# Microsoft's AI doubles doctors' accuracy on complex diagnoses

> Microsoft's AI gets 83% of complex diagnoses right in a test, compared with 36% of the participating general practitioners
> By Covadonga Fernandez · Artificial Intelligence · 2026-10-23

An AI system developed by Microsoft got 83% of diagnoses right in an evaluation of complex medical cases, compared with 36% of the participating general practitioners, who were able to consult books and the internet. The results, announced yesterday, 22 October, and published in Science, come from a test using documented cases. Evaluation with patients will be the next step.

The system, called Microsoft AI Diagnostic Orchestrator (MAI-DxO), works as a panel of virtual doctors: it proposes possible diseases, decides what information to request and revises its hypotheses as it receives new data. The evaluation included 304 complex cases published in The New England Journal of Medicine.

## Asking questions and deciding on tests

The procedure reproduces part of the diagnostic process and, instead of receiving all the symptoms and results from the outset, participants start from a brief description and request information successively. The project documentation explains that an intermediary agent provides the history, examination findings and the results of the requested tests. Another agent compares the final diagnosis with the reference one. This evaluates both the answer and the path taken to reach it.

In the cases drawn from the MIMIC-IV-ED hospital database, the system reached 80% accuracy, compared with 47% for the doctors. [Microsoft now announces](https://microsoft.ai/news/building-medical-superintelligence/) a prospective study with Beth Israel Deaconess Medical Center, in Boston, to compare how the tool performs against independent assessments by professionals.

## Emergency department data

The MIMIC documentation specifies that its selected cases come from patients admitted after going to the emergency department and do not represent all consultations in that service. The simulation uses text and does not capture all the uncertainty of talking to a patient, observing their progress or interpreting medical images directly. Nor does it measure treatment safety or health outcomes. The percentages therefore show diagnostic performance within that evaluation; their clinical usefulness will have to be verified in medical practice.
