Overview
- The Science paper published Thursday reported Microsoft’s Diagnostic Orchestrator (DxO) reached about 83% accuracy on 304 New England Journal of Medicine clinicopathological cases and about 80% on 325 emergency department cases versus roughly 36% and 47% for generalist physicians respectively.
- DxO works sequentially by simulating a panel of clinicians: multiple AI agents ask history questions, choose tests, update diagnoses as results arrive, and coordinate next steps through an orchestration layer.
- Researchers measured performance with a new Sequential Diagnosis Benchmark (SDBench) and found the orchestration produced the biggest gains when paired with weaker base models while achieving its strongest overall scores using OpenAI’s GPT‑5 and GPT‑5.5.
- The study was run on controlled, text‑only cases and the authors and Microsoft stress that the findings do not prove superiority over specialists or care teams, note disclosed Microsoft conflicts of interest, and say clinical validation and regulatory review remain required.
- Microsoft plans real‑world evaluations with Beth Israel Deaconess Medical Center in the coming months to compare DxO’s recommendations against standard care, and the results could affect testing costs, clinician workflow, liability rules, and how hospitals adopt AI decision support.