Microsoft said an experimental diagnostic system correctly identified 83% of complex cases in simulated testing, more than double the rate of generalist physicians using reference materials. The tool, called MAI-DxO, coordinates multiple AI “agents” to ask follow-up questions, propose tests and debate differential diagnoses, and was assessed on New England Journal of Medicine case reports and historical emergency-department data. The results, published in Science, hint at potential cost savings from fewer unnecessary tests in certain configurations, but the study relied on simulations, compared the AI to individual non-specialists, and involved Microsoft-affiliated authors—limitations that call for independent, real-world validation. Microsoft plans hospital trials with Beth Israel Deaconess and is collaborating with Mayo Clinic, as rivals including Google advance conversational medical AIs. Broader adoption would hinge on clinical performance, physician oversight, data governance and evolving regulatory scrutiny of AI-enabled medical tools.
Related articles:
Expert-level medical question answering via large language models (Med-PaLM 2)
Articulate Medical Intelligence Explorer (AMIE): Towards AI for Clinicians
Ethics and governance of artificial intelligence for health (WHO guidance)
