Elamur Research⌕
☰
home / papers / 2026 / CURA: Clinical Uncertainty Risk Alignment for Language Model–Based Risk Prediction

CURA: Clinical Uncertainty Risk Alignment for Language Model–Based Risk Prediction

Paper · 2026 · Sizhe Wang, Ziqi Xu, Claire Najjuuko et al.

Summary

Fine-tuned clinical language models can predict patient risk from free-text notes, but their confidence scores are often miscalibrated. They tend to be overconfident even when wrong, which is dangerous in hospital settings. The authors present CURA, a two-stage calibration method that needs only binary outcome labels, not textual rationales. First, a domain-specific clinical LM is fine-tuned with class-weighted cross-entropy, and its embeddings are frozen. Second, an ensemble of lightweight MLP heads is trained on those embeddings using three losses:

  • a base risk loss;
  • an individual-level term that ties each patient's uncertainty to the model's estimated chance of error;
  • a cohort-aware regularizer that nudges predictions toward event rates among embedding-space neighbours, with extra weight on ambiguous cases near the decision boundary.

The cohort term can be read as cross-entropy with soft labels derived from neighbourhoods, a form of label smoothing. On MIMIC-IV risk prediction tasks and several clinical LMs, CURA improves calibration with little or no loss in discrimination. It also lowers the rate of confidently wrong "safe" predictions, which makes uncertainty-based triage more reliable.

Key points

  • Evaluated on five MIMIC-IV tasks: in-hospital, 7-day and 30-day mortality, ICU stay, and early discharge. CURA was compared against a standard fine-tuned baseline, Deep Ensembles and MC Dropout, using several clinical LMs including BioGPT.
  • CURA works as a plug-in loss. It requires no architectural changes and needs only a single shared backbone at inference, unlike methods that rely on multiple stochastic passes.
  • The false reassurance rate counts truly positive patients whose uncertainty and predicted risk both fall below τ = 0.1. CURA achieved the lowest rate in 4 of 5 tasks, with large drops for 7-day and 30-day mortality. On in-hospital mortality, MC Dropout was slightly lower.
  • In selective deployment, CURA mostly gave higher AUPRC on the retained low-uncertainty cohorts. It also had fewer missed positive events per 1,000 automatically handled patients.
  • Limitations: experiments are retrospective and use single-center notes, and closed-source API models were not evaluated because of the MIMIC-IV data use agreement. Code is available at github.com/sizhe04/CURA.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Washington University in St. LouisSt Louis · United States15 works · 1,129 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksMedicine · 275 worksHealth Informatics · 214 worksHealth Professions · 75 worksHealth Information Management · 56 works

Cite

Cite as Sizhe Wang, Ziqi Xu, Claire Najjuuko, Charles Alba and Chenyang Lu. 2026. CURA: Clinical Uncertainty Risk Alignment for Language Model–Based Risk Prediction. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-long.1567

@inproceedings{wang2026cura,
  title = {{CURA: Clinical Uncertainty Risk Alignment for Language Model–Based Risk Prediction}},
  author = {Sizhe Wang, Ziqi Xu, Claire Najjuuko, Charles Alba and Chenyang Lu},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1567},
  url = {https://doi.org/10.18653/v1/2026.acl-long.1567}
}

References

4 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Moloud Abdar, Farhad Pourpanah, Sadiq Hussain, Dana Rezazadegan, Li Liu, Mohammad Ghavamzadeh, Paul Fieguth, Xiaochun Cao, Abbas Khosravi, U Rajendra Acharya, and 1 others. 2021. A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information fusion, 76:… find
  2. Charles Alba, Bing Xue, Joanna Abraham, Thomas Kannampallil, and Chenyang Lu. 2025. The foundational capabilities of large language models in predicting postoperative risks using clinical notes. npj Digital Medicine, 8(1):95. find
  3. Emily Alsentzer, John Murphy, William Boag, Wei-Hung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical bert embeddings. In Proceedings of the 2nd clinical natural language processing workshop, pages 72–78.
  4. Ziyi Chen, Mengyuan Zhang, Mustafa Mohammed Ahmed, Yi Guo, Thomas J George, Jiang Bian, and Yonghui Wu. 2025a. Narrative feature or structured feature? a study of large language models to identify cancer patients at risk of heart failure. In AMIA Annual Symposium Proceedings, volume 2024, page 242. find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wang-2026-cura-clinical-uncertainty-risk-alignment/