Microsoft’s new AI tool diagnosed complex medical cases with more than double the accuracy of general physicians, according to a study published this week in Science. The system works by mimicking the step-by-step problem-solving process doctors use, rather than relying on a single, all-encompassing assessment.
The model uses an “orchestrator” that coordinates a simulated panel of clinical experts. It decides what questions to ask patients, which physical exams to prioritise, and what further tests to order based on the information gathered. This iterative process continues until the AI narrows down the diagnosis.
Existing medical AI systems typically analyse symptoms, lab results, and imaging all at once to generate a diagnosis. Microsoft’s approach differs by updating its working hypotheses continuously as it collects new data. This simulates how a doctor might learn more about a patient during an appointment.
The researchers put the tool to the test against more than 300 difficult and unusual medical cases. The model was over twice as accurate as general physicians who relied on textbooks and internet resources to diagnose the same cases. It also outperformed doctors on cases drawn from a major database of emergency room admissions.
“AI has become very good at getting a 100% pass rate on those [multiple-choice diagnostic tests],” said Dr Dominic King, corporate vice president of Microsoft AI Health, whose team developed the technology. “But of course, that’s not real-world. No patient ever came to see me with a nice summary.”
The study authors pointed out that typical AI diagnostic evaluations use multiple-choice questions, a format that doesn’t reflect real clinical complexity. Microsoft’s tool instead works through the diagnostic process in a way that more closely resembles actual medical practice.
The orchestrator also designs diagnostic work-ups that cost less than those devised by human doctors, the researchers said. “We’re not just measuring the accuracy of the systems. We’re also measuring how judiciously and carefully it goes through a diagnostic work-up,” said Harsha Nori, head of AI at Microsoft AI Health.
The next step for the team is to test the AI’s ability to triage patients in emergency department waiting rooms, in collaboration with Boston’s Beth Israel Deaconess Medical Center. This trial will assess how well the system prioritises urgent cases in a live clinical setting.
Microsoft’s research suggests the AI’s iterative, consultative process could address limitations of earlier tools that generate diagnoses from static snapshots of patient information. The system’s stepwise questioning and testing mirrors how doctors gather evidence in complex cases.
The findings come amid growing interest in AI applications in healthcare, where diagnostic errors remain a common source of patient harm. Whether this technology can be deployed safely and effectively in real hospitals will depend on ongoing clinical trials and regulatory review.
For now, the orchestrator offers a model of how AI might assist clinicians by coordinating multiple lines of inquiry, rather than providing a single, fixed answer. This approach reflects a move toward more interactive and transparent AI decision-making in medicine.
According to Joy Online.
