Microsoft is preparing a hospital study in Boston to assess how the technology performs alongside clinicians. #
Microsoft researchers have developed an artificial intelligence system that diagnosed 83% of complex medical cases in a controlled test, compared with 36% for general physicians using textbooks and online references.
The findings, published in the journal Science, come from a system designed to investigate medical problems step by step rather than receive all the information needed for a diagnosis at once.
Called Microsoft AI Diagnostic Orchestrator, or MAI-DxO, the system coordinates multiple AI agents that act as a team of virtual physicians, assessing possible conditions and deciding which information is needed before reaching a diagnosis.
Researchers evaluated the technology using 304 complicated cases published in The New England Journal of Medicine. Participating generalist physicians correctly diagnosed 22% without outside reference materials and 36% when they could consult them, according to Microsoft AI's announcement.
Traditional medical AI tests often use multiple-choice questions that provide all the relevant clinical information upfront. Microsoft's researchers instead developed the Sequential Diagnosis Benchmark, or SDBench, to assess how AI investigates an unfamiliar patient.
Starting with a short clinical description, MAI-DxO can request medical history, physical examination findings and laboratory or imaging tests. Its virtual specialists propose diagnoses, challenge one another's assessments and determine whether further investigation is necessary. The benchmark measures diagnostic accuracy, the information requested and the estimated cost of testing.
Researchers also evaluated the system using emergency department cases from MIMIC-IV-ED, a database containing deidentified hospital records. MAI-DxO achieved approximately 80% accuracy, compared with 47% for generalist physicians under the study conditions.
The results extended beyond diagnostic performance. Microsoft examined whether the system could reach accurate conclusions while limiting unnecessary medical investigations. Each requested test was assigned an estimated cost, allowing researchers to compare different approaches.
One configuration achieved 83.2% accuracy on complex cases at an estimated diagnostic cost of $5,431. When stronger cost constraints were introduced, accuracy remained at 80.9%, while expenses fell to $2,241.
The physicians' comparative workups averaged $4,725. These were modeled testing costs, not savings demonstrated in hospitals. Microsoft tested its orchestration method with language models from several developers, including OpenAI, Google, Anthropic, Meta and DeepSeek.
Some models performed well independently. For example, GPT-5 achieved approximately 79% accuracy on the complex cases without the orchestration system, compared with just over 83% when used with MAI-DxO.
Microsoft said subsequent evaluations involving newer models, including GPT-5.5, also showed improvements when the orchestration method was applied.
However, the research was conducted using historical, text-based cases rather than live clinical consultations. The physician comparison involved generalists working individually, not specialist teams.
Microsoft is now working with Beth Israel Deaconess Medical Center in Boston to test the system in clinical settings.
The planned prospective study will involve patients interacting directly with the technology while receiving conventional medical care. Researchers will compare its assessments with physicians' independent evaluations to identify where the system performs well and where it encounters difficulties.
© Copyright IBTimes 2026. All rights reserved.