Microsoft publishes new results from MAI-DxO, its artificial intelligence system dedicated to medical diagnosis. Introduced in 2025, the tool is the subject of a study published October 22, 2026 in Science. The researchers expanded the assessment to emergency files in addition to the 304 complex cases used to evaluate reasoning skills.
From complex cases to emergency files
The first work was based on cases published in the New England Journal of Medicine, selected for their difficulty. This time, Microsoft adds MIMIC-IV-ED situations, a database of unidentified emergency files. The objective is to observe the behavior of the system when faced with another set of clinical problems.
Of the 304 complex cases, MAI-DxO now reaches approximately 83% correct diagnosesaccording to the results published by Microsoft. Doctors who participated in the evaluation scored 22% without external resources and 36% when they could consult reference materials, including medical books and the Internet. These conditions count: scores describe the performance of participants in a supervised exercise, not that of the profession as a whole.
In emergencies, the system obtains 80% correct answers, compared to 47% of the doctors evaluated, an approximate ratio of 1.7.
Behind these figures, MAI-DxO organizes the work of several specialized agents. Some offer diagnoses, others choose which tests to order or look for inconsistencies in hypotheses. The information arrives gradually: the system must decide what it wants to know before receiving the answer. This sequential approach seeks to reproduce part of the clinical reasoning, when it is necessary to choose the next useful examination with a file that is still incomplete.
Microsoft also indicates that it has repeated the evaluations with more recent models, including GPT-5.5, maintaining the same orchestration method. Therefore, performance also benefits from the evolution of the models that power the system.
Diagnostic accuracy and estimated average cost of examinations in MIMIC-IV-ED emergency cases. The chart compares MAI-DxO, medical, and AI models evaluated in October 2025.
An 81% setting to limit exams
The test bank also takes into account the estimated cost of the requested exams. A MAI-DxO configuration aimed at limiting these expenses achieves 81% correct diagnoses in complex cases, at less than half the cost of comparison systems. The point is to find the correct answer without unnecessarily multiplying the analyzes and images. However, these are costs assigned in the protocol: this result does not demonstrate that a hospital would reduce its bill in the same proportions.
In complex NEJM cases, different MAI-DxO configurations are compared with doctors and AI models based on their diagnostic accuracy and the estimated average cost of the exams. Points toward the top and left combine greater accuracy with lower cost. October 2025 evaluations.
The transition to patients is yet to be evaluated
Publication in science magazine subjects the method to scientific examination, but does not constitute authorization for clinical use. The tests are still simulations based on files already created. The testbed documentation even specifies that when a requested result is missing from the original case, the module responsible for providing the information can produce a plausible synthetic response. A real consultation adds discussions with the patient, time constraints, and the consequences of decisions made.
Microsoft announces a prospective study with Beth Israel Deaconess Medical Center in Boston. Patients must interact directly with the system, whose evaluations will be compared with those of doctors working independently. This step should allow us to examine their behavior in the new cases collected throughout the study.
To facilitate further work, Microsoft releases certain components of its SDBench Frameworkincluding tools that provide information during the test, evaluate the diagnosis and estimate the cost of the exams. Clinical data remains subject to access requirements: New England Journal of Medicine cases require a license, while MIMIC requires PhysioNet credentials, training, and acceptance of a user agreement.
