Microsoft’s Medical AI Is Over Twice as Good Diagnosing Diseases as Doctors

Key Points:

• Microsoft researchers developed MAI-DxO, an AI diagnostic system powered by GPT-5, which demonstrated over twice the accuracy of physicians on complex medical cases and nearly four times the accuracy compared to doctors relying solely on their own knowledge.
• In testing with 304 complex New England Journal of Medicine cases and typical emergency room scenarios, MAI-DxO achieved 83% accuracy on difficult cases and 80% on ER cases, significantly outperforming experienced physicians who had about 22-47% accuracy.
• MAI-DxO not only improved diagnostic accuracy but also reduced costs, maintaining high accuracy while spending roughly $1,300 per case compared to over $4,700 by doctors, by strategically ordering tests and considering expenses during diagnosis.
• The AI system simulates a panel of specialists collaborating and iteratively refining diagnoses, with an orchestrator managing test orders and cost-benefit decisions, outperforming standalone GPT-5 and other language models like Claude, Gemini, and Grok.
• While promising, MAI-DxO's current validation is limited to controlled, text-based simulations; real-world deployment faces challenges such as incomplete data, patient preferences, and integrating into clinical workflows, with ongoing trials planned at Beth Israel Deaconess Medical Center.