Microsoft MAI-DxO lands in Science: diagnostic accuracy of about 83% on complex cases, higher than the 36% of the physician group

Top journal vindicates the "AI doctor": consultation accuracy far outstrips humans

Letter Ranking Download the app for exclusive early access October 27, 2026, 17:21:41 From Beijing Updated October 27, 2026, 17:34:29

When the other roles lean toward a certain diagnosis, it has to find the holes in it, so that everyone doesn't drift further and further down a wrong line of thinking.

This system, called MAI-DxO, recently appeared in the journal Science.

Results Microsoft announced on October 22 show that, in tests using 304 difficult cases from the New England Journal of Medicine, its diagnostic accuracy was about 83%. By comparison, the group of doctors allowed to consult reference materials had an accuracy of about 36%.

The consultation panel also includes one member dedicated to costs. If a cheaper test can provide the same information, it will suggest a substitution, trying to cut out expensive items that add little value. The research team hopes this division of labor will both help the AI pinpoint the cause of illness and reduce unnecessary testing costs.

01

Five virtual doctors seated in a row

In this simulated diagnosis based on historical cases, a patient had been repeatedly falling for months, had leg weakness, bruised easily, and had bleeding gums and loose teeth.

When ordinary GPT-5 was asked to "diagnose," it focused its attention on neuropathy and autonomic problems and ordered a series of tests one after another.

In the end, it gave an explanation involving multiple factors—saying everything amounted to saying nothing, and it failed to pinpoint the cause.

MAI-DxO, handling the same case, noticed another set of clues worth connecting: the patient had undergone gastric bypass surgery, rarely ate fruit, and had bleeding gums and lost teeth. The system therefore considered scurvy and ordered a plasma vitamin C test, which confirmed the diagnosis.

Why can MAI-DxO do this?

The core secret lies in five virtual doctors, all played by the same underlying model, each taking on one part of the diagnostic process.

"Dr. Hypothesis" is responsible for proposing diagnoses; it keeps the three most likely diseases, ranked by likelihood. When new evidence appears, it readjusts. This way, possibilities temporarily ranked lower remain within consideration.

"Dr. Test-Chooser" is responsible for choosing which tests to do. It selects at most three tests per round. For example, if two diagnoses could both explain the existing symptoms, and a test cannot help determine which one it is regardless of the result, its value is limited.

"Dr. Challenger" specializes in playing devil's advocate and pushing back; it has to watch for whether the other roles have settled on a disease too early, and raise evidence that contradicts the current judgment.

Take the earlier symptoms for a hypothetical example. If one diagnosis can explain the leg weakness but not the bleeding gums, the discussion cannot stop at "the cause of the leg weakness has been found." The remaining symptoms may mean that other causes still need to be considered. At that point, "Dr. Challenger" has to speak up.

The role of this "Dr. Challenger" is crucial. When the researchers reviewed some of the cases, they found that ordinary GPT-5 or o3 was sometimes led astray by its initial judgment and failed to adjust the diagnostic direction in time. "Dr. Challenger's" job is to remind the other roles to pay attention to clues that do not match the current judgment.

The former compares the value of different tests; when diagnostic utility is comparable, it recommends the cheaper option, and if a test is expensive and unlikely to add useful information, it can propose vetoing it. The latter checks in the background whether the reasoning is internally contradictory and whether the names of the ordered tests are valid.

After the five AI doctors discuss, the system then decides whether to continue asking about the condition, order tests, or give a diagnosis. After the test results come out, these "doctors" hold another round of discussion. If the results do not support the original judgment, other possible causes must be considered.

02

To test the effectiveness of this consultation, Microsoft designed a test called SDBench.

The difficulty of this test lies in incomplete information.

Many previous medical AI tests gave the model symptoms, medical history, and test results all at once, then asked it for an answer. In SDBench, the test subject can only see a short introduction to the condition at the start; to learn more, it has to ask follow-up questions itself. If it thinks it needs tests to obtain more information, it also has to proactively request them.

Another model responsible for providing information holds the complete case and returns the corresponding content based on the questions. For example, only after a certain blood test is requested can the corresponding result be seen. Results not recorded in the original medical record are supplemented by system simulation. After the participant makes a final diagnosis, the answer is checked, and an estimate is made of how much the previously ordered tests and visits would cost.

