Overview
- The peer-reviewed study published on Thursday, October 22, 2026, found MAI‑DxO reached about 83% accuracy on challenging New England Journal of Medicine cases versus roughly 36% for individual generalist physicians in the same text‑only simulations.
- DxO is an orchestration layer that coordinates multiple AI agents to mimic a team of clinicians that ask questions, order tests, debate findings, and update diagnoses iteratively.
- Microsoft tested DxO with models from several vendors and reported the largest gains when DxO ran on weaker base models, with the best absolute results using OpenAI’s GPT‑5 and GPT‑5.5.
- The study also tracked estimated testing costs and found some DxO configurations cut per‑case diagnostic cost while raising accuracy, but gains were largest on complex cases and small or absent on many emergency cases.
- Microsoft says it will begin real‑world comparative evaluations with Beth Israel Deaconess and train a healthcare foundation model with Mayo Clinic, and it plans to open‑source the SDBench benchmark to let outside researchers validate results.