Microsoft announced on October 22 that the research results of its artificial intelligence diagnostic system MAI-DxO have been officially published in the journal Science. Tests show that the system's accuracy in diagnosing complex cases is more than twice that of general practitioners, and it can also reduce tests and surgeries. Microsoft plans to launch clinical trials at Beth Israel Deaconess Medical Center in Boston in the coming months.
In tests on complex cases, MAI-DxO achieved a diagnostic accuracy of 83.2%, compared with 36.1% for general practitioners with access to reference materials; when doctors could not use reference materials, the accuracy was about 22%. In tests on emergency admission cases, the system's accuracy was 80.3%, versus 46.8% for general practitioners. The system performed best when using OpenAI GPT-5, and further tests adopting GPT-5.5 also showed improved performance.
MAI-DxO adopts a multi-agent architecture that simulates a team of doctors, progressively refining the differential diagnosis by proposing diagnoses, selecting tests, challenging hypotheses, and weighing costs. Microsoft's research team also proposed the Sequential Diagnosis Benchmark, SDBench, to evaluate models' diagnostic capabilities in step-by-step consultation and testing processes.
Microsoft says the system has the potential to reduce diagnostic and treatment costs to one half or even one third of the traditional approach, but the relevant cost estimates are based on simulated cases, not actual medical expenses. The research results still need to be validated by clinical trials, and AI is currently better suited to assisting medical staff rather than replacing doctors.