Decision Curve Analysis (DCA)
Assess a prediction model's clinical usefulness: compute the model's net benefit across different "threshold probabilities" and compare it with the "treat all" and "treat none" reference lines. Where the model curve is above both reference lines, deciding by that model yields a net benefit. Multiple models can be compared at once. Computed locally in your browser; data are not uploaded.
① Paste data
First row is column names, then one subject per row: an outcome column (1=event/0=no) and one or more predicted-probability columns (values 0~1, from logistic/nomogram). You can paste from Excel.
How to use & methodology
What is DCA decision curve analysis for?
It assesses whether a prediction model/marker is 'useful' in actual decisions. AUC and calibration measure how good the model itself is, but don't directly answer 'does deciding by it bring net benefit'. DCA computes net benefit across risk thresholds and compares with the 'treat all' and 'treat none' default strategies to see where the model is superior.
What is the threshold probability?
The critical probability at which you think risk is high enough to intervene (e.g. biopsy, medication). It implicitly weighs the relative costs of 'missed diagnosis' vs 'over-treatment': the higher the threshold, the less willing you are to intervene on low-risk patients. DCA shows a range of thresholds at once to cover different clinical preferences.
How do I understand net benefit?
Net benefit = true-positive proportion − false-positive proportion × (pt/(1−pt)), i.e. the benefit of correct intervention minus the cost of 'treating one who shouldn't be treated', discounted by the threshold probability. It puts benefit and cost on the same scale, so the values are directly comparable across strategies.
Where do the predicted probabilities come from, and how do I compare models?
Usually the disease/event probability (0~1) computed for each subject by logistic regression or a nomogram. Put each model's probabilities in a separate column; this tool draws one curve per model, and the model whose curve is overall higher — especially across the clinically relevant threshold range — is better.