Microsoft’s medical diagnostic coordination system achieved an 83% accuracy rate in diagnosing complex medical conditions compared to 36% for general practitioners in a study published in the journal Science, in a move described by Mustafa Suleyman as paving the way for artificial superintelligence in medicine.
Striking results in the journal Science and the concept of medical superintelligence
How diagnostic agents work and the limits of comparison with doctors
From research preprint to clinical trials at Beth Israel Hospital
Mayo Clinic partnership and expansion of medical decision support systems
Frequently asked questions
Striking results in the journal Science and the concept of medical superintelligence
Microsoft’s AI-powered medical diagnosis system successfully solved and diagnosed 83% of complex disease cases drawn from New England Journal of Medicine records correctly during a simulation experiment. In contrast, general practitioners who were allowed to consult medical references achieved a correct answer rate of only 36%, according to a peer-reviewed scientific study published on Thursday, October 22, in the prestigious journal Science.
Mustafa Suleyman, CEO of AI at Microsoft, described this result as an important step toward achieving what he called “medical superintelligence.” However, Suleyman stressed that the tool has not yet proven its effectiveness on actual patients in a clinical setting, telling the tech publication The Verge: “It needs clinical validation, and that is a big leap, but what has been achieved represents a major milestone along that journey.”
How diagnostic agents work and the limits of comparison with doctors
The system, named “MaiDxO,” coordinates the work of multiple AI agents that collaborate together like a panel of physicians, asking the patient questions, requesting lab tests, and discussing options to arrive at an accurate diagnosis. Doctors who worked without outside references scored only about 22% on the same cases. In emergency department cases derived from admission records at Beth Israel Deaconess Medical Center, the system achieved 80% accuracy, compared to an average of 47% for human physicians, typically accomplishing this at a lower estimated cost for the requested tests.
However, this comparison involves methodological limitations; cases of this level of complexity are handled in reality by specialized physicians or integrated medical teams, and neither group was included in the comparison. The study’s own data also indicate that the bulk of the gains stem from the underlying AI models behind the system; running the “GPT-5” model alone with a simple prompt scored 78.9% on complex cases, and adding the MaiDxO coordinator raised accuracy by about four percentage points while reducing estimated costs by about $3,000 per case. In emergency cases, the coordinator added little to the powerful model, but it helped weaker models significantly, with Google’s “Gemini 2.5 Flash” model performance jumping from 46.4% to nearly 69%, noting that the paper did not test the newest frontier models.
From research preprint to clinical trials at Beth Israel Hospital
Microsoft first described the MaiDxO system in June 2025, stating at the time that the system diagnosed up to 85.5% of standard test cases, compared to 20% for some 21 physicians working without outside sources—results released at the time in a research preprint before peer review. Microsoft now plans to open-source its test benchmark, “SDBench.”
The next step is to conduct a field clinical study at Beth Israel Hospital, where patients will receive standard routine medical care, while patients who agree to participate will also interact with the MaiDxO system to compare its recommendations with the decisions of the treating physicians. Dominic King, vice president of healthcare AI at Microsoft, told TechCrunch: “We believe this will be the largest study of its kind,” indicating it is likely to begin “over the next couple of months,” without the company setting a date for the commercial release of the tool.
Mayo Clinic partnership and expansion of medical decision support systems
Mustafa Suleyman revealed that Microsoft is currently working in collaboration with the renowned Mayo Clinic to develop a healthcare foundation model planned for launch next year.
This trend puts Microsoft in direct competition alongside Google and OpenAI, which have also announced that their systems can match or outperform doctors in diagnostic tests. Suleyman concluded his remarks by saying: “We need to have the absolute best clinical medical decision support products over the next few years.”
Frequently asked questions
Question: What result did Microsoft’s MaiDxO system achieve in the Science journal study?
Answer: The system successfully diagnosed 83% of complex New England Journal of Medicine cases compared to 36% for general practitioners using references and 22% for doctors without references, and also scored 80% in emergency cases versus 47% for doctors.
Question: How does the MaiDxO system work in diagnosing medical conditions?
Answer: The system coordinates several AI agents acting as a panel of doctors that ask questions, request tests, and discuss to reach a diagnosis while reducing the estimated cost by about $3,000 per case.
Question: What are the scientific limitations pointed out by the study in comparing the system to doctors?
Answer: The study explained that complex cases are usually treated by specialized physicians or medical teams not included in the comparison, and running the GPT-5 model alone achieved 78.9% accuracy.
Question: What is the next step to verify the system’s efficiency on actual patients?
Answer: Microsoft will begin a clinical study over the next two months at Beth Israel Deaconess Medical Center to compare the system’s recommendations with physician decisions, parallel to developing a health foundation model with Mayo Clinic.
