Microsoft said the Microsoft AI Diagnostic Orchestrator (MAI-DxO) identified the expected diagnosis in 83% of complex medical cases in a simulation benchmark, with a hospital study planned.
The company’s announcement accompanied a paper published Oct. 22 in Science. The SDBench evaluation included two case groups: 304 complex cases from The New England Journal of Medicine and a separate set of more typical emergency department cases.
Microsoft said MAI-DxO could seek additional case information, request diagnostic tests and revise its working diagnosis as evidence arrived. It identified the expected diagnosis in 83% of the complex cases. Physicians scored 36% with textbooks and online references and 22% without those resources.
On the emergency department set, MAI-DxO scored 80% for diagnostic accuracy, compared with 47% for physicians. The complex and emergency cases were two parts of one benchmark, rather than separate studies.
SDBench recorded which evidence the system sought and estimated the costs of tests it requested, alongside diagnostic accuracy. Microsoft compared MAI-DxO with individual physicians; specialist teams were not included. The case simulations did not establish treatment quality or patient outcomes in clinical care. The scores do not show whether the system would improve decisions or health outcomes in a hospital.
A cost-optimized version scored 81% accuracy on the complex cases, Microsoft said. Its estimated cost was less than half that of comparison systems under benchmark conditions. Those calculations do not represent hospital spending or savings observed in practice.
Microsoft plans a prospective study with Beth Israel Deaconess Medical Center in Boston. Patients are expected to interact with MAI-DxO, while researchers compare its assessments with physicians’ independent evaluations. The planned work would move the evaluation from simulated cases to a hospital setting. Microsoft did not provide a start date.