Through this back-and-forth, the SDBench test is completed.

The study prepared two sets of cases. One came from the New England Journal of Medicine and contained 304 difficult cases. MAI-DxO's diagnostic accuracy was about 83%. Doctors' accuracy was only about 22% when they could not consult reference materials, and about 36% when they were allowed to.

The other set consisted of 325 emergency patient cases; the system's accuracy was about 80%, and doctors' about 47%.

Both sets showed a gap, especially the difficult cases. But these cases are by nature complex and hard to diagnose, and the doctors who took part in the test were mainly from general practice and general internal medicine, completing the task independently. At the very least, we can say that, judging from this test, the AI consultation team beats human doctors working alone.

The results listed in the author's public draft are that ordinary GPT-5 had an accuracy of 78.9% on the NEJM cases, which rose to 83.2% after using MAI-DxO, an increase of 4.3 percentage points.

The change in cost was even greater. According to the study's estimates, the diagnostic process arranged by ordinary GPT-5 required an average of $8,409 per case; after adding the consultation process, this dropped to $5,430, a reduction of nearly $3,000. However, the group of doctors who were allowed to consult reference materials had an average cost of $4,725, which was still lower.

But if a budget constraint is added to the AI, MAI-DxO wins again. The system estimates costs before ordering tests, checks cumulative spending, and can cancel some tests accordingly. In a budget-controlled configuration, MAI-DxO's accuracy was 80.9%, with an average cost of $2,241, less than half that of the group of doctors allowed to consult reference materials.

03

Next stop, the emergency room

The next task Microsoft has arranged for this system is to be tested in the emergency room. The company has announced a collaboration with Beth Israel Deaconess Medical Center (BIDMC) in Boston.

Researchers plan to test whether MAI-DxO can help with triage, that is, determining which patients need priority treatment.

Google is also pushing to bring consultation AI before real patients, and its partner happens to be BIDMC as well.

On October 9, this hospital announced the AMIE study conducted in collaboration with Google. In this feasibility study, 98 patients completed AI interactions and follow-up outpatient visits. Patients first had a text conversation with the AI at home, during which it asked about symptoms and collected medical history, and the treating physician could review the conversation record or summary in advance. The entire AI conversation process was supervised in real time by a doctor.

OpenAI, the company behind GPT-5, is integrating its products into the medical record systems used by doctors. In September of this year, OpenAI announced that medical institutions could connect patient information from the electronic medical record vendor Epic's system to ChatGPT for Healthcare. Clinicians can use AI to sort through patients' test results and medication changes, and to look back at the corresponding medical records. Such features target the preparatory work doctors must do every day, allowing AI to be useful during patient consultations.

Peers are advancing real-patient research and also striving to get doctors to use their products in daily work. For Microsoft to push MAI-DxO further, it likewise needs to connect validation of diagnostic capabilities with product development.

Microsoft already has a batch of tools that have entered hospitals. Take the medical AI assistant Dragon Copilot as an example; it can help medical staff organize clinical records and generate documents such as referral letters and post-visit summaries.

In March of this year, Microsoft disclosed that more than 100,000 clinicians use Dragon Copilot in their daily work. This group of users, along with the product team's experience working with hospitals, forms the foundation for Microsoft to continue developing medical AI.

This also explains why Microsoft, while publishing the research results, announced the integration of medical-related teams, bringing scientists from Microsoft Research and product teams such as Dragon Copilot and PowerScribe together into Microsoft AI.

Microsoft's larger ambition is "medical superintelligence." According to the company's vision, diagnosis is only one step; AI must also help choose treatment plans, adjust recommendations based on subsequent changes in the condition, and take into account patients' personal wishes and the medical services actually available to them. It hopes that future systems will be able to surpass today's strongest clinical teams and medical systems.

Obviously, there is still a long way to go to reach this goal. But it must be said that the decisive and accurate yet somewhat cold cyber-diagnosis in science fiction seems to be coming into reality.

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of the Phoenix News self-media platform 'Dafeng Hao.' This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."