跳到正文
原文
Nature 健康护理主题·· 2026-10-09AI 评分48

GPT-5多模态医学诊断的复杂性悖论:161例真实病例系统评估

The complexity paradox of generalist AI agents in medicine: exploring GPT-5 for real-world multimodal medical diagnosis

AI 导读

研究团队对10种GPT-5配置在161例经临床医生验证的真实病例上进行系统定量评估,考察诊断准确性、推理质量、检索准确性、证据整合及对低质量输入的鲁棒性。研究发现这些AI系统存在图像误读、网络检索不一致导致性能下降、智能体工作流透明度不足等局限,同时也展现出有效利用病史、清晰的逐步推理、合理的鉴别诊断等优势。

正文

npj Digital Medicine (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

The next generation of medical AI systems that integrate language understanding, web data retrieval, and multi-step reasoning are demonstrating increasing promise as potential diagnostic aids in technical evaluations. However, their real world diagnostic performance, explainability, and reliability remain unclear. We conducted a systematic quantitative analysis of ten GPT-5-based configurations on 161 clinician validated real world cases. Each model was assessed for diagnostic accuracy, reasoning quality, retrieval accuracy, evidence integration, and robustness to low-quality inputs. We found important limitations in these AI systems, including image misinterpretation, inconsistent web search that sometimes degraded performance, and limited transparency in agentic workflows. Yet the models also showed strengths, including effective use of patient history, clear stepwise reasoning, plausible differentials, improved specificity with explicit reasoning, and relevant evidence retrieval. To support clinical deployment and future model development, we recommend prioritizing reasoning-optimized AI models, improving information retrieval mechanisms for reflection and error correction, strengthening visual-textual integration, and establishing practical standards for image quality and clarity. Overall, reliable real world multimodal diagnostic performance depends more on transparent reasoning and high-quality inputs than on aggressive retrieval or complex agentic workflows. Persistent gaps in evidence integration and visual reasoning remain major barriers to safe routine clinical use.

Acknowledgements

This research was partly supported by the NIH Intramural Research Program, National Library of Medicine (ZL). The findings and conclusions presented in this paper are those of the author(s) and do not necessarily reflect the views of the NIH or the U.S. Department of Health and Human Services.

Author information

Author notes

  1. These authors contributed equally: Shaohui Zhang, Jiarong Qian.

Authors and Affiliations

  1. Department of Biostatistics, Epidemiology, and Informatics, University of Pennsylvania, Philadelphia, PA, USA

    Shaohui Zhang, Jiarong Qian & Jing Huang

  2. Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA

    Jiarong Qian

  3. Department of Computer Science and Engineering, Lehigh University, Bethlehem, PA, USA

    Kai Zhang, Zhiling Yan, Lifang He & Lichao Sun

  4. Department of Radiation Oncology, Stanford University, Stanford, CA, USA

    Kai Zhang & Lei Xing

  5. Department of Pediatrics, Children’s Hospital of Philadelphia, Philadelphia, PA, USA

    Michael A. Catalano, Adam Kiridly, Omar Toubat, Gregory E. Tasian & Jing Huang

  6. Department of Surgery, University of Pennsylvania, Philadelphia, PA, USA

    Michael A. Catalano, Adam Kiridly, Omar Toubat & Gregory E. Tasian

  7. Department of Radiology, Massachusetts General Hospital, Boston, MA, USA

    Quanzheng Li & Xiang Li

  8. Department of Health Outcomes and Biomedical Informatics, University of Florida, Gainesville, FL, USA

    Yonghui Wu

  9. Department of Radiation Oncology, Mayo Clinic, Rochester, MN, USA

    Wei Liu

  10. Department of Biomedical Informatics and Data Science, Yale University, New Haven, CT, USA

    Hua Xu

  11. National Library of Medicine, National Institutes of Health, Bethesda, MD, USA

    Zhiyong Lu

Authors

  1. Shaohui Zhang
  2. Jiarong Qian
  3. Kai Zhang
  4. Zhiling Yan
  5. Michael A. Catalano
  6. Adam Kiridly
  7. Omar Toubat
  8. Gregory E. Tasian
  9. Quanzheng Li
  10. Yonghui Wu
  11. Lei Xing
  12. Wei Liu
  13. Hua Xu
  14. Lifang He
  15. Xiang Li
  16. Zhiyong Lu
  17. Lichao Sun
  18. Jing Huang

Corresponding authors

Correspondence to Lichao Sun or Jing Huang.

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

Zhang, S., Qian, J., Zhang, K. et al. The complexity paradox of generalist AI agents in medicine: exploring GPT-5 for real-world  multimodal medical diagnosis. npj Digit. Med. (2026). https://doi.org/10.1038/s41746-026-03325-7

Download citation

  • Received: 24 December 2025

  • Accepted: 17 September 2026

  • Published: 09 October 2026

  • DOI: https://doi.org/10.1038/s41746-026-03325-7

来源:Nature 健康护理主题 · nature.com