文 | 字母AI
微软让同一个大模型扮演五位“医生”组成会诊小组，其中一位AI医生专门负责抬杠。
在其他角色倾向于某个诊断时，它要找出其中的漏洞，免得大家沿着一条错误的思路越走越远。
这套名为MAI-DxO的系统，最近登上了《科学（Science）》杂志。
微软10月22日公布的结果显示，在用《新英格兰医学杂志》304个疑难病例开展的测试中，它的诊断准确率约为83%。作为对照，允许查阅资料的医生组，准确率约为36%。
会诊小组里，还有一位专门管花费。如果便宜的检查能提供同样的信息，它就会建议替换，尽量省掉花钱多、作用小的项目。研究团队希望，这样的分工既能帮助AI找准病因，也能减少不必要的检查费用。
接下来，微软计划与美国一家医疗中心合作，让真实患者直接与系统互动。这场由AI自己给自己挑错的会诊，究竟怎么进行？
五个虚拟医生排排坐
给AI安排几个分工不同的角色，诊断过程会有什么变化？研究团队在论文中举了一个例子。
在这次根据历史病例进行的模拟诊断中，一位患者几个月来反复跌倒、腿部无力，还容易瘀青，伴有牙龈出血和牙齿脱落。
让普通GPT-5“诊断”的时候，它把注意力放到了神经病变和自主神经问题上，接连安排了多项检查。
查到最后，它给出了一个涉及多种因素的解释，啥都说等于啥都没说，没能找准病因。
同样处理这份病例的MAI-DxO，注意到了另一组值得联系起来的线索：患者做过胃旁路手术，平时很少吃水果，又出现了牙龈出血、掉牙的情况。系统由此考虑坏血病，并安排血浆维生素C检测，确认了这一诊断。
为什么MAI-DxO能做到？
核心秘诀就在于五位虚拟医生，它们由同一个底层模型扮演，各自承担诊断过程中的一项工作。
“诊医生（Dr. Hypothesis）”负责提出诊断，它会保留最可能的三种疾病，按可能性排序。有新证据出现，就重新调整。这样一来，暂时排在后面的可能性也还留在考虑范围内。
“查医生（Dr. Test-Chooser）”负责挑选做什么检查。它每轮最多选择三项检查。比如，假设两个诊断都能解释现有症状，一项检查无论结果如何，都不能帮助判断究竟是哪一种，它的价值就有限。
“杠医生（Dr. Challenger）”专门唱反调、抬杠，它要留意其他角色是不是太早认定了某种病，并把与当前判断矛盾的证据提出来。
拿前面的症状举个假设性的例子。如果一种诊断能解释腿部无力，却解释不了牙龈出血，讨论就不能停在“腿无力的原因找到了”。剩下的症状可能意味着，还需要考虑别的病因。这时候“杠医生”就得开麦了。
这个“杠医生”的角色很关键。研究人员复盘部分病例时发现，普通GPT-5或o3有时会被最初的判断带偏，没能及时调整诊断方向。“杠医生”的任务，就是提醒其他角色留意那些与当前判断不符的线索。
另外两位AI医生分别负责盯成本和做复核。
前者会比较不同检查的价值，诊断作用相当时，建议选择便宜的方案，如果某项检查价格高，又很难增加有用信息，它可以提出否决。后者在后台检查推理是否前后矛盾，以及开出的检查项目名称是否有效。
五位AI医生讨论之后，系统再决定下一步继续问病情、安排检查，还是给出诊断。检查结果出来后，这几位“医生”会再讨论一轮。如果结果不支持原先的判断，就要考虑其他可能的病因。
“AI医生会诊”还真有用
为了检验这场会诊的效果，微软设计了一套叫SDBench的测试。
这个测试难就难在信息不完整。
以前有不少医疗AI测试会把症状、病史和检查结果一起交给模型，再让它给答案。SDBench里被测者开局只能看到一小段病情介绍，想了解更多就得自己追问。如果觉得需要通过检查来获得更多信息，也得自己主动提出。
另一个负责提供信息的模型掌握完整病例，根据提问返回相应内容。比如，申请了某项血液检查，才能看到对应的结果。原始病历没有记载的部分结果，会由系统模拟补充。等参与者作出最终诊断，再核对答案，并估算此前安排的检查和就诊需要多少钱。
这来来回回之间，SDBench测试就完成了。
首先是AI与人类的对决。
研究准备了两组病例。一组来自《新英格兰医学杂志》，包含304个疑难病例。MAI-DxO的诊断准确率约为83%。医生不能查资料时准确率仅约为22%，允许查资料后约为36%。
另一组是325个急诊患者的病例，系统准确率约为80%，医生约为47%。
两组成绩都拉开了差距，尤其是疑难病例。但这些病例本就以复杂、难诊断著称，参加测试的医生主要来自全科和综合内科，独立完成任务。我们至少可以说，从这个测试来看，AI会诊团队要胜过单打独斗的人类医生。
其次是AI会诊团队与单个AI的对决。
作者公开稿列出的结果是，普通GPT-5在NEJM病例上的准确率为78.9%，使用MAI-DxO后升至83.2%，提高4.3个百分点。
费用变化更大。按研究的估算，普通GPT-5安排的诊断过程，平均每例需要8409美元；加入会诊流程后，降到5430美元，减少了近3000美元。不过，允许查资料的医生组平均费用为4725美元，仍然更低。
不过如果再给AI加上预算约束，MAI-DxO又赢了。系统在开检查前估算费用，查看累计花费，并可以据此取消部分检查。在一个控制预算的配置中，MAI-DxO的准确率为80.9%，平均费用为2241美元，不到允许查资料的医生组一半。
下一站，急诊室
微软给这套系统安排的下一项任务，是到急诊室接受测试。公司已宣布与波士顿贝斯以色列女执事医疗中心（BIDMC）合作。
研究人员准备检验MAI-DxO能否帮助分诊，也就是判断哪些患者需要优先接受诊治。
毕竟在现实生活中，候诊患者的病情轻重缓急不同。AI如果能通过追问发现提示病情严重的线索，就有机会帮助医护人员更早识别这些患者。
按照微软公布的计划，患者将直接与系统互动，研究人员再把AI的表现与医生独立作出的评估比较。目前，这项合作尚未公布临床研究结果。
把问诊AI带到真实患者面前这件事，谷歌也在推进，合作方恰好也是BIDMC。
10月9日，这家医院公布了与谷歌合作的AMIE研究。在这项可行性研究中，98名患者完成了AI互动及后续门诊就诊。患者先在家里与AI进行文字对话，由它询问症状、收集病史，接诊医生可以提前查看对话记录或摘要。整个AI对话过程都有医生实时监督。
GPT-5背后的OpenAI，则在把产品接入医生使用的病历系统。今年9月，OpenAI宣布，医疗机构可以将电子病历厂商Epic系统中的患者信息接入ChatGPT for Healthcare。临床人员可以用AI梳理患者的检查结果、用药变化，并回查对应的病历记录。这类功能瞄准的是医生每天都要做的准备工作，让AI在接诊时派上用场。
从这些动作看，医疗AI的竞争已经延伸到了具体的诊疗环节。
同行在推进真实患者研究，也在争取让医生日常使用自己的产品。微软要把MAI-DxO继续往前推，同样需要把诊断能力的验证和产品开发联系起来。
微软手里已经有一批进入医院的工具。以医疗AI助手Dragon Copilot为例，它可以帮助医护人员整理诊疗记录，生成转诊信、就诊后摘要等文书。
微软今年3月披露，超过10万名临床人员在日常工作中使用Dragon Copilot。这批用户，以及产品团队与医院合作的经验，构成了微软继续开发医疗AI的基础。
这也解释了微软为什么在发布研究成果的同时，宣布整合医疗相关团队，将微软研究院的科学家，以及Dragon Copilot、PowerScribe等产品团队汇集到Microsoft AI。
从这次整合看，微软希望把诊断研究和医疗产品开发结合起来，让AI的判断能力用于医生的日常工作。
微软更大的野心是“医疗超级智能”。按照公司的设想，诊断只是其中一步，AI还要帮助选择治疗方案，根据后续病情变化调整建议，并考虑患者的个人意愿和实际能获得的医疗服务。它希望未来的系统能够超越当今最强的临床团队和医疗系统。
很明显，这个目标还有很长一段路要走。但不得不说，科幻作品中果决准确但同时略显冰冷的赛博诊疗，似乎正在照进现实。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Letter AI
Microsoft has given the same LLM a team of five “doctors”, one of whom is a doctor specializing in the challenge.
When other actors prefer a diagnosis, it identifies the gaps in it, lest you go further and further along the wrong lines.
The system, called MAI-DxO, was recently published in Science.
The results released by Microsoft on 22 October showed that its diagnosis accuracy rate was about 83 per cent of the 304 difficult cases tested in the New England Medical Journal. In contrast, the group of doctors allowed access to information had an accuracy rate of approximately 36 per cent.
There is also a dedicated management fee in the group. If a cheap inspection provided the same information, it would recommend replacement, with as little as possible cost-effective and ineffective projects. The team hoped that such a division of labour would help AI to identify the cause and reduce unnecessary examination costs.
Next, Microsoft worked with a United States medical centre to allow real patients to interact directly with the system. How do we do this by the AI.'s own mistake?
Five virtual doctors sitting in rows.
How will the diagnostic process change by assigning several different roles to AI? The research team gave an example in its paper.
During this simulation, based on historical cases, one patient fell repeatedly for several months, had weak legs and was prone to bruises, with haemorrhage from teeth and loss of teeth.
When regular GPT-5 “diagnostics”, it focuses on neuropsychiatric and autonomous neurons, with a series of examinations.
Finally, it gave a multi-factor explanation that nothing was said and that no cause could be identified.
The MAI-DxO, who also handled the case, noted another group of clues that deserved to be linked: the patient had stomach bystander surgery, was rarely eating fruit, and had haemorrhages in his teeth. The system thus considered sepsis and arranged plasma vitamin C testing, confirming the diagnosis.
Why does MAI-DxO do that?
The core secret lies in the five virtual doctors, each of whom plays a role in the diagnostic process, with the same bottom model.
The "Dr. Hypothesis" is responsible for the diagnosis, which retains the three most likely diseases, in order of likelihood. When new evidence emerges, recalibrate. In this way, the possibility of being left behind for the time being remains within consideration.
Dr. Test-Chooser is responsible for selecting what to do. It selects a maximum of three inspections per round. Assuming, for example, that both diagnoses explain existing symptoms, an examination, whatever its outcome, does not help to determine which one is, and its value is limited.
Dr. Challinger, who sings inverted and rebuffs, has to be careful whether the other players are too early to identify a disease and present evidence that contradicts the current judgement.
Take the symptoms in front and give a hypothetical example. If a diagnosis can explain the weakness of the leg, but not the bleeding of the tooth, the discussion cannot stop at “the reason why the leg is weak”. The remaining symptoms may mean that other causes need to be considered. It's time for Dr. Bark to start eating.
The role of Dr. Bark is crucial. When the researcher resets some of the cases, it was found that ordinary GPT-5 or o3 were sometimes biasd by the initial judgement and did not adjust the diagnosis in time. The mission of Dr. Bar is to alert other actors to leads that do not correspond to current judgements.
The other two Dr. AI were responsible for follow-up and review, respectively.
The former compares the value of different examinations and, when diagnostics are comparable, suggests a cheap option, which can be rejected if an examination is expensive and it is difficult to add useful information. The latter checked backstage for inconsistencies in reasoning and for validity of the project name.
Following a discussion among the five AI doctors, the system decided whether next step would continue to ask questions, arrange examinations or give a diagnosis. Once the results of the examination are available, the “doctors” will discuss another round. If the result does not support the original judgement, other possible causes of the disease will be considered.
Dr. Ai's meeting really works.
To test the results of the consultation, Microsoft designed a set of tests called SDBench.
This test is difficult to complete.
There have been a number of medical AI tests before which symptoms, medical history and the results of the tests have been presented to the model for the answer. Only a small number of medical reports are available at the start of the SDBench test, and if you want to know more, you have to ask yourself. If it is felt that more information needs to be obtained through an inspection, it will also have to be offered on its own initiative.
Another model, which provides information, captures the full number of cases and returns the corresponding content upon enquiry. For example, a blood test was requested to see the corresponding results. Partial results not documented in the original medical records will be supplemented by a system simulation. Once the final diagnosis is made by the participants, the answers will be checked and the cost of the previously scheduled examinations and consultations estimated.
Between then and again, the SDBench test was completed.
First is the battle between AI and humanity.
Two groups of cases were prepared. A group from the New England Journal of Medicine contains 304 suspected cases. The diagnostic accuracy rate for MAI-DxO is about 83 per cent. The accuracy rate when doctors are unable to access information is only about 22 per cent and about 36 per cent when information is allowed.
The other group was 325 cases of emergency cases, with a system accuracy of about 80 per cent and doctors of about 47 per cent.
Both groups of achievements have resulted in gaps, especially in difficult cases. However, these cases are well known for their complexity and difficulty of diagnosis, and the doctors participating in the tests are mainly from general and general medicine and perform their tasks independently. At least we can say that from this test, the AI team is better than a single-handed human doctor.
This is followed by an AI consultation team against individual AI.
As a result of the author ' s public draft, the standard GPT-5 accuracy rate for NEJM cases was 78.9 per cent, which rose to 83.2 per cent using MAI-DxO, an increase of 4.3 percentage points.
Costs change even more. Based on the study estimates, an average of $8409 per case is required for the diagnostic process to be arranged by regular GPT-5; after joining the consultation process, it has fallen to $5430, a reduction of nearly $3,000. However, the average cost of the group of doctors allowed to access the information was $4725, which is still lower.
But if the budget constraints were added to AI, MAI-DxO would win again. The system estimates the costs prior to the opening of an inspection, looks at the cumulative costs and can cancel some of the checks accordingly. In a controlled budget configuration, the MAI-DxO accurate rate was 80.9 per cent, at an average cost of $2241, less than half the group of doctors allowed to check.
Next stop, emergency room.
The next task that Microsoft has assigned to the system is to be tested in the emergency room. The company has announced its cooperation with the Israeli Women ' s Medical Centre (BIDMC) in Boston Beth.
Researchers are prepared to test whether MAI-DxO can help with the diagnosis, i.e. to determine which patients need priority treatment.
After all, in real life, patients waiting for treatment have different priorities. If AI is able to detect serious signs by following up, it has the opportunity to help medical personnel identify these patients earlier.
Under Microsoft's announced plan, patients will interact directly with the system, and researchers will compare AI's performance with the independent assessment of doctors. Currently, the results of the clinical study have not been published.
The idea of bringing the consultation to the real patient, Google is also advancing, and the partner happens to be the BIDMC.
On 9 October, the hospital published the AMIE study in collaboration with Google. In this feasibility study, 98 patients completed AI interaction and follow-up consultations. Patients first speak to AI at home, asking questions about symptoms and collecting medical history, and the attending physician can visit the interview record or summary in advance. The entire AI dialogue process is supervised by a doctor in real time.
Open AI, behind GPT-5, is accessing the medical records system used by doctors. In September of this year, Open AI announced that medical institutions could access information on patients in the electronic medical recordmaker Epic for Healthcare. Clinicians can comb through the results of an AI patient ' s examination, changes in the use of medication, and recheck the corresponding medical records. This type of functionality is aimed at the daily preparation of doctors to be used by AI during consultations.
In the light of these actions, the competition for medical AI has been extended to specific clinical sessions.
Peers are promoting real patient research and are seeking to get doctors to use their products on a daily basis. Microsoft continues to push MAI-DxO forward, as does the need to link the validation of diagnostic capacity with product development.
Microsoft already has a set of access tools in its hands. In the case of Dragon Copilot, a medical assistant to AI, it can help medical staff to prepare medical records, produce referral letters, post-diagnosis documents, etc.
In March of this year, Microsoft disclosed that more than 100,000 clinicians were using Dragon Copilot in their daily work. This group of users, and the experience of the product team working with the hospital, form the basis for Microsoft ' s continued development of medical AI.
This explains why Microsoft announced, in parallel with the publish study, the integration of medical-related teams, bringing together Microsoft Institute scientists and product teams like Dragon Copilot and PowerScribe into Microsoft AI.
From this integration point of view, Microsoft hopes to combine diagnostic research and the development of medical products, so that AI ' s judgement capabilities can be used in the daily work of doctors.
Microsoft's greater ambition is “medical super-intelligence”. According to the company ' s vision, the diagnosis is only one step in the process, and AI also helps to choose treatment options, adjusts recommendations to subsequent changes in the condition, and takes into account the patient ' s personal wishes and actual availability of medical services. It hoped that future systems would go beyond the strongest clinical teams and medical systems of the day.
Clearly, there is a long way to go. It must be said, however, that the sci-fi exercise, which is accurate and at the same time somewhat cold, seems to be taking reality into account.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.