Microsoft’s AI Is Better Than Doctors at Diagnosing Disease

Medicine can be a combination of art and science, but Microsoft has just shown that a large part of the two can be learned – by a bot.

The company reports in a study published on the preparation site arxiv That its AI-based medical program, the Microsoft AI diagnostic orchestrator (May-Dxo), properly diagnosed 85% of the cases described in the New England Journal of Medicine. This is four times higher than the precision rate of human doctors, which found the right diagnoses of around 20% of the time.

The cases are part of the weekly series of the journal designed to extract the doctors: complicated and difficult scenarios where the diagnosis is not obvious. Microsoft took approximately 300 of these cases and compared the performance of its May-DXO to that of 21 general practices in the United States and in the United Kingdom in order to imitate the iterative way of doctors generally addresses these cases-by collecting information, by analyzing it, by ordering the tests, then making decisions according to these results. This allowed the doctors and the AI ​​system to ask questions and make decisions concerning the next steps, such as the order tests, on the basis of the information they have learned at each step – similar to a flow graph for decision -making, with questions and subsequent actions based on information gleaned by the previous ones.

The 21 doctors were compared to a group group of standard AI models which included Claude, Deepseek, Gemini, GPT, Grok and Llama. To further reflect the way in which human doctors tackle such difficult cases, the Microsoft team has also built an orchestrator: a virtual emulation of the series of colleagues and consultations that doctors often look for in complex cases.

In the real world, ordering medical tests costs money, so Microsoft has followed the tests that the AI ​​system and human doctors have ordered what method could do it more.

Not only did Mai-dxo outdoor doctors to land on the correct diagnosis, but the IA bot was able to do so at a lower cost of 20% on average.

Learn more:: If the Thimérosal is sure, why is it removed from vaccines?

“The increase in quadruple accuracy was more than that of previous studies,” explains Dr. Eric Topol, president of Translational Medicine and director and founder of the Research Translational Institute, which provided information on the project. “Most of the time, there is an absolute percentage difference of 10%, so it’s a very big leap.” But what really caught his attention was the cost. “Not only was AI more precise, but it was much cheaper,” he said.

Mai-dxo is still in development and is not yet available for use outside of research. But the integration of such a model in medicine could potentially lead to reductions in medical errors, which explains a significant part of health care costs and increase the efficiency of human doctors – which could in turn lead to better results for patients.

“This is a surprising result,” explains Mustafa Suleyman, CEO of Microsoft AI. “I think it gives us a clear line of view to put the best expert diagnoses available to everyone in the world at an incredibly affordable price.”

Ten years ago, when AI algorithms were introduced for the first time in medicine, they focused on binary tasks, known as Suleyman, such as digitization of images to detect tumors. “Today, these models have common conversations to a very high quality, asking the right questions and survey in the right way, suggesting the right tests and interventions at the right time,” he said.

Another advantage that an AI system can have is that it is free from many biases inherent in human experience. “We all have a confirmation bias,” said Dr. Dominic King, vice-president of Microsoft AI. “Sometimes clinicians will see something and think,” I’m sure it’s like the patient I saw last week. But AI thinks slightly differently. »»

Learn more:: The surprising reason why rural hospitals close

Mai-dxo not only spits an answer. He shows his work, so that doctors can potentially study and examine his reasoning process. “It is available for real -time surveillance by the human clinician,” explains Suleyman. “It is a level of transparency and visibility in the reflection process that we have not seen before.” This, in turn, could improve the education and training that doctors receive to further increase the accuracy of the diagnosis and ultimately the results for patients.

However, some experts in the field of AI and medicine note that Microsoft’s approach is not entirely new, because its diagnoses depended on the combined performance of several AI models. “In my mind, they do not test any individual model optimized for health care,” explains Keith Dreyer, Director of Data Sciences at Massachusetts General Hospital and Brigham and Women’s Hospital Center for Clinical Data Science. “They test the concept of testing all the models today and combining their decision -making. This part for me is not surprising.”

Dreyer also underlines that the results do not necessarily bring these systems of approval by regulatory agencies like the Food and Drug Administration of the United States, which has still not weighed whether these systems are medical devices.

Learn more:: What obtaining 105 blood tests from a health startup taught me

Microsoft is not the only company to follow an AI -based medical program to diagnose the disease. Google is developing a system based on conversation to imitate the back and forth-and more and more and back and pass, imitating the reasoning of human doctors in collecting information from patients and interpreting these symptoms to land on a diagnosis. In the first tests, the system surpassed doctors to accurately diagnose the case studies of simulated patients. In a 2024 test similar to that which Microsoft has performed using case studies, the previous version of the Google system diagnosed 59% of cases with precision, compared to the rate of human doctors of 33%.