Sequential diagnosis with language models

IF 47.3 1区综合性期刊 Q1 MULTIDISCIPLINARY SCIENCES

Science Pub Date : 2026-10-22 DOI:10.1126/science.aeh7930

Harsha Nori, Mayank Daswani, Christopher Kelly, Scott Lundberg, Marco Tulio Ribeiro, Marc Wilson, Xiaoxuan Liu, Viknesh Sounderajah, Jonathan M. Carlson, Matthew P. Lungren, Bay Gross, Peter Hames, Mustafa Suleyman, Dominic King, Eric Horvitz

{"title":"Sequential diagnosis with language models","authors":"Harsha Nori, Mayank Daswani, Christopher Kelly, Scott Lundberg, Marco Tulio Ribeiro, Marc Wilson, Xiaoxuan Liu, Viknesh Sounderajah, Jonathan M. Carlson, Matthew P. Lungren, Bay Gross, Peter Hames, Mustafa Suleyman, Dominic King, Eric Horvitz","doi":"10.1126/science.aeh7930","DOIUrl":null,"url":null,"abstract":"<div >We explored the capabilities of frontier language models in sequential diagnosis, a diagnostic paradigm that captures the iterative nature of clinical reasoning. Given a brief clinical presentation, clinicians selectively acquired information to reach a diagnosis, enabling assessment of diagnostic accuracy and workup cost. We developed an interactive benchmark and an artificial intelligence (AI) system that simulates a panel of physicians, continually revising the differential diagnosis and selecting cost-effective tests. Using this benchmark, the system achieved 83.2% accuracy on <i>New England Journal of Medicine</i> clinicopathological conference cases and 80.3% on MIMIC-IV-ED emergency department admissions, outperforming generalist physicians (36.1 and 46.8%, respectively). Although these findings were obtained in a controlled benchmark setting, they highlight the promise of language models for supporting more accurate, lower-cost diagnoses.</div>","PeriodicalId":21678,"journal":{"name":"Science","volume":"394 6822","pages":""},"PeriodicalIF":47.3000,"publicationDate":"2026-10-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.science.org/doi/reader/10.1126/science.aeh7930","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Science","FirstCategoryId":"103","ListUrlMain":"https://www.science.org/doi/10.1126/science.aeh7930","RegionNum":1,"RegionCategory":"综合性期刊","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"MULTIDISCIPLINARY SCIENCES","Score":null,"Total":0}

引用次数: 0

Abstract

We explored the capabilities of frontier language models in sequential diagnosis, a diagnostic paradigm that captures the iterative nature of clinical reasoning. Given a brief clinical presentation, clinicians selectively acquired information to reach a diagnosis, enabling assessment of diagnostic accuracy and workup cost. We developed an interactive benchmark and an artificial intelligence (AI) system that simulates a panel of physicians, continually revising the differential diagnosis and selecting cost-effective tests. Using this benchmark, the system achieved 83.2% accuracy on New England Journal of Medicine clinicopathological conference cases and 80.3% on MIMIC-IV-ED emergency department admissions, outperforming generalist physicians (36.1 and 46.8%, respectively). Although these findings were obtained in a controlled benchmark setting, they highlight the promise of language models for supporting more accurate, lower-cost diagnoses.

查看原文本刊更多论文

用语言模型进行顺序诊断

我们探索了前沿语言模型在顺序诊断中的能力，这是一种捕捉临床推理迭代本质的诊断范式。给出简短的临床表现，临床医生选择性地获取信息以达到诊断，从而能够评估诊断的准确性和检查成本。我们开发了一个交互式基准和一个人工智能（AI）系统，模拟一组医生，不断修改鉴别诊断并选择具有成本效益的测试。使用该基准，该系统对新英格兰医学杂志临床病理会议病例的准确率为83.2%，对MIMIC-IV-ED急诊科入院的准确率为80.3%，优于全科医生（分别为36.1%和46.8%）。尽管这些发现是在一个受控的基准设置中获得的，但它们强调了语言模型在支持更准确、更低成本诊断方面的前景。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

求助全文

约1分钟内获得全文求助全文

来源期刊

Science 综合性期刊-综合性期刊

CiteScore

61.10

自引率

0.90%

发文量

审稿时长

2.1 months

期刊介绍： Science is a leading outlet for scientific news, commentary, and cutting-edge research. Through its print and online incarnations, Science reaches an estimated worldwide readership of more than one million. Science’s authorship is global too, and its articles consistently rank among the world's most cited research. Science serves as a forum for discussion of important issues related to the advancement of science by publishing material on which a consensus has been reached as well as including the presentation of minority or conflicting points of view. Accordingly, all articles published in Science—including editorials, news and comment, and book reviews—are signed and reflect the individual views of the authors and not official points of view adopted by AAAS or the institutions with which the authors are affiliated. Science seeks to publish those papers that are most influential in their fields or across fields and that will significantly advance scientific understanding. Selected papers should present novel and broadly important data, syntheses, or concepts. They should merit recognition by the wider scientific community and general public provided by publication in Science, beyond that provided by specialty journals. Science welcomes submissions from all fields of science and from any source. The editors are committed to the prompt evaluation and publication of submitted papers while upholding high standards that support reproducibility of published research. Science is published weekly; selected papers are published online ahead of print.