The multi-agent system, built on GPT-5, mimics a panel of specialists who choose the exams to prescribe. It remains to be validated in real-life conditions, starting with the Beth Israel Deaconess Hospital in Boston.
Microsoft publishes in the journal Science the results of its Diagnostic Orchestrator, a system where several AI agents play the role of a committee of experts: they examine symptoms and history, decide which tests to request and converge on a diagnosis, replicating the doctors' investigative approach. The tests are based on SDBench, a test bed built by Microsoft from published cases and emergency data, which must be open to everyone.
Net variances, at a lower estimated cost
Out of hundreds of cases from the New England Journal of Medicine, the system achieves about 83 percent accurate diagnoses, compared to nearly 36 percent for general practitioners who are allowed to consult referrals and 22 percent without assistance. On admissions to the emergency room at Beth Israel Deaconess Medical Center, it scores 80 percent compared to an average of 47 percent for doctors, with a lower estimated cost of examinations. It also exceeds state-of-the-art models used alone; designed at the time of GPT-4.1, it now works with GPT-5, and the study does not include newer models.
Microsoft acknowledges that everything is based on past cases: Mustafa Suleyman speaks of a major milestone but insists on the need for clinical validation. Hospital trials with the Beth Israel Deaconess should begin in the coming months, according to Dr. Dominic King, co-author of the study, without a firm timetable or announced marketing modalities. The company includes this work in its "medical superintelligence" project.
Read the full article on the original source
Go to the source →