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
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。
- Article
- Open access
- Published: 09 October 2026
- Shaohui Zhang1 na1,
- Jiarong Qian1,2 na1,
- Kai Zhang3,4,
- Zhiling Yan3,
- Michael A. Catalano5,6,
- Adam Kiridly5,6,
- Omar Toubat5,6,
- Gregory E. Tasian5,6,
- Quanzheng Li7,
- Yonghui Wu ORCID: orcid.org/0000-0002-6780-61358,
- Lei Xing4,
- Wei Liu9,
- Hua Xu10,
- Lifang He ORCID: orcid.org/0000-0001-7810-90713,
- Xiang Li7,
- Zhiyong Lu11,
- Lichao Sun3 &
- …
- Jing Huang ORCID: orcid.org/0000-0002-6133-69881,5
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
These authors contributed equally: Shaohui Zhang, Jiarong Qian.
Authors and Affiliations
Department of Biostatistics, Epidemiology, and Informatics, University of Pennsylvania, Philadelphia, PA, USA
Shaohui Zhang, Jiarong Qian & Jing Huang
Department of Electrical and Computer Engineering, Rice University, Houston, TX, USA
Jiarong Qian
Department of Computer Science and Engineering, Lehigh University, Bethlehem, PA, USA
Kai Zhang, Zhiling Yan, Lifang He & Lichao Sun
Department of Radiation Oncology, Stanford University, Stanford, CA, USA
Kai Zhang & Lei Xing
Department of Pediatrics, Children’s Hospital of Philadelphia, Philadelphia, PA, USA
Michael A. Catalano, Adam Kiridly, Omar Toubat, Gregory E. Tasian & Jing Huang
Department of Surgery, University of Pennsylvania, Philadelphia, PA, USA
Michael A. Catalano, Adam Kiridly, Omar Toubat & Gregory E. Tasian
Department of Radiology, Massachusetts General Hospital, Boston, MA, USA
Quanzheng Li & Xiang Li
Department of Health Outcomes and Biomedical Informatics, University of Florida, Gainesville, FL, USA
Yonghui Wu
Department of Radiation Oncology, Mayo Clinic, Rochester, MN, USA
Wei Liu
Department of Biomedical Informatics and Data Science, Yale University, New Haven, CT, USA
Hua Xu
National Library of Medicine, National Institutes of Health, Bethesda, MD, USA
Zhiyong Lu
Authors
- Shaohui Zhang
- Jiarong Qian
- Kai Zhang
- Zhiling Yan
- Michael A. Catalano
- Adam Kiridly
- Omar Toubat
- Gregory E. Tasian
- Quanzheng Li
- Yonghui Wu
- Lei Xing
- Wei Liu
- Hua Xu
- Lifang He
- Xiang Li
- Zhiyong Lu
- Lichao Sun
- 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/.
About this article
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
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