Microsoft Publishes Peer-Reviewed Study on MAI-DxO Medical AI Orchestrator
Microsoft published peer-reviewed findings on MAI-DxO in Science. Coordinating multi-agent clinical debates, it achieved 83% accuracy on NEJM cases versus 36% for doctors with resources, and open-sourced benchmark and judge code.
Microsoft published its MAI-DxO AI orchestrator study in Science, coordinating multi-agent medical panels to debate diagnoses and evaluate test costs. Tested on SDBench using GPT-5.5 and Gemini 2.5 Flash, it achieved 83% accuracy on complex NEJM cases versus 36% for doctors with external resources (and 22% without), and 80% on emergency cases. Microsoft noted the architecture is model-agnostic and open-sourced benchmark and judge components.
Key takeaways
- MAI-DxO coordinates multi-agent clinical debates, achieving 83% accuracy in complex NEJM cases and 80% in emergency cases, exceeding tested human physicians.
- Evaluations across GPT-5.5 and Gemini 2.5 Flash demonstrated model independence; Microsoft open-sourced SDBench benchmarks and judge code.
Source：iThome Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.