跳到正文
原文
Nature 健康护理主题·· 22 天前AI 评分36

大型推理模型在临床诊断中的准确性-安全性权衡

The accuracy-safety trade-off of large reasoning models in clinical diagnostics

AI 导读

基于真实电子健康记录构建的AIMS-LDR基准显示,大型推理模型(LRM)较标准大语言模型(LLM)假阳性诊断增加逾34%、可重复性约低40%,但主要诊断更准确、鉴别诊断更具临床相关性。15名医师验证中,LRM扩展诊断覆盖、医师核验可靠性的结构化协作使诊断准确率提高20%、假阳性减少44%。作者称该结果来自受控回顾性条件,临床部署前仍需前瞻性验证。

正文

Nature Communications (2026) Cite this article

We’re sharing this article early to provide faster access to peer-reviewed, accepted research. It is citable and carries a permanent DOI. This version is subject to further edits and will be replaced automatically by the final Version of Record. All legal disclaimers apply.

Abstract

Large reasoning models (LRMs) offer potential advantages for clinical diagnosis over standard large language models (LLMs), yet their utility and safety remain unclear. Here we show, on AIMS-LDR, a long-context diagnostic reasoning benchmark built upon real-world electronic health records, covering six matched model pairs and 12 specialties, that LRMs generate over 34% more false positive diagnoses and show approximately 40% lower reproducibility than their LLM counterparts, while making more accurate primary diagnoses and more clinically relevant differentials. Experiments confirm this trade-off across prompt designs, model hyperparameters, and evaluation paradigms. A structured collaboration model, in which the LRM broadens diagnostic coverage and the clinician verifies reliability, increases diagnostic accuracy by 20% and reduces false positives by 44% in a 15-physician validation study. These results, obtained under controlled retrospective conditions, suggest that pairing LRM reasoning with physician oversight may support safer and more accurate diagnostic decision-making, warranting prospective validation before clinical deployment.

Acknowledgements

The authors thank the developers and maintainers of the MIMIC-IV and eICU Collaborative Research Database resources, as well as the PhysioNet platform, for making these de-identified clinical datasets available for research.

Funding

S.W., Yunxiang F., M.L. and Y.Y. were supported by the Hong Kong Research Grants Council through the General Research Fund (grant no. 17207722).

Author information

Author notes

  1. These authors contributed equally: Shuai Wu, Linwei Chen.

Authors and Affiliations

  1. School of Computing and Data Science, The University of Hong Kong, Hong Kong SAR, China

    Shuai Wu, Yunxiang Fu, Meng Lou & Yizhou Yu

  2. Hong Kong Generative AI Research & Development Center, Hong Kong SAR, China

    Linwei Chen & Yizhou Yu

  3. School of Biomedical Engineering, Tsinghua Medicine, Tsinghua University, Beijing, China

    Linwei Chen, Tien Yin Wong & Hong-Yu Zhou

  4. Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA

    Xiaoman Zhang & Luyang Luo

  5. School of Nursing, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong SAR, China

    Yangxi Huang

  6. Department of Statistics and Data Science, Southern University of Science and Technology, Shenzhen, China

    Guanhua Chen

  7. School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China

    Ying Fu

  8. Beijing Visual Science and Translational Eye Research Institute (BERI), School of Clinical Medicine, Beijing Tsinghua Changgung Hospital Eye Center, Tsinghua Medicine, Tsinghua University, Beijing, China

    Ya Xing Wang & Tien Yin Wong

  9. Vanke School of Public Health, Tsinghua University, Beijing, China

    Qian Di

  10. Department of Computer Science and Engineering, University of California, Santa Cruz, CA, USA

    Yuyin Zhou

  11. Department of Radiology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA

    Xiang Li

Authors

  1. Shuai Wu
  2. Linwei Chen
  3. Yunxiang Fu
  4. Xiaoman Zhang
  5. Luyang Luo
  6. Meng Lou
  7. Yangxi Huang
  8. Guanhua Chen
  9. Ying Fu
  10. Ya Xing Wang
  11. Qian Di
  12. Yuyin Zhou
  13. Xiang Li
  14. Tien Yin Wong
  15. Yizhou Yu
  16. Hong-Yu Zhou

Corresponding authors

Correspondence to Tien Yin Wong, Yizhou Yu or Hong-Yu Zhou.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Wu, S., Chen, L., Fu, Y. et al. The accuracy-safety trade-off of large reasoning models in clinical diagnostics. Nat Commun (2026). https://doi.org/10.1038/s41467-026-78520-x

Download citation

  • Received: 18 November 2025

  • Accepted: 28 September 2026

  • Published: 29 October 2026

  • DOI: https://doi.org/10.1038/s41467-026-78520-x

来源:Nature 健康护理主题 · nature.com