Key points from article :
Microsoft has developed an AI system that correctly diagnosed complex medical cases far more often than generalist doctors, according to a peer-reviewed study published in the journal Science. The tool, called the Microsoft AI Diagnostic Orchestrator (MAI-DxO), works like a virtual panel of doctors. It starts with a patient's first symptoms, then asks follow-up questions, orders tests and revises its diagnosis step by step, keeping track of what each test would cost.
The system was tested on 304 challenging case reports from the New England Journal of Medicine and 325 emergency department cases from a US hospital database. On the journal cases, it reached the correct diagnosis 83.2% of the time, compared with 36.1% for generalist physicians who were allowed to use reference materials. On the emergency cases, it scored 80.3% against 46.8% for doctors. Microsoft says the AI also reached its diagnoses for around a half to a third of the cost of the tests doctors ordered.
Mustafa Suleyman, CEO of Microsoft AI, said he expects to see "extremely impressive deployments of intelligent systems" in hospitals and clinics next year.
However, the results come from a controlled test rather than real clinical practice. A second AI played the role of the patient, answering questions using detailed case files that doctors had already compiled. This meant the system did not have to draw out information from real patients, an important part of a doctor's job. Studies in real healthcare settings will be needed to show whether the tool improves patient care.
The study, led by Harsha Nori at Microsoft AI, is published in Science.
