Researchers compared two approaches for evaluating clinical prediction models: decision curve analysis and cost curves, specifically the Brier curve. Both methods assess how a model performs across decision thresholds, weighting correct and incorrect classifications by their clinical importance. The paper shows that decision curves are essentially a counterpart to Brier curves, and that reference lines such as “treat none” can be translated between the two frameworks.
The authors demonstrate that at any single threshold, net benefit (used by decision curves) and Brier loss (used by Brier curves) will always select the same model as optimal. This equivalence means practitioners can choose either approach for threshold-specific comparisons without changing conclusions. However, the measures cannot be directly compared or aggregated across different thresholds unless specific assumptions hold; doing so otherwise requires custom weightings.
They also propose an “upper envelope decision curve” as a useful benchmark, showing the maximum achievable net benefit through recalibration of the model. Since the article is based on one preprint, these findings are presented as the authors' own analysis rather than a consensus view. The paper offers practical guidance for choosing between these evaluation tools in clinical machine learning.
Source: arXiv:2509.24608 (v2, 20 Oct 2026).