What changes compared with general-purpose models
Microsoft researchers have published in Science the results of an artificial intelligence system that formulates diagnoses sequentially, examining symptoms one step at a time instead of processing all the data in a single pass. The architecture is based on a central orchestrator that reads the initial medical history, evaluates the symptoms, and determines which questions to ask or which tests to request before reaching a conclusion.
The point is the progressive decision-making logic: according to the researchers, real medicine does not offer already-synthesized clinical pictures, and it is precisely the sequence of decisions that is decisive for reconstructing how a diagnosis is reached.
The numbers from the simulations
In simulations on complex cases drawn from the New England Journal of Medicine, the MAI-DxO model reached a diagnostic accuracy of 83%, versus 36% for a sample of human generalist physicians. In tests on typical emergency department cases, the system achieved 80%, versus an average of 47% for the physicians involved in the comparison, according to the source.
In terms of costs, the orchestrator-guided pathway proved to be between half and one-third of traditional assessments, thanks to a reduction in unnecessary tests ordered by the system. These are simulation data: they do not yet prove effectiveness in clinical practice.
Statements and checks still to be done
Mustafa Suleyman, head of Microsoft's AI division, described the discovery as a building block toward medical superintelligence, specifying that the goal is not to replace healthcare personnel but to provide physicians with tools capable of increasing precision and preventing errors. It is a company statement, not a clinical result.
The tool is not yet used in daily practice: the next step is a field collaboration with Boston's Beth Israel Deaconess Medical Center, to test the algorithm in the management and triage of emergency patients. The interaction with people and integration into staff working rhythms therefore remain to be verified; only afterward will it be possible to say whether the benefits observed in the simulations will transfer to the ward.
