IT Home reported on October 23 that on October 22 local time, Microsoft announced that when evaluating common visit cases in US emergency departments based on the MIMIC-IV-ED datasetcommon visit cases, its AI diagnostic system MAI-DxO achieved a diagnostic accuracy 1.7 times that of human physicians.
The system is named Microsoft AI Diagnostic Orchestrator, or DxO for short, and simulates the process of a multidisciplinary expert panel consultation by uniformly orchestrating multiple AI agents — the agents work together to interview the patient, order lab tests, and conduct in-depth analysis and debate over potential diagnoses.
In a simulated evaluation based on complex cases from the New England Journal of Medicine (NEJM)complex cases,MAI-DxO achieved a diagnostic accuracy as high as 83%. By comparison, primary care general practitioners and general internists with access to reference materials achieved only 36% accuracy.
Notably, such complex cases in real clinical practice are usually handled by specialist experts or multidisciplinary medical teams, and these specialists were not included in this controlled experiment.
Officially, the system costs only half or even one-third of what physicians cost, and can precisely pinpoint the correct diagnosis with fewer tests and clinical procedures.The research results have been formally published in the journal Science。
The study's own data also show that the leap in the system's diagnostic capability is largely attributable to improvements in the underlying foundation AI models, while Microsoft's orchestration architecture further amplifies these performance gains on top of that.
Microsoft tested MAI-DxO against models from multiple institutions including OpenAI, Google, Anthropic, Meta, and DeepSeek,with the best performance coming from OpenAI's GPT-5 and GPT-5.5。
IT Home attaches the paper link: