Microsoft MAI-DxO, Microsoft’s AI agent system, achieved about 83% accuracy diagnosing complex cases from the New England Journal of Medicine, according to a paper published in Science. Generalist physicians with access to references scored 36%, and those without external resources scored 22%.
Led by Microsoft AI CEO Mustafa Suleyman, the Microsoft AI Diagnostic Orchestrator simulates a panel of specialists. Agents work sequentially: they ask questions, order tests, and reach a diagnosis, mimicking real clinical workflow. On emergency cases from Beth Israel Deaconess Medical Center, accuracy was around 80% versus 47% for physicians.
How Microsoft MAI-DxO works
Unlike single prompts that deliver the full case at once, Microsoft MAI-DxO operates iteratively. Agents with distinct roles—one challenges hypotheses, another controls costs—debate, update the differential diagnosis, and weigh the value of each test. The strongest results came with OpenAI models, especially GPT-5 and GPT-5.5.
Microsoft says that in some configurations the system reaches the correct diagnosis at lower cost than physicians. In others, unrestricted cost was higher, but budget-limited versions maintained superior accuracy. The study notes that much of the gain comes from the strength of the base model, with orchestration adding additional margin.
Limitations and next steps
Tests occurred in controlled, text-based scenarios, not with real-time patients. Microsoft plans to begin clinical trials at Beth Israel Deaconess soon. Suleyman described the result as a step toward “medical superintelligence,” but the paper itself and independent reporting highlight that performance still depends on the underlying model and does not replace human clinical judgment.
The SDBench benchmark assembling the cases will be open-sourced by the company. The full paper is in Science (volume 394).
Sources
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and sources available on the internet. Consult the original sources for full context.
Por GeekikiBot
