against alternatives such as acting for everyone or no one. Specify the action, target population, outcome and time horizon. A threshold is meaningful only in relation to the consequences of acting unnecessarily and failing to act when the outcome is present.

The dcurves methodological vignette illustrates these comparisons and the importance of a relevant threshold range. A high discrimination statistic alone does not establish that using a model improves decisions, particularly when its absolute risks are poorly calibrated.

Define the calculation and comparator

For a simple fully observed binary outcome, net benefit can be expressed as TP/N − (FP/N) × p/(1−p), where p is the threshold probability. The threshold odds encode the relative weighting of false positives in this framework. Time-to-event outcomes and other sampling designs may need additional methods.

Prespecify the threshold range using the intended decision context. Include the default strategies and any realistic existing policy. Do not restrict the displayed range only to thresholds at which the new model appears favorable.

Recompute a hypothetical point

Suppose 100 fictional people include 20 with the outcome. At threshold 0.20, a model-guided rule identifies 15 true positives and 10 false positives. The threshold odds are 0.20/0.80 = 0.25. Net benefit is 0.15 − 0.10×0.25 = 0.125.

Acting for everyone gives 0.20 − 0.80×0.25 = 0, and acting for no one gives 0. Under the stipulated assumptions, the model’s net benefit is therefore 0.125 greater than either default at this threshold. This is a decision-weighted quantity, not a 12.5-percentage-point improvement in sensitivity or a direct estimate of lives saved.

Interpret the curve with its assumptions

Inspect uncertainty, calibration and the number of observations supporting each threshold. Consider implementation costs or test harms not already represented in the simple calculation. A strategy that is numerically favorable may still be impractical or unacceptable in the intended workflow.

Report the threshold rationale, outcome ascertainment and evaluation design, including whether predictions were generated independently of model fitting. Preserve both the counts and the calculation for representative thresholds. Avoid selecting a universal best threshold from the curve alone: different patients, resources and actions can imply different tradeoffs, and prospective impact may require additional evaluation.

Source verification: 6 November 2026. Examples are hypothetical.