Microsoft’s Medical AI Takes Aim at the Hardest Diagnoses

Microsoft’s newest medical AI has delivered a striking result: in a controlled simulation, it diagnosed complex cases more than twice as often as generalist physicians. The system did not receive every symptom at once. It had to question the patient, choose tests, weigh costs, and build a diagnosis step by step.

That system is the Microsoft AI Diagnostic Orchestrator, or MAI-DxO, known as DxO for short. A peer-reviewed study published in Science tested DxO on hundreds of difficult medical cases, and Microsoft says the results point toward a new kind of clinical decision support. The bigger question now is whether performance in a benchmark can survive the messy pressure of real healthcare.

A Panel of AI Agents Works Through Each Case

DxO coordinates several AI agents that act like a panel of doctors. The system begins with a patient’s initial complaint, forms a working diagnosis, asks follow-up questions, decides which laboratory tests to order, and stops when it has enough evidence to reach a conclusion.

The system uses five AI agents based on OpenAI systems. Each agent handles a different part of the process: generating hypotheses, selecting tests, reviewing results, controlling costs, and checking the final diagnosis. Together, they create an orchestrated workflow instead of relying on one model and one prompt.

This approach addresses a major weakness in earlier AI medical evaluations. Previous studies often fed a complete case into a model in a single prompt, giving it all the facts at once. DxO instead had to gather information sequentially, a process closer to the way doctors investigate a case during a consultation.

The researchers tested DxO on 304 clinicopathological conference cases from The New England Journal of Medicine and 325 emergency department-to-admission encounters from the MIMIC-IV-ED database. Microsoft compiled the cases into its own benchmark, SDBench, which it plans to open-source.

The Accuracy Numbers Are Powerful — With Important Limits

On the complex cases from The New England Journal of Medicine, DxO achieved 83.2 percent diagnostic accuracy. Generalist physicians examining the same cases with reference materials reached 36.1 percent, while physicians working without reference materials reached 22.3 percent.

DxO also scored 80.3 percent on the emergency department cases, compared with 46.8 percent for human physicians. Those numbers explain why Microsoft describes the system as more than an exam-answering tool: it is designed to reason through uncertainty, gather evidence, and balance the cost of investigation.

Cases this complex are usually handled by specialists or physician teams, and those specialists were not part of the comparison. The physician group consisted of generalist doctors, including primary care doctors and general internists, who worked individually with reference materials.

Cost formed a second part of the evaluation. DxO reached its diagnoses at half or even one third of the estimated cost of human doctors, Microsoft said. With a wide budget, the system reached 83.2 percent accuracy at an average cost of 5,431 dollars per case; when spending fell to an average of 1,303 dollars, accuracy dropped to 79.6 percent.

The results also changed depending on the underlying AI model. On complex cases, DxO lifted GPT-5.5 from 72 percent accuracy to nearly 87 percent at a slightly cheaper cost. Gemini 2.5 Flash rose from 46.4 percent on its own to nearly 69 percent with DxO.

GPT-5 scored 78.9 percent on complex cases at about $8,400 per case, while DxO cut the cost by about $3,000 and raised accuracy by about four points. On emergency department cases, the gains were much smaller. GPT-5.5 scored about 85 percent with or without DxO at a similar cost, showing that orchestration helps most on complex cases and with less capable models.

Clinical Validation Is the Next Test

The study’s controlled setting leaves a major gap between impressive benchmark results and medical use. DxO interacted with a “gatekeeper agent” that mediated requests from an underlying case record, so the patient in the simulation was played by an AI. Human doctors had also done the work of creating the detailed case files that DxO analyzed.

That limitation matters because real care involves more than identifying a diagnosis. Microsoft wants to move beyond individual tasks and connect AI with clinicians and healthcare systems across the full arc of care. The company says that path will require clinical validation, regulatory compliance, and a responsible rollout.

Through a partnership with Beth Israel Deaconess Medical Center, Microsoft plans to evaluate DxO across different hospital settings. Dominic King, vice president of Health at Microsoft AI, said the testing will “get underway in the next couple of months.” Microsoft is also working with the Mayo Clinic to train a healthcare foundation model, with release planned for next year.

Mustafa Suleyman, CEO of Microsoft AI’s AI lab, called the result a step toward “medical superintelligence.” He said the age of AI improving lives through healthcare “is finally here,” and predicted, “Next year we’re going to see extremely impressive deployments of intelligent systems in production.”

Suleyman also said healthcare models could become “an order of magnitude better in capability within the next 24 to 36 months.” He believes that “in 2027, we will be on the verge of a coding style breakthrough for medical records,” with highly accurate predictions about how a medical record might unfold.

Microsoft’s ambition is clear, but so is the challenge. DxO must prove that its accuracy, cost control, and multi-agent reasoning can support real clinicians without turning cost optimization into a new source of harm. As Suleyman put it, “It needs clinical validation.” The next stage will show whether this powerful simulation marks the start of a healthcare transformation or only the opening move.

Stay connected via Google News

Follow us for the latest AI News