Microsoft MAI-DxO Hits 83.2% Diagnostic Accuracy, Beating Physicians at 36.1% in Science Study
Updated
Updated · Unite.AI · Oct 22
Microsoft MAI-DxO Hits 83.2% Diagnostic Accuracy, Beating Physicians at 36.1% in Science Study
3 articles · Updated · Unite.AI · Oct 22
Summary
83.2% accuracy on complex NEJM cases put Microsoft’s MAI-DxO far ahead of generalist physicians’ 36.1% open-book average in a peer-reviewed Science study, with the best individual physician reaching 51.8%.
304 NEJM cases and 325 emergency-department encounters were converted into stepwise diagnostic sessions in SDBench, letting the system ask questions, order tests and decide when to stop rather than answer from a static prompt.
80.9% accuracy at $2,241 per case and 79.6% at $1,303 showed the orchestrated system could stay highly accurate under budget constraints; on emergency-department cases it scored 80.3% versus physicians’ 46.8%.
Microsoft said no tested human or AI approach was both more accurate and cheaper, and it open-sourced parts of SDBench while launching a prospective Beth Israel Deaconess study and expanding work with Mayo Clinic.
The paper also flags important limits: all 15 authors are Microsoft employees, the benchmark is a controlled simulation, and the results do not establish superiority over specialists, teams or real-world clinical care.