The Microsoft AI Diagnostic Orchestrator (MAI-DxO) uses multiple AI agents to simulate a medical team. The system iteratively analyzes patient data, selects tests based on cost-effectiveness, and debates potential diagnoses.
In a study published in the journal Science, MAI-DxO achieved approximately 83% accuracy on complex cases from the New England Journal of Medicine, while generalist physicians scored between 22% and 36%.
For more than 300 emergency department cases, the system reached 80% accuracy compared to 47% for physicians with web access. Additionally, diagnostic costs were up to 20% lower in certain configurations.
Led by CEO Mustafa Suleyman, the project is model-agnostic. Microsoft plans to open-source the SDBench benchmark and intends to begin real-world hospital testing in the coming months.
The company positions MAI-DxO as a tool to assist clinicians rather than replace them. Future plans include integrating the technology into consumer products like Copilot, pending regulatory compliance.
