Microsoft researchers have developed an AI diagnostic tool that follows a doctor-like, step-by-step process, showing promising accuracy in complex medical cases while potentially reducing diagnostic costs.
Microsoft researchers are developing an artificial intelligence (AI) system designed to diagnose medical conditions through a structured, step-by-step process similar to the way doctors assess patients. According to a study published in Science, the tool demonstrated significantly higher diagnostic accuracy than general physicians in several complex medical scenarios.
Unlike conventional medical AI systems that analyse all available information at once, the new approach gradually gathers information, evaluates findings and adjusts its diagnostic plans as additional evidence becomes available.
The AI system uses an orchestration mechanism to organise the diagnostic process. Instead of immediately producing a diagnosis from a complete set of symptoms and test results, it determines which questions should be asked and what information needs to be collected first.
The system evaluates a patient’s medical history, identifies relevant findings from physical examinations and determines which investigations may be necessary. After reviewing the results, it can recommend further tests and revise its initial assumptions.
This iterative process allows the AI to develop and update possible explanations for a patient’s condition as new evidence emerges.
The system’s orchestrator coordinates a simulated panel of clinical experts to work through diagnostic problems. This structure is intended to organise the reasoning process, helping the AI assess different possibilities and determine the next appropriate steps.
By coordinating these activities, the technology aims to replicate important aspects of clinical decision-making rather than simply matching symptoms to a likely diagnosis.
In the reported evaluation, the AI system was tested on more than 300 unusual and challenging medical cases. Researchers found that it was more than twice as accurate as general physicians who used textbooks and internet resources to investigate the cases.
The model also performed better than human physicians when evaluated against scenarios derived from a major database of emergency department admissions.
These findings suggest that sequential diagnostic reasoning could help AI systems handle complicated medical problems that require careful investigation and the consideration of multiple possible conditions.
However, results from research evaluations do not necessarily establish how well the technology will perform in routine clinical practice. Further testing in real healthcare settings will be important.
Many existing medical AI evaluations rely on multiple-choice questions or structured clinical summaries. According to Microsoft’s researchers, AI systems have become highly capable of answering such questions, but these tests do not fully reflect the uncertainty doctors face when examining actual patients.
In real-world situations, patients may provide incomplete histories, have several overlapping symptoms or require additional examinations before a diagnosis becomes clear.
Microsoft’s approach attempts to address this challenge by allowing the system to gather information progressively and change its diagnostic strategy as the investigation develops.
The researchers also reported that diagnostic work-ups designed by the orchestrator cost less than those conducted by human doctors in the evaluated scenarios.
The system is intended to consider not only whether a diagnosis is correct but also how carefully and efficiently the necessary investigations are selected.
If further studies confirm these results, the approach could help healthcare providers explore ways to improve diagnostic efficiency and reduce unnecessary testing. Its real-world value will depend on clinical validation, safety and how effectively it integrates with existing medical workflows.
Microsoft is collaborating with Beth Israel Deaconess Medical Center in Boston to investigate the technology’s potential in emergency care.
The next stage of research will examine whether the AI tool can help triage patients in emergency department waiting rooms. Triage involves assessing patients’ conditions and prioritising care according to the urgency of their medical needs.
This research could provide important evidence about whether sequential AI reasoning can support healthcare professionals in time-sensitive environments. However, the system’s performance in this setting remains to be established through further evaluation.
Although the reported findings are promising, an AI diagnostic tool should not be treated as a replacement for qualified medical professionals. Clinical decisions require careful consideration of patient circumstances, medical history, examination findings and potential risks.
Researchers will need to assess the system’s reliability across different patient populations, its ability to manage incomplete or conflicting information, and the consequences of incorrect recommendations.
Human oversight, appropriate clinical safeguards and evidence from real-world trials will be essential before the technology can be considered for wider deployment.
Conclusion
Microsoft’s sequential AI diagnostic tool represents a potential development in medical artificial intelligence by focusing on how diagnoses are investigated, rather than simply predicting an answer from existing information. Its reported performance in complex cases and potential to reduce diagnostic costs make it a promising area of research. Further clinical testing will determine whether this approach can safely and effectively support doctors in everyday healthcare and emergency departments.