Conformal selective prediction with cost aware deferral for safe clinical triage under distribution shift

在分布变化的情况下,采用成本感知延迟的一致性选择性预测来实现安全的临床分诊。

阅读:1

Abstract

We propose a selective prediction framework for clinical triage that combines calibrated probabilistic modeling, conformal prediction, and cost-aware deferral to prioritize patient safety. The system produces set-valued predictions with finite-sample coverage control and defers low-confidence cases to clinicians when doing so reduces an explicit expected clinical cost. We construct prediction sets using split conformal prediction, a group-conditional Mondrian variant for gender-stratified coverage, and an importance-weighted variant to improve robustness under distribution shift, and we select a single deferral threshold by minimizing expected cost on a held-out calibration set. On temporally separated in-distribution and out-of-distribution test splits for early sepsis prediction, selective deferral yields a favorable risk-coverage trade-off, reducing error on retained cases by 49.6% on the in-distribution (ID) test split and 46.7% on the out-of-distribution (OOD) test split, both at 80% coverage, while achieving low expected cost on both splits with only a moderate increase under shift. Calibration remains strong with low expected calibration error on both test sets, and rule-out performance is conservative, attaining near-perfect negative predictive value at a 95% sensitivity target. Coverage stays close to the nominal 90% target in-distribution and degrades only slightly out-of-distribution, with the weighted method most robust, and the Mondrian method reduces the gender coverage gap to 1.4 percentage points. These results indicate that conformal uncertainty quantification combined with cost-aware deferral can provide transparent and safer clinical decision support that degrades gracefully under temporal distribution shift.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。