Microsoft presented new results for its MAI-DxO diagnostic system on October 22, 2026. On complex medical cases, the company announces 83% correct responses, compared with 36% for doctors with documentary resources. These scores come from a file exercise, not from a trial demonstrating improved patient care.
A medical investigation step by step
The device does not immediately receive all the elements of the folder. It starts with a short description, asks for information or reviews, and then revises its assumptions. Several virtual roles confront possible diagnoses and discuss the usefulness of investigations.
The work published on 22 October includes 304 complex cases from medical publications. Microsoft also reports 80% correct diagnoses on a set of cases from emergencies, compared to 47% for physicians participating in the comparison. These two sets do not measure exactly the same clinical difficulty.
Selected files, not all emergency patients
The public literature of the game MMIC-IV-Ext-SDBench describes 488 records of patients hospitalized after their transition to an emergency. Of these, 325 are used for published evaluations; The remaining 163 are reserved for development. A subset of 71 cases was reviewed by clinicians, who corrected 13 baseline diagnoses.
This selection includes situations with an identifiable complaint and a final diagnosis. It therefore does not represent all those arriving in the emergency room. Direct admissions to resuscitation are excluded.
Another limitation: exchanges are simulated in writing. Where the file does not contain a response, the system responsible for providing the information may produce a summary. This is not a measure actually observed in the patient.
Clinical outcome still to be demonstrated
Cost estimates are based on standardized US prices. They do not reproduce local rates, the availability of examinations or all their deadlines. The protocol evaluates the correctness of the diagnosis, but not the safety of the treatment or the subsequent evolution of the patient.
Microsoft announces a prospective study with the Beth Israel Deaconess Medical Center in Boston, where patients will interact directly with the system. Evaluations will be compared to independent physicians. The announcement describes a next step, without yet delivering the results of this use in real situations.
Comments