Decision Curve Analysis
A method that asks whether using a test or model actually benefits patients, rather than only how accurately it classifies them.
What it is
A test can be statistically accurate and clinically useless. Decision curve analysis (DCA) addresses that gap by incorporating the consequences of decisions rather than only their correctness.
The mechanism is a threshold probability. If you would act (treat, biopsy, resect) when the predicted probability of disease exceeds some value p, then choosing p reveals how you trade off harms: a low threshold means you consider missing disease much worse than overtreating. Net benefit is then true positives minus false positives, with false positives weighted by the odds implied by that threshold, all divided by the sample size. It is expressed in units of “true positives per patient, net of the false positives you had to accept.”
The plot shows net benefit across a range of thresholds for your model, alongside two reference strategies: treat everyone, and treat nobody. A model is useful over the threshold range where its curve sits above both references. Because reasonable clinicians disagree about thresholds, the honest output is a range, not a point.
The key reading pitfall — explicitly flagged in the group’s own documentation — is that a DCA plot is not a ROC curve. Curves crossing is normal and informative: a model may be worth using for a screening-minded clinician at a low threshold and not for a conservative one at a high threshold.
Why it matters for my work
meddecide implements DCA, net benefit, and clinical-impact curves, and its chapter frames this as one of the three core questions for evaluating any new test, biomarker, or AI model: “does using the test help the patient?” That module is the group’s recommended framing for AI versus pathologist comparisons specifically because raw accuracy does not answer it.
Where it actually lives, checked from source 2026-07-26. There is no standalone decision-curve analysis in meddecide — it is inside the ROC analyses. psychopdaROC performs “clinical utility analysis including net benefit curves, decision curve analysis, and clinical impact assessment” with a configurable treatment-threshold range, and enhancedROC adds clinical-impact metrics (NNT, NND) and net-benefit regression. Worth knowing before hunting the menu for a panel that is not there.
ClinicoPath jamovi Module adds three standalone variants the sub-module does not have: decisioncurve, bayesdca (Bayesian) and timedependentdca (time-dependent, for time-to-event outcomes — the one that matters for a survival endpoint, where a fixed-horizon net benefit is the wrong quantity).
That makes DCA directly relevant to the vendor evaluations — Paige Prostate and Aiforia Breast — where the real question is not whether the model is accurate but whether inserting it into sign-out changes outcomes for the better.
How it connects
Diagnostic Test Accuracy — DCA builds directly on sensitivity and specificity but adds the weighting that accuracy metrics deliberately omit.
Class Imbalance and Accuracy — DCA is naturally prevalence-aware, which makes it more robust than accuracy in the imbalanced settings that dominate pathology.
Biomarker Cut Points — the threshold probability in DCA is conceptually the decision-theory version of a biomarker cut-point, and reasoning about one clarifies the other.
Open questions
- Have any of the group’s AI evaluations run a DCA? Nothing recorded. [unverified]
- What threshold range is clinically defensible for the prostate and breast AI use cases? That needs a clinician-stated range before the analysis, not after.