Microsoft researchers have announced that an artificial intelligence tool which diagnoses diseases step-by-step, like a doctor, was more than twice as accurate as primary care physicians in resolving complex medical cases. The study was published on Thursday in the journal Science.
The model determines which questions to ask the patient and which examinations and tests to conduct. It is assisted by a coordination system that organises the approach to the problem, using a simulated team of clinical experts who collaborate on diagnostic cases.
Existing artificial intelligence models in medicine generally analyse all symptoms, laboratory results and images simultaneously before making a diagnosis. The new system directs this process: the model first decides which questions to ask regarding medical history and which physical examinations are important. Based on the information obtained, it determines tests, considers results and orders further investigations if necessary. Throughout the process, it updates working hypotheses and plans.
When the tool was given more than 300 very unusual and complex medical cases, the model with the coordination system was more than twice as accurate in assessing patient conditions and determining diagnoses than primary care physicians, who used textbooks and the internet. It also outperformed them in cases with templates from a large patient database obtained from Emergency Departments.
Researchers said that estimates of older systems of this type are usually based on multiple-choice questions, where results are compared with those of doctors. Dr Dominik King is corporate vice president of the health unit in Microsoft’s AI division, whose team is developing the new technology. Harsha Nori is the head of the healthcare AI division at Microsoft AI. Researchers found that the diagnostic process is cheaper when the coordination system plans examinations and analyses than when doctors do so.
In collaboration with Boston’s Beth Israel Deaconess Medical Center, scientists will in the future test the tool’s ability to assess the urgency of patient conditions waiting for examination in Emergency Departments.
