Microsoft's MAI-DxO AI: superiority in diagnosing complex cases
Microsoft's new artificial intelligence diagnostic tool, MAI-DxO, has demonstrated significantly higher accuracy in diagnosing complex clinical cases compared to a group of experienced physicians. In internal testing, MAI-DxO correctly diagnosed 85.5% of cases, whereas 21 practicing physicians from the United States and the United Kingdom were only able to manage 20% of the tasks.
The testing was conducted on 304 complex clinical cases previously published in the New England Journal of Medicine. MAI-DxO emulates a physician's workflow by analyzing symptoms and proposing necessary diagnostic procedures, while simultaneously optimizing costs by eliminating redundant tests.
The tool leverages a combination of leading language models, including GPT, Llama, Claude, and Gemini, with optimal results observed when utilizing OpenAI's o3 model. Microsoft emphasizes that the test focuses on "sequential diagnosis"—a critical skill requiring profound comprehension rather than mere memorization. It is important to note that, unlike real-world practice, the physicians in the study operated without the possibility of consultation.
MAI-DxO is not yet ready for clinical deployment. Further testing, clinical trials, and regulatory approval are pending. Microsoft envisions MAI-DxO as a supportive tool for physicians, automating routine tasks and assisting in diagnosis, rather than replacing medical professionals. The study's findings are detailed in a scientific paper that has not yet undergone peer review.
ORIENT
