Microsoft researchers unveiled a new AI system that diagnoses patients in a step‑by‑step manner, much like a human clinician, and performed more than twice as well as general physicians on a set of 300 complex medical cases published in Science on Thursday.
The model uses an "orchestrator" that coordinates questions, physical exam findings and laboratory tests, simulating a panel of clinical experts working together to reach a diagnosis. Unlike conventional AI that processes all data at once, this approach first asks about a patient’s history, then decides which tests to order, updates hypotheses as results arrive, and may request additional investigations if needed.
In tests drawn from a major emergency‑room database, the orchestrator also outperformed human doctors. Researchers noted that the tool’s diagnostic work‑ups were not only more accurate but also cost less than those performed by clinicians. Dr. Dominic King, corporate vice president of Microsoft AI Health, said the system’s success on real‑world scenarios marks a significant step beyond the 100% pass rates seen on multiple‑choice questions.
Harsha Nori, head of AI at Microsoft AI Health, emphasized that the study measured both accuracy and the prudence of the diagnostic process. The team plans to partner with Beth Israel Deaconess Medical Center to test the tool’s ability to triage patients in emergency department waiting rooms.
As AI continues to evolve, Microsoft’s orchestrated approach could reshape how complex medical cases are handled, potentially improving patient outcomes while reducing healthcare costs.
