Microsoft MAI-DxO in Science: What the Diagnostic AI Showed

Diagnosing a hard medical case is less like answering a quiz question and more like detective work: ask the right questions, order the right tests, and update your thinking as results arrive. Microsoft AI has now published peer-reviewed results in Science showing how its MAI-DxO system handles that process. Announced on October 22, 2026, the study reports that the diagnostic AI reached the correct answer on complex New England Journal of Medicine cases far more often than generalist physicians in a simulated setting, while ordering fewer, cheaper tests.

• Complex cases: MAI-DxO was correct on 83% of NEJM cases, versus 36% for physicians using reference materials and 22% without them
• Emergency cases: 80% accuracy on MIMIC-IV-ED emergency department cases, versus 47% for physicians
• Cost: Microsoft says the system reached diagnoses at half to one third of the cost, with fewer tests and procedures
• Next step: a prospective real-world study with Beth Israel Deaconess Medical Center in Boston

How Does the MAI-DxO Diagnostic AI Work?

MAI-DxO, short for Microsoft AI Diagnostic Orchestrator, coordinates several AI agents that act like a panel of doctors. They question a simulated patient, choose which tests to order, debate competing hypotheses over several rounds and weigh each test's likely value against its cost. It is evaluated on the Sequential Diagnosis Benchmark, or SDBench, which reveals information step by step instead of presenting everything upfront the way licensing exams do.

Importantly, MAI-DxO is model-agnostic orchestration. Microsoft tested it on top of models from several labs, and TechCrunch reports the best results came with OpenAI's GPT-5 and GPT-5.5.

What Do the Science Results Really Show?

The headline numbers are strong, and the fine print is useful. TechCrunch's analysis of the study data notes that much of the improvement comes from more capable underlying models, with the orchestration adding gains on top. Those gains were largest on complex cases and with weaker models: Gemini 2.5 Flash rose from 46.4% on its own to nearly 69% with DxO, and GPT-5.5 climbed from 72% to nearly 87% on complex cases at a slightly lower cost. On emergency department cases, the orchestration added little over GPT-5.5 alone, with both scoring about 85%.

Two caveats keep the result in perspective. The physicians in the comparison were primary care doctors and general internists working alone, while cases this complex are usually handled by specialists or teams. And this remains a research system tested in simulation, not a clinical product.

Why the Open SDBench Matters

Microsoft is open-sourcing SDBench for select NEJM cases, including the benchmark, its Gatekeeper and Judge components and the cost-estimation code. That lets outside researchers reproduce the evaluation, test other models and challenge the findings, which is exactly what healthcare AI needs to earn trust. Microsoft AI CEO Mustafa Suleyman also told TechCrunch the company is working with the Mayo Clinic on a healthcare foundation model planned for next year.

When Will Diagnostic AI Reach Real Hospitals?

The Beth Israel Deaconess study is the real test. In that prospective study, consenting patients will receive standard care while also interacting with the system, and researchers will compare its recommendations against physicians' independent assessments. It follows Google's own real-world work, covered in our look at the AMIE Lancet clinic study. Together, these studies mark a shift from benchmark scores toward careful clinical evaluation. Keep up with the research in our artificial intelligence coverage.

Sources: Microsoft AI — October 22, 2026; TechCrunch — October 22, 2026.