The research, published in the journal Science on Thursday, examined a model designed to gather information progressively rather than produce a diagnosis from a single assessment of a patient's symptoms, test results, and medical records. In tests involving more than 300 unusual and challenging medical cases, the system was more than twice as accurate as general physicians using textbooks and the internet to investigate the cases. It also outperformed doctors on scenarios drawn from a major emergency department admissions database.
The system uses an AI component described as an “orchestrator” to coordinate simulated clinical experts. It determines which questions should be asked about a patient's history, assesses relevant examination findings, selects further tests and interprets the results. It then updates its working hypotheses as new information becomes available.
This approach differs from conventional medical AI systems that analyse an entire set of symptoms and clinical information at once. By reproducing elements of the investigative process used by doctors, the researchers aim to evaluate not only the accuracy of a final diagnosis but also how carefully and efficiently the system reaches it.
The researchers also found that diagnostic work-ups designed by the orchestrator cost less than those conducted by human physicians in the tests. That finding could be relevant to healthcare providers seeking to manage workloads and resources, particularly in settings where clinicians face pressure to assess complex cases quickly.
However, the results do not establish that the tool is ready to replace doctors or can safely be deployed across routine clinical practice. The test cases and simulated assessments differ from the unpredictable conditions of real healthcare, where patients may provide incomplete information, have multiple conditions, or require urgent intervention. Clinical validation, patient safety, accountability, and integration with existing systems remain important requirements.
Microsoft's next planned step is to work with Beth Israel Deaconess Medical Center in Boston to test the tool's ability to triage patients in emergency department waiting rooms.
That work could provide evidence of how the system performs in a live clinical environment, where speed, accuracy and the prioritisation of urgent cases matter directly to patient outcomes.
The wider significance lies in the possibility of AI supporting clinical reasoning rather than simply interpreting data. Whether that promise translates into practical benefits will depend on further testing, transparent evaluation and demonstrated safety in real-world healthcare settings.
