Microsoft's MAI-DxO diagnostic AI published in Science

Microsoft AI announced that its research on MAI-DxO, an AI system that works through diagnoses step by step, has been published in Science. The paper builds on Microsoft's 2025 preprint and introduces two pieces: the Sequential Diagnosis Benchmark (SDBench) and the MAI Diagnostic Orchestrator (MAI-DxO).

SDBench turns cases into interactive exercises. Instead of seeing everything at once, as on licensing exams, a model or physician starts with limited information and chooses which questions to ask and which tests to order. The benchmark scores the final diagnosis, the evidence requested and the estimated cost of tests. It draws on 304 complex cases from The New England Journal of Medicine (NEJM) and on more typical emergency-department presentations from the MIMIC-IV-ED dataset.

MAI-DxO acts like a panel of virtual physicians. Several model instances propose diagnoses, pick tests and challenge each other over multiple rounds, weighing each test's value against its cost. Microsoft reports:

• NEJM cases: 83% accuracy, against 22% for physicians working without reference resources and 36% for physicians with them.
• Emergency-department cases: 80% accuracy, against 47% for physicians.
• Cost-optimised setup: 81% accuracy on NEJM cases at less than half the cost of comparable systems.
• Newer models: the gains held, and improved, when the same orchestration method was rerun on newer models, including GPT-5.5.

Microsoft is open-sourcing SDBench for selected NEJM cases. The release includes its Gatekeeper, Judge and cost-estimation code.

All of these results come from benchmarks. Microsoft says a prospective study with Beth Israel Deaconess Medical Center will have patients use the system directly, with its performance compared against physicians' independent assessments. No timeline or results have been given. The company also cited a Mayo Clinic collaboration and a companion paper in BMJ Digital Health and AI.