微软宣布，其人工智能医疗诊断系统Diagnostic Orchestrator在复杂病例测试中取得了优于人类医生的成绩。微软AI首席执行官穆斯塔法·苏莱曼表示，公司正在朝着医疗超级智能的目标迈进，并预计未来几年内，人工智能系统在医学领域的表现有望达到与人类专家难以区分的水平。
这项研究成果于10月22日发表在国际学术期刊《科学》上。微软通过一系列医疗病例和急诊科数据，将自研AI诊断系统与人类医生进行比较，并建立了名为SDBench的评估基准。与只需要回答一道医学选择题的传统测试不同，该基准要求参与者逐步分析病情、提出问题、选择检查项目，最终给出诊断结果，试图更真实地模拟临床医生的诊断过程。
Diagnostic Orchestrator的核心思路，是让多个AI智能体协同工作，模拟由不同专业背景的医生组成的会诊团队。系统会根据患者的症状和病史，判断下一步需要询问哪些问题、进行哪些检查，并根据新获得的信息不断调整判断，直到形成最终诊断。微软希望借此突破单一AI模型在复杂医学推理方面的局限。
在以《新英格兰医学杂志》发表的数百份病例报告为基础的测试中，该系统的诊断准确率约为83%。相比之下，能够查阅参考资料的全科医生准确率约为36%，无法查阅参考资料的医生则约为22%。在另一组来自美国波士顿贝斯以色列女执事医疗中心急诊患者数据库的病例中，AI系统准确率达到80%，参与比较的人类医生平均准确率约为47%。
除了诊断准确率更高，微软还表示，该系统在测试中往往能够以较低的估算成本得出诊断结果。原因之一是系统能够根据现有信息选择检查项目，而不是不加区分地进行大量检查。不过，这些成本结果基于研究中的统一估算方法，并不意味着系统已经在现实医疗环境中证明能够直接降低医院的实际运营开支。
该系统目前由OpenAI的GPT-5提供底层能力，但微软强调，真正发挥作用的不只是基础模型本身，还有负责组织诊断流程、协调多个智能体并验证推理结果的编排机制。研究团队表示，即使基础模型持续升级，多智能体协作仍可能带来额外收益。该系统最初是在GPT-4.1仍属先进模型的阶段开发的，后续测试也显示，模型能力提升并不必然削弱协同诊断架构的价值。
不过，这些成绩仍需要谨慎看待。测试主要围绕既往病例和经过筛选的复杂诊断案例展开，尚不能证明系统在真实临床环境中也能取得同样的表现。研究还需要进一步评估其面对普通、常见症状时的可靠性，以及在信息不完整、病情不断变化等现实条件下的表现。微软也承认，系统要真正投入临床使用，还必须经过严格的临床验证。
苏莱曼将这项研究视为微软迈向“医疗超级智能”的重要一步。微软所说的医疗超级智能，不仅是让AI在某些诊断任务中超过人类医生，更是希望未来的AI系统能够在医疗服务的多个关键环节提供超越传统医疗体系的能力。
微软负责AI医疗业务的多米尼克·金表示，诊断只是公司医疗战略的一部分。微软还希望AI能够协助协调临床护理、制定治疗方案，并处理医疗服务中涉及患者沟通和人际互动的复杂问题。相关研究和公司文章还探讨了AI如何在医疗体系中承担更广泛的任务，而不仅仅是根据症状判断疾病。
微软计划与贝斯以色列女执事医疗中心合作，在医院环境中开展测试，进一步观察该系统在真实医疗流程中的表现。金表示，相关工作可能在未来几个月启动，但微软尚未公布确定的时间表。至于商业化，公司同样没有给出具体日期或产品发布计划。医疗领域涉及严格的监管、合规和患者安全要求，系统必须在完成必要验证后，才有可能逐步扩大应用范围。
苏莱曼预计，未来一年将出现一些令人印象深刻的智能系统实际部署案例，并认为AI将显著改变医疗服务的提供方式。不过，他强调，人工智能并不意味着医生将从医疗体系中消失。医生的工作不仅包括诊断，还涉及处理模糊信息、与患者及家属建立信任，以及根据具体情况作出综合判断。微软希望AI能够减少医生在重复性、繁琐事务上耗费的时间，让他们把更多精力用于患者护理。
如果微软的研究成果能够在真实临床环境中得到验证，多智能体AI诊断系统就可能成为医生处理疑难病例的重要辅助工具，并有机会提高复杂疾病的识别效率。但从实验室和既往病例测试走向日常医疗实践，仍需要经过临床验证、监管审查和长期安全评估。现阶段，这项研究展现了AI辅助诊断的潜力，却还不能据此认定人工智能已经可以普遍取代医生。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Microsoft announced that its artificially intelligent medical diagnostic system, Diagnostic Orchestrator, had achieved better results than human doctors in complex case testing. Mr. Mustafa Suleiman, Chief Executive Officer of Microsoft AI, stated that the company was moving towards the goal of medical super-intelligence and that it was expected that the performance of artificial intelligence systems in the field of medicine would reach levels that were hard to distinguish from human experts in the coming years.
The results were published in the international academic journal Science on 22 October. Microsoft compares self-study AI diagnostic systems with human doctors through a series of medical case and emergency medical data, and has established assessment benchmarks known as SDBench. Unlike the traditional test, which requires only one medical choice, the benchmark requires participants to gradually analyse the condition, ask questions, choose the examination project and eventually give the diagnosis, in an attempt to simulate the clinical doctor ' s diagnosis more realistically.
The core idea of Diagnostic Orchestrator is to bring together several AZAI agents to simulate a team of doctors from different professional backgrounds. The system will determine which questions next step needs to ask, which examinations to carry out, based on the patient ' s symptoms and history, and will adjust the judgement on the basis of newly acquired information until the final diagnosis is made. Microsoft hopes to break the limitations of a single AI model in complex medical reasoning.
The diagnostic accuracy of the system is about 83 per cent in tests based on hundreds of case reports published in the New England Medical Journal. In contrast, the accuracy rate of general practitioners who have access to reference materials is about 36 per cent, while those who do not have access to reference materials are about 22 per cent. In another group of cases from the database of emergency patients at the Israeli Deacon Medical Centre in Beth, Boston, United States, the AIS accuracy rate was 80 per cent, and the average accuracy rate of human doctors involved in the comparison was approximately 47 per cent.
In addition to higher diagnostic accuracy, Microsoft also indicated that the system was often able to produce diagnostic results at lower estimated cost during testing. One reason for this is that the system is able to select items for inspection based on available information, rather than conducting a large number of inspections without distinction. However, these cost results are based on a uniform estimation methodology in the study and do not mean that the system has proved to be able to directly reduce actual hospital operating costs in a realistic medical environment.
The system is currently provided by Open AI GPT-5, but Microsoft emphasizes that what really works is not just the underlying model itself, but also the organization of diagnostic processes, coordination of multiple AI agents and validation of reasoning. The research team indicated that even if the base model were to continue to be upgraded, additional benefits could result from multi-AI agent collaboration. The system was first developed at the stage of the GPT-4.1 still advanced model, and follow-up tests have shown that the enhancement of model capabilities does not necessarily diminish the value of the synergetic diagnostic architecture.
However, these achievements need to be viewed with caution. The tests are mainly based on past cases and screened complex diagnostic cases, which do not yet prove that the system can perform the same way in a real clinical environment. Research also needs to further assess their reliability in the face of common and common symptoms and their performance under realistic conditions such as incomplete information and changing conditions. Microsoft also recognizes that the system has to be strictly clinically validated if it is to be truly clinically used.
Sleiman saw the study as an important step towards Microsoft's “medical super-intelligence”. Microsoft's medical super-intelligence not only allows AI to outpace human doctors in some diagnostic missions, but also hopes that the future AI system will be able to provide capacity beyond the traditional health system in a number of critical stages of medical services.
According to Dominic King, who is in charge of medical operations in AI, diagnosis is only part of the company ' s medical strategy. Microsoft also hopes that AI will assist in coordinating clinical care, developing treatment programmes and addressing the complex issues of patient communication and interpersonal interaction in medical services. Research and corporate articles also explored how AI could take on a broader role in the health system, rather than just judging diseases on the basis of symptoms.
The Microsoft project, in collaboration with the Bess Israeli Women's Medical Centre, conducted testing in hospital settings to further observe the system's performance in real medical processes. Kim indicated that work might start in the coming months, but Microsoft had yet to announce a defined timetable. As for commercialization, the company also did not give a specific date or a product publish plan. In the area of health care, which involves strict supervision, compliance and patient safety requirements, the system must be able to gradually expand its scope of application after the necessary validation has been completed.
Sleiman expects some impressive cases of actual deployment of smart systems in the coming year and believes that AI will significantly change the way medical services are delivered. He stressed, however, that artificial intelligence did not mean that doctors would disappear from the medical system. The doctor ' s work involves not only diagnosis but also handling vague information, building trust with patients and their families and making comprehensive judgements on a case-by-case basis. Microsoft hopes that AI will reduce the time spent by doctors on repetitive and cumbersome matters and allow them to devote more energy to patient care.
If Microsoft ' s research results are validated in real clinical settings, the multi-AI agent AI diagnostic system can become an important support tool for doctors in dealing with difficult cases and have the opportunity to improve the effectiveness of identifying complex diseases. However, the transition from laboratory and ex-case testing to routine medical practice still requires clinical validation, supervisory review and long-term security assessment. At this stage, the study demonstrates the potential for AI-aided diagnosis, but it is not yet possible to conclude that artificial intelligence has become a general substitute for doctors.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.