Overview
- A peer-reviewed paper published Oct. 22 found Microsoft’s Diagnostic Orchestrator (DxO) diagnosed curated New England Journal of Medicine cases about 83% of the time versus roughly 36% for individual generalist physicians and scored about 80% on emergency cases versus about 47% for physicians.
- DxO coordinates multiple AI agents that ask questions, decide which tests to order, update hypotheses and debate diagnoses in a sequential loop that simulates a panel of clinicians rather than a single all-at-once pass.
- Microsoft reported the largest gains when DxO ran on more capable underlying models and said OpenAI’s GPT-5 and GPT-5.5 produced the strongest results, while DxO was designed to be model-agnostic and was tested with models from several labs.
- The study is text-only, compared DxO to individual generalists rather than specialists or care teams, used modeled cost estimates, and was authored and funded by Microsoft with disclosed equity holdings, so the authors and outlets say clinical trials, regulation and governance are required before deployment.
- Microsoft plans to begin real-world evaluations with Beth Israel Deaconess Medical Center in the coming months, will work with Mayo Clinic on a healthcare foundation model, and says it will open-source the study’s SDBench so independent researchers can test the claims.