Building Medical Superintelligence | Microsoft AI

The multi-agent orchestrator achieved 83% accuracy on complex NEJM cases versus 36% for physicians with reference materials, while reducing diagnostic costs.
Key points
- SDBench tests multi-step diagnostic reasoning, evidence requests, and costs across 304 NEJM cases and MIMIC-IV-ED emergency data.
- MAI-DxO coordinates virtual physician agents that debate diagnoses and weigh test value against expense over multiple rounds.
- On typical emergency department cases from MIMIC-IV-ED, MAI-DxO reached 80% accuracy compared to 47% for physicians.
- Microsoft open-sourced SDBench benchmark, Gatekeeper, Judge, and cost-estimation code for select NEJM cases.
- Microsoft partnered with Beth Israel Deaconess Medical Center to evaluate the system directly with patients in prospective hospital studies.
Why it matters
Medical AI researchers gain a standardized sequential benchmark and multi-agent architecture designed to improve clinical accuracy while lowering the cost of diagnostic workups.
Sources
- Building Medical Superintelligence | Microsoft AI (Microsoft AI, Oct 22, 2026, 6:04 PM)