# Microsoft AI 发布 SDBench 基准与 MAI-DxO 诊断系统,登 Science
> 原标题:Building Medical Superintelligence
- 来源:Microsoft AI:官方博客
- 发布时间:2026-10-22T18:01:00.000Z
- AIHOT:https://aihot.news/items/r7uwz6n0sd6jr2r0tb1fu3o0u
- 原文:https://microsoft.ai/news/building-medical-superintelligence/
## 摘要
Microsoft AI 发布 Sequential Diagnosis Benchmark(SDBench)和 AI 诊断系统 MAI-DxO,论文发表于 Science。
## 推荐理由
原文给出 SDBench 基准与 MAI-DxO 的具体准确率和成本数据,并开放部分评测代码,读者可据此评估其诊断推理方法。
## 正文 · 中文译文
健康
MAI 健康团队
2026年10月22日
健康
MAI 健康团队

如果 AI 能帮助临床医生逐步完成复杂的医学诊断,并以更低的成本、更少的医学检查实现更高的准确率,会怎样?
这一未来正逐渐成为现实。在美国急诊科的常见就诊病例中,使用 MIMIC-IV-ED 的病例进行评估时,我们的 AI 诊断系统 MAI-DxO 的准确率是医生的 1.7 倍。
而在发表于_The New England Journal of Medicine_的医学最复杂病例中,MAI-DxO 的准确率是使用教科书和互联网的医生的两倍以上——当医生在没有这些资源的情况下进行诊断时,其准确率接近四倍。它还能以一半甚至三分之一的成本做到这一点,用更少的检查和操作得出正确诊断。
我们的研究成果今天发表在_[Science](https://www.science.org/doi/10.1126/science.aeh7930)。_
本周还发表在[_BMJ Digital Health and AI__。_](https://www.microsoft.com/en-us/research/publication/medical-superintelligence-foundations-and-aspiration/)上的文章介绍了我们对医疗超级智能的愿景,可阅读更多内容。
## **重新思考 AI 的构建与测试方式**
AI 有潜力变革健康与医疗保健,相关进展正在迅速加快。当前的挑战在于构建并严格评估 AI 系统,将其临床推理能力的进步转化为有意义的影响。
今天的论文[建立在早期工作之上](https://microsoft.ai/news/the-path-to-medical-superintelligence/),i提出了这一转变的两项基础:顺序诊断基准(SDBench),用于测试逐步诊断推理;以及 Microsoft AI Diagnostic Orchestrator(MAI-DxO),这是我们在该基准上表现领先的 AI 系统。
诸如美国医师执照考试(USMLE)等医学执照考试旨在评估医生的知识——而非进行临床诊断的现实情况。在这些考试中,所有信息都会预先给出,并选择一个答案;而在实践中,临床医生必须逐步构建整体情况,决定要询问什么、开具哪些检查,以及每项新发现如何改变可能的诊断。此类考试已成为评估健康 AI 性能的常见基准。
这就是我们开发 SDBench 的原因,它通过两组互补的病例评估性能:取自_The New England Journal of Medicine_(NEJM)病例报告的 304 个复杂病例,以及来自 MIMIC-IV-ED 的一组更典型的急诊科就诊病例。每个病例都会成为一个交互式病例,AI 模型和/或医生从有限信息开始,并选择接下来请求什么。该基准衡量诊断准确率、请求的证据和估计的检查成本——同时评估答案和所采取的路径。
我们构建了 MAI-DxO,并在 SDBench 中对其进行了测试。它并非依赖单一模型立即给出答案,而是像一组虚拟医生那样运作:每位医生提出诊断、选择有用的问题或检查,并对逐渐形成的观点提出质疑。MAI-DxO 在数轮辩论中协调这一交流,随着证据出现更新其评估,并权衡每项检查的可能价值与成本。
**我们的发现:**
- 在复杂的 NEJM 病例上,MAI-DxO 达到了 83% 的诊断准确率。医生在没有外部参考资料的情况下达到 22%,有参考资料时为 36%,这意味着 MAI-DxO 的准确率分别约为其 3.7 倍和 2.3 倍。
- 在来自 MIMIC-IV-ED 的更典型的急诊科病例上,MAI-DxO 达到了 80% 的准确率,而医生为 47%——高出 1.7 倍以上。
- 当更侧重成本优化时,MAI-DxO 的表现依然强劲,同时选择了更具成本效益的检查方案。在 NEJM 病例上,一种配置以不到对比系统一半的成本达到了 81% 的准确率。
- 在两个数据集上,没有其他人类或 AI 方法能比 MAI-DxO 既更准确又成本更低。
- 当我们使用包括 GPT-5.5 在内的更新模型重复评估时,这些提升依然存在——并且有所改善——且未改变编排方法,这表明该方法并不绑定于某一代模型。
我们正在开源 SDBench 用于部分 NEJM 病例——包括基准、Gatekeeper、Judge 和成本估算代码——以便研究人员能够复现评估、测试其他模型,并对这些发现提出挑战或加以扩展。
## **从更好的诊断到更好的照护**
诊断只是照护旅程的一部分。更好的照护还需要跟上医学知识、选择正确的治疗、理解随时间发生的变化,并支持反映个人价值观和照护可及性的选择。机会在于将这些能力统一为一种当今尚未广泛可用的照护标准。在发表于 [_BMJ Digital Health and AI_](https://www.microsoft.com/en-us/research/publication/medical-superintelligence-foundations-and-aspiration/) 的 _The Building Blocks of Medical Superintelligence_ 一文中,我们探讨了 AI 需要具备什么才能支持这一更宏大的愿景——并最终超越当今最强临床团队和卫生系统的能力。
序贯诊断提供了一个早期示例:AI 医生协作、质疑假设,并随着证据出现完善其临床推理。我们的目标是构建在整个照护旅程中连接知识、情境和决策的 AI——而不仅仅是在某一项临床任务上表现出色。通过与 [Mayo Clinic](https://news.microsoft.com/source/2026/06/02/mayo-clinic-and-microsoft-collaborate-to-develop-a-frontier-ai-model-for-healthcare/) 的合作,我们旨在结合前沿 AI、临床专业知识和纵向多模态数据,构建能够应对真实世界照护复杂性和多样性的系统。
## **在真实世界中测试**
基准测试固然重要,但真正的考验是在真实临床环境中的表现。通过与波士顿的**Beth Israel Deaconess Medical Center**建立新的合作关系,我们将在不同医院环境中评估临床推理能力。在一项前瞻性研究中,患者将直接与系统互动,其表现将对照医生的独立评估进行衡量。与临床团队并肩工作将揭示这项技术在哪些方面取得成功、在哪些方面存在不足,以及它如何融入繁忙的工作流程。
## **微软的健康事业**
医疗 AI 的快速进步正在为患者和临床医生带来新的可能性。为了将这一进展转化为实际影响,我们正汇聚微软各健康团队——包括 Microsoft Research 的科学家以及 Dragon Copilot 和 PowerScribe 等产品背后的团队——在 Microsoft AI 形成一体化协作。通过携手合作,我们将加速交付值得信赖的 AI 系统,推动更好的临床结果和护理体验。
## 与我们共创未来
我们是一个精干、行动迅速的实验室,汇聚了世界上一些最有才华的人才。我们在 MAI 拥有一份令人振奋的算力路线图,正在快速且大规模地扩展。我们还有一个我们真正相信的雄心勃勃的使命。我们也很幸运能与出色的产品团队合作,让我们的模型有机会触达数十亿用户并创造巨大的积极影响。如果你是一位才华横溢、雄心勃勃且不摆架子的人,你会非常适合这里——快来加入我们,一起打造我们的下一代模型!
[查看所有职位](https://microsoft.ai/careers/#join-the-team)
## 正文 · 原文
Health
MAI Health Team
October 22, 2026
Health
MAI Health Team

What if AI could help clinicians work through a complex medical diagnosis, step by step, and achieve higher accuracy, at lower cost, with fewer medical tests?
That future is now coming into view. On common presentations to a US emergency department, MAI-DxO – our AI diagnostic system – is 1.7 times more accurate than physicians when evaluated using cases from MIMIC-IV-ED.
And on medicine’s most complex cases, published in _The New England Journal of Medicine_, MAI-DxO is more than twice as accurate as physicians using textbooks and the internet – and nearly four times as accurate when they diagnose without those resources. It also manages to do this at half or even one third of the cost, getting to the right diagnosis with fewer tests and procedures.
Our work is published today in _[Science](https://www.science.org/doi/10.1126/science.aeh7930)._
Read more about our vision for medical superintelligence also published this week in [_BMJ Digital Health and AI__._](https://www.microsoft.com/en-us/research/publication/medical-superintelligence-foundations-and-aspiration/)
## **Rethinking How AI is Built and Tested**
AI has the potential to transform health and healthcare, and progress is accelerating rapidly. The challenge now is to build and rigorously evaluate AI systems that translate advances in their clinical reasoning capabilities into meaningful impact.
Today’s paper, [building on earlier work](https://microsoft.ai/news/the-path-to-medical-superintelligence/), introduces two foundations for that shift: the Sequential Diagnosis Benchmark (SDBench), which tests step-by-step diagnostic reasoning, and the Microsoft AI Diagnostic Orchestrator (MAI-DxO), our leading AI system on the benchmark.
Medical licensing exams such as the United States Medical Licensing Exams (USMLEs) are designed to assess physicians’ knowledge – not the realities of making a clinical diagnosis. In these tests, all information is presented upfront and a single answer is selected; in practice, clinicians must build the picture step by step, deciding what to ask, which tests to order and how each new finding changes the likely diagnosis. Such exams have become the common benchmarks for evaluating the performance of health AI.
That’s why we developed SDBench, which evaluates performance across two complementary sets of cases: 304 complex cases taken from _The New England Journal of Medicine_ (NEJM) case proceedings and a set of more typical emergency-department presentations from MIMIC-IV-ED. Each becomes an interactive case in which AI models and/or physicians start with limited information and choose what to request next. The benchmark measures diagnostic accuracy, evidence requested and estimated test costs – assessing both the answer and the path taken.
We built MAI-DxO and tested it within SDBench. Rather than relying on one model for an immediate answer, it operates like a panel of virtual physicians: each proposing diagnoses, selecting useful questions or tests, and challenging the emerging view. MAI-DxO coordinates this exchange over several rounds of debate, updating its assessment as evidence arrives and weighing each test’s likely value against its cost.
**What we found:**
- On complex NEJM cases, MAI-DxO achieved 83% diagnostic accuracy. Physicians achieved 22% without external reference resources and 36% with them, making MAI-DxO about 3.7 and 2.3 times more accurate, respectively.
- On more typical emergency-department cases from MIMIC-IV-ED, MAI-DxO achieved 80% accuracy versus 47% for physicians – more than 1.7 times higher.
- When optimised more heavily for cost, MAI-DxO performance remained strong while choosing more cost-effective workups. On NEJM cases, one configuration achieved 81% accuracy at less than half the cost of comparative systems.
- Across both datasets, no other human or AI approach was both more accurate and less costly than MAI-DxO.
- These gains persisted – and improved – when we repeated the evaluation with newer models, including GPT-5.5, without changing the orchestration method, suggesting the approach is not tied to one model generation.
We are open-sourcing SDBench for select NEJM cases – including the benchmark, Gatekeeper, Judge and cost-estimation code – so researchers can reproduce the evaluation, test other models, and challenge or extend the findings.
## **From Better Diagnosis to Better Care**
Diagnosis is only one part of the care journey. Better care also requires keeping pace with medical knowledge, selecting the right treatment, understanding change over time, and supporting choices that reflect a person’s values and access to care. The opportunity is to unite these capabilities into a standard of care not widely available today. In _The Building Blocks of Medical Superintelligence_, published in [_BMJ Digital Health and AI_](https://www.microsoft.com/en-us/research/publication/medical-superintelligence-foundations-and-aspiration/), we explore what it would take for AI to support this broader vision – and ultimately exceed the capabilities of today’s strongest clinical teams and health systems.
Sequential diagnosis offers an early example: AI physicians collaborating, challenging assumptions and refining their clinical reasoning as evidence emerges. Our ambition is to build AI that connects knowledge, context and decisions across the care journey – not simply excels at one clinical task. Through our collaboration with [Mayo Clinic](https://news.microsoft.com/source/2026/06/02/mayo-clinic-and-microsoft-collaborate-to-develop-a-frontier-ai-model-for-healthcare/), we aim to combine frontier AI, clinical expertise and longitudinal multimodal data to build systems for the complexity and diversity of real-world care.
## **Testing in the Real World**
Benchmarks matter, but the true test is performance in real clinical environments. Through a new partnership with **Beth Israel Deaconess Medical Center** in Boston, we will evaluate clinical reasoning capabilities across different hospital settings. In a prospective study, patients will interact with the system directly, with performance measured against physicians’ independent assessments. Working alongside clinical teams will show where the technology succeeds, where it falls short and how it can fit into busy workflows.
## **Health at Microsoft**
Rapid advances in medical AI are opening new possibilities for patients and clinicians. To turn this progress into practical impact, we are bringing together health teams from across Microsoft – including scientists from Microsoft Research and the teams behind products like Dragon Copilot and PowerScribe – into an integrated effort at Microsoft AI. Working together, we will accelerate the delivery of trusted AI systems that drive better clinical outcomes and care experiences.
## Build the Future With Us
We’re a lean, fast-moving lab made up of some of the world’s most talented minds. We have an exciting roadmap of compute at MAI, which is ramping quickly and extensively. And we have an ambitious mission we truly believe in. We’re also fortunate to partner with incredible product teams giving our models the chance to reach billions of users and create immense positive impact. If you’re a brilliant, highly-ambitious and low ego individual, you’ll fit right in—come and join us as we work on our next generation of models!
[Explore all jobs](https://microsoft.ai/careers/#join-the-team)