Microsoft AI diagnostic system outperforms doctors on complex cases in study

Microsoft has developed an AI diagnostic system called MAI-DxO that outperforms generalist doctors in solving complex case simulations. In a study published in Science on October 22, 2026, MAI-DxO correctly diagnosed 83.2% of challenging New England Journal of Medicine cases compared to 36.1% for generalist physicians with references. The system simulates a panel of clinical experts to order fewer tests at lower estimated costs, and Microsoft plans clinical evaluation trials at Beth Israel Deaconess Medical Center in Boston.

Diagnostic performance on NEJM complex cases

• ▪Generalist physicians scored 36.1% diagnostic accuracy when given access to reference materials on the New England Journal of Medicine cases, and 22% accuracy without reference materials.
• ▪In the Science study, 304 complex cases published in the New England Journal of Medicine were evaluated without physical examinations or medical imaging available to the AI or human doctors.
• ▪Microsoft's MAI-DxO diagnostic system achieved an 83.2% accuracy rate on complex medical cases from the New England Journal of Medicine in a study published in Science on October 22, 2026.

Emergency department performance

• ▪Microsoft's MAI-DxO system achieved an 80.3% accuracy rate on emergency department admissions, compared to 46.8% for generalist physicians with internet access.
• ▪Running on OpenAI's GPT-5.5 model, MAI-DxO achieved about 85% accuracy on emergency department cases both with and without orchestration.

How the MAI-DxO agent architecture works

• ▪MAI-DxO coordinates multiple AI agents that simulate a panel of clinical experts to sequentially ask questions, order tests, challenge assumptions, and revise working diagnoses.
• ▪MAI-DxO works with different underlying language models, achieving its highest diagnostic performance using OpenAI's GPT-5 and GPT-5.5.
• ▪MAI-DxO increased the diagnostic accuracy of Gemini 2.5 Flash on complex cases from 46.4% on its own to nearly 69% with orchestration.

Cost reductions from test selection efficiency

• ▪MAI-DxO running on OpenAI's GPT-5 reduced the estimated per-case running cost by about $3,000 on complex cases while raising accuracy from 78.9% to over 83%.
• ▪MAI-DxO achieves lower estimated costs by judiciously ordering fewer unnecessary medical tests and procedures during sequential patient evaluations.
• ▪Researchers measured how effectively MAI-DxO selected diagnostic tests rather than evaluating only final diagnostic accuracy.
• ▪Estimated diagnostic work-up costs for MAI-DxO were up to 20% to 50% lower than those designed by human physicians in case simulations.

SDBench diagnostic benchmark

• ▪Microsoft compiled scenarios from medical case reports and emergency department data to create its SDBench benchmark.
• ▪Microsoft plans to open-source SDBench, the Sequential Diagnosis Benchmark created to evaluate diagnostic capabilities step-by-step.

Upcoming trial at Beth Israel Deaconess Medical Center

• ▪Microsoft plans to conduct real-world clinical testing of MAI-DxO with Beth Israel Deaconess Medical Center in Boston to evaluate patient triage in emergency department waiting rooms.
• ▪In the Beth Israel trial, consenting patients will interact with MAI-DxO alongside standard medical care to compare its recommendations against decisions made by human doctors.