Microsoft’s AI diagnostic system, MAI-DxO, appropriately recognized 83% of complicated New England Journal of Drugs circumstances in a peer-reviewed examine in Science, towards 36% for generalist physicians who might use reference instruments.
The paper, “Sequential diagnosis with language models”, was printed on October 22. As a substitute of handing an AI a whole case write-up, it makes the system begin from a brief presentation and resolve what to ask and which assessments to order, as a physician would.
The way it works
The staff constructed a benchmark referred to as SDBench from 304 NEJM clinicopathological circumstances and 325 emergency division admissions from the MIMIC-IV-ED dataset. A “Gatekeeper” agent releases solely what a query or take a look at would realistically reveal, and a “Judge” agent scores the ultimate prognosis.
MAI-DxO, the Microsoft AI Diagnostic Orchestrator, works like a panel of virtual physicians: every proposes diagnoses, picks questions or assessments and challenges the rising view, whereas the system weighs every take a look at’s price.
The numbers
On the NEJM circumstances, the system operating on GPT-5 scored 83.2% at a median estimated take a look at price of $5,431. Physicians averaged 36.1% with search and reference instruments and 22.3% with out. Even the perfect particular person doctor, at 51.8%, was crushed on each accuracy and value.
On extra routine emergency circumstances, MAI-DxO reached 80.3% versus 46.8% for physicians. Microsoft says the positive factors held, and grew, when it re-ran the analysis on GPT-5.5 with out altering the orchestration.
The caveats
The authors are upfront concerning the limits. The assessments had been text-only, in contrast the AI with particular person generalists slightly than specialists or groups, and measured “diagnostic reasoning under idealized information conditions rather than bedside performance or patient outcomes.” The entire authors are Microsoft workers.
That’s why the subsequent step issues. Microsoft says it’ll run a potential examine with Beth Israel Deaconess Medical Middle in Boston, the place sufferers work together with the system straight. It’s additionally open-sourcing SDBench code so others can take a look at different fashions and problem the findings.
Why it issues
GPs are already turning to ChatGPT and other tools for diagnoses, and instruments just like the NHS’s Mia cancer tool have proven AI can catch what medical doctors miss. A structured, cost-aware panel like MAI-DxO factors to a extra rigorous model of that, however till it’s confirmed with actual sufferers, it’s a benchmark outcome, not a medical one.

