The Zhitong Finance App learned that Microsoft (MSFT.US) dropped a heavy weight on the medical AI circuit. On October 22, local time, the research results of the Microsoft AI diagnostic system MAI-DxO (Microsoft AI Diagnostic Orchestrator) were officially published in “Science” magazine. This is an AI tool that can gradually diagnose patients like a doctor. The accuracy rate in solving complex medical cases is more than twice that of a general internist.
In a simulated evaluation based on the New England Journal of Medicine (NEJM) of 304 complex cases, MAi-DxO's diagnostic accuracy rate reached 83.2%, while the accuracy rate for general practitioners and general physicians who could independently access reference data was only 36.1%; when the use of reference materials was not permitted, the doctors' accuracy rate was further reduced to about 22%. In another set of tests based on MIMIC-IV-ED emergency admission data, the system accuracy rate was 80.3%, and 46.8% for general practitioners, narrowing the advantage to 1.7 times.
According to the report, the model uses an “orchestrator” to identify and coordinate questions to be asked to patients, as well as tests and test items to be arranged; the orchestrator organizes the way AI handles problems by simulating a team of clinical experts collaborating to resolve diagnosed cases.
Existing medical AI models usually use a “one-off” approach, that is, a comprehensive analysis of all of the patient's symptoms, laboratory results, and imaging tests to generate a diagnosis.
With the orchestrator coordinating the diagnostic steps, the AI model first determines what medical history information should be asked and what is important in the physical examination. Based on the information obtained, it decides which tests to schedule, evaluates the results of the tests, and arranges further inspections where appropriate. The study leader said in an interview that the model will continuously update its working assumptions and plans throughout the process.
When diagnosing more than 300 highly rare and confusing medical problems, the diagnostic orchestrator model is more than twice as accurate in “examining and evaluating” patients and determining diagnoses as general physicians using textbooks and the internet. The tool also outperforms human doctors in scenarios designed based on a large database of emergency admissions cases.
Researchers say assessments that compare such earlier systems to human doctors usually use multiple-choice questions.
“AI is already very good at this, and the pass rate can reach 100%,” said Dr. Dominic King, corporate vice president of Microsoft's AI health division and developing this new technology. “But of course it's not the real world. No patient comes to me with a nice summary.”
The researchers also found that when the orchestrator designed the examination process, the diagnostic process was less expensive than when designed by a human doctor. “We measure not only the accuracy of the system, but also how careful and meticulous it is in carrying out diagnostic tests,” said Hasha Nori, AI director at Microsoft AI Health.
In collaboration with Boston's Beth Israel Deaconess Medical Center, the researchers will next test the tool's ability to treat patients in emergency department waiting areas.
In terms of behavioral characteristics, doctors only asked an average of 5.6 questions before issuing the first test, while DxO asked an average of 25.5 questions, showing a more systematic and structured medical history collection model. In terms of the underlying model, the system performed best on OpenAI GPT-5, and performance was further improved after switching to GPT-5.5, indicating that the benefits of the orchestration layer are transferable across models.
Looking at it from a commercial perspective, the more noteworthy sign of this study is the shift in evaluation standards: Microsoft simultaneously proposed SDBench's “sequential diagnosis benchmark” to transform cases into a “consultation site” with step-by-step interaction, and for the first time included examination costs in addition to accuracy. This means that the competitive dimension of medical AI is shifting from the accuracy of individual algorithms to a comprehensive competition of “clinical reasoning orchestration capability+cost efficiency”.
This shift is in line with the overall pace of the industry. According to Grand View Research data, the global medical AI market is about 42.62 billion US dollars in 2025, and is expected to reach 56.2 billion US dollars in 2026. It is expected to expand to 1.05 trillion US dollars by 2036, with a compound growth rate of about 34% from 2026 to 2036.
On the supply side, since Microsoft acquired Nuance for nearly $20 billion in 2022, it has embedded Dragon Copilot into clinical documentation and voice workflows, covering a large number of hospitals and doctor groups in the US. MAi-DxO forms a natural puzzle extending from “document automation” to “diagnostic decision support.”
The next key verification node has surfaced. Microsoft plans to collaborate with Boston's Beth Israel Deaconess Medical Center within the next few months to launch forward-looking tests in real clinical environments, focusing on evaluating performance in emergency triage scenarios. At the same time, the complete puzzle of commercializing Microsoft's healthcare is being closed by superimposing the basic medical model plan of cooperation with the Mayo Clinic.
As far as the secondary market is concerned, short-term sentiment catalysis has been released under the endorsement of “Science,” but the valuation center of medical AI ultimately depends on whether real-world blind comparison can reproduce the advantages in a simulated environment, and the pricing and responsibility definition of “AI-assisted diagnosis” by payers and regulatory frameworks.