Microsoft MAI-DxO Beats Doctors in Complex Diagnosis
Microsoft's MAI-DxO AI tool correctly diagnosed 83% of complex medical cases in a Science study, more than double the 36% scored by general physicians.
Introduction: Microsoft's AI team has built a diagnostic tool that diagnoses patients step by step, the way a doctor does. In a new study, it solved 83% of hard medical cases, more than twice the rate of general physicians.
What is MAI-DxO
The tool is called MAI-DxO, short for Microsoft AI Diagnostic Orchestrator. It was announced on October 22 by Microsoft's MAI health team. Instead of one AI giving an instant answer, MAI-DxO coordinates several AI agents that act like a panel of doctors with different specialties. They ask questions, order tests, review the results, and update their ideas until they reach a diagnosis.
The headline result
The team tested MAI-DxO on 304 difficult case records from the New England Journal of Medicine. The AI got 83% right. General practitioners who worked alone without reference material scored 22%. Those who could use textbooks and internet searches scored 36%. So the AI was more than twice as accurate as doctors who could look things up. The research paper, titled "Sequential diagnosis with language models," was published in the journal Science.
A second test with emergency rooms
The team also ran MAI-DxO on 325 cases drawn from emergency department admission data at Beth Israel Deaconess Medical Center in Boston. There the AI reached 80% accuracy, compared with 47% for human physicians. The researchers also found the AI's diagnostic workups cost less, because it ordered fewer unnecessary tests. Microsoft estimates diagnostic costs fall to between one-half and one-third of existing systems.
How sequential diagnosis works
Most medical AI tools work "all at once." They look at every symptom, lab result, and scan at the same time and then give one answer. MAI-DxO copies the real clinical process. It first decides what to ask about the patient's history and what matters in the physical exam. Then it picks which tests to order, studies the results, and decides what to check next. All the while, it updates its working hypotheses. Alongside the tool, Microsoft released SDBench, a new benchmark that tests AI the same sequential way instead of using simple multiple-choice questions.
Why it matters
This is a shift in how medical AI is judged. Older tests mostly used multiple-choice questions. As Microsoft AI health chief Dominic King noted, AI tools already get a 100% pass rate on those, but no patient ever walks into a clinic with a tidy summary. Microsoft says the system is model-agnostic, meaning it can run on top of language models from different companies rather than depending on one model. The company will also open the evaluation code and cost-estimation method to researchers. Microsoft AI CEO Mustafa Suleyman called the result a meaningful step toward "medical artificial superintelligence." The next test will happen with real patients: the team will work with Beth Israel Deaconess Medical Center to try the tool in emergency department waiting rooms.
The limits of the study
There are real caveats. The tests used published past cases, not real patients in live hospitals. Also, the human doctors in the comparison were internists and general practitioners, while in the real world very difficult cases usually go to specialists or full medical teams. So the headline numbers do not mean AI is ready to replace doctors. The value is narrower: a step toward tools that can support clinicians on hard cases while ordering fewer wasteful tests. Real-world results are still ahead.
Sources
Reuters via BusinessWorld - https://bworldonline.com/technology/2026/10/23/788896/microsoft-develops-ai-tool-that-can-diagnose-sequentially-like-a-doctor/ CHOSUNBIZ - https://biz.chosun.com/en/en-it/2026/10/23/SHJ5IJ6H4JBP5I5CYAIBBOT7WQ/ PBX Science - https://pbxscience.com/microsofts-mai-dxo-diagnoses-83-of-complex-cases-in-science-study-but-real-world-tests-are-still-ahead/