Microsoft (MSFT.US) AI diagnostic "Orchestrator" featured in Science: Accuracy is more than twice that of general practitioners, marking a step toward clinical productivity for medical AI

According to Zhitong Finance APP, Microsoft (MSFT.US) has made a significant move in the medical AI field. On October 22, local time, Microsoft's AI diagnostic system MAI-DxO (Microsoft AI Diagnostic Orchestrator) research results were officially published in the journal Science. This is an AI tool capable of diagnosing patients step by step like doctors do, and its accuracy in solving complex medical cases is more than twice that of general internists.

In a simulated evaluation based on 304 complex cases from the New England Journal of Medicine (NEJM), MAI-DxO achieved a diagnostic accuracy rate of 83.2%, while general practitioners and internists with access to reference materials achieved only 36.1%; when references were not allowed, doctors' accuracy further dropped to about 22%. In another test based on MIMIC-IV-ED emergency admission data, the system's accuracy was 80.3%, general practitioners' was 46.8%, with the advantage narrowing to 1.7 times.

The report states that the model uses an "orchestrator" to identify and coordinate which questions to ask the patient and which examinations and tests to arrange; this orchestrator organizes the AI’s problem-solving approach by simulating a collaborative clinical expert team tackling diagnostic cases.

Current medical AI models typically use a "one-off" approach, meaning they analyze all patient symptoms, lab results, and imaging studies at once to produce a diagnosis.

With the orchestrator coordinating diagnostic steps, the AI model first decides which medical history information to inquire about and which components of the physical examination are important. Based on the information obtained, it determines which tests to arrange, evaluates the test results, and orders further examinations when appropriate. The study lead said that throughout the process, the model continuously updates its working hypotheses and plans.

When diagnosing over 300 highly rare and perplexing medical challenges, the model with the diagnostic orchestrator had an accuracy rate in "reviewing and evaluating" patients and confirming the diagnosis that was more than twice that of ordinary internists relying on textbooks and the internet. In scenarios based on a large emergency admission case database, this tool also outperformed human doctors.

Researchers stated that earlier systems of this type have typically been evaluated using multiple-choice questions in comparisons with human doctors.

“AI is now extremely proficient at this, achieving a pass rate of up to 100%,” said Dr. Dominic King, Corporate Vice President at Microsoft AI Health, who is developing this new technology. “But that is certainly not the real world. No patient ever comes to me carrying a perfectly summarized case.”

Researchers also found that when the orchestrator designed the testing process, the cost of diagnosis was lower than when human doctors designed the process. “We’re not only measuring the system’s accuracy, but also how prudent and meticulous it is in conducting diagnostic testing,” said Harsa Nouri, Head of AI at Microsoft AI Health.

In collaboration with Beth Israel Deaconess Medical Center in Boston, researchers will next test the tool's ability to triage patients in emergency department waiting areas.

In terms of behavioral traits, before ordering the first test, doctors on average asked only 5.6 questions, while DxO asked 25.5 questions on average, reflecting a more systematic and structured approach to taking medical histories. Regarding the underlying model, the system performed best on OpenAI GPT-5, with performance further improving upon switching to GPT-5.5, indicating that the benefits of the orchestration layer are transferable across models.

From a commercialization perspective, the more noteworthy signal from this study is the shift in evaluation standards: Microsoft simultaneously introduced SDBench (Sequential Diagnostic Benchmark), which transforms cases into step-by-step interactive "consultation scenarios," and for the first time, includes examination costs in addition to accuracy. