REDMOND, WASHINGTON / RankWire.AI / – Microsoft researchers said an AI tool that diagnoses patients step by step, as doctors do, was more than twice as accurate as general physicians at solving complex medical cases. The findings appeared in a study published on Thursday, October 22, in the journal Science. The paper, titled “Sequential Diagnosis with Language Models,” describes the Microsoft AI Diagnostic Orchestrator, known as MAI-DxO. The system ran on OpenAI’s GPT-5 model in the study.
The tool identifies and coordinates the questions that should be asked of a patient and the exams and tests that should be ordered. It uses an orchestrator that organizes how the AI tackles a problem through a simulated panel of clinical experts. Existing medical AI models typically take an all-at-once approach. They examine all of a patient’s symptoms, lab results and imaging studies together to generate a diagnosis.
With the orchestrator coordinating each step, the model first decides what to ask about the patient’s history and what matters in the physical exam. Based on what it learns, it chooses which tests to order, reviews the results and orders further testing when needed. It updates its working hypotheses and plans throughout the process, the research leaders said in an interview. The approach mirrors how clinicians work through difficult cases.
AI doubles accuracy on difficult cases
The researchers tested the model on 304 highly unusual and confusing cases published in the New England Journal of Medicine. MAI-DxO reached the correct diagnosis in 83.2% of them. Generalist physicians who could use textbooks and the internet were correct in 36.1% of cases. The tool also outperformed doctors on 325 scenarios drawn from a major database of emergency room admissions. It scored 80.3% on those cases, compared with 46.8% for the physicians.
The researchers said assessments comparing older AI systems with human doctors usually rely on multiple-choice questions. Dominic King, corporate vice president of Microsoft’s AI health unit, said real patients do not arrive with neat case summaries. Harsha Nori, head of AI at Microsoft AI Health, said the team measured how judiciously and carefully the system completes a diagnostic work-up. That included how well it chose which tests to order.
Hospital trial planned as limits remain
The study has several limitations. The physicians were generalists rather than specialists, and the cases were presented as text. Neither the AI nor the doctors could perform physical exams or review medical imaging directly. Researchers said budget-limited versions of the system ran lower test costs, but those figures were estimates rather than actual medical bills. MAI-DxO is not yet used in clinical practice. Other tools, such as an AI-driven headset that detects epileptic seizures, are also moving toward patient care.
Microsoft plans to work with Beth Israel Deaconess Medical Center in Boston on the next phase. That study will test whether the tool can help triage patients in emergency department waiting rooms. Microsoft first described MAI-DxO in a research preprint in June 2025, which showed the orchestrator could work with models from Google, Anthropic and Meta. OpenAI has since released its GPT-5.6 family of models. Further evidence will be needed to show whether the tool improves patient outcomes.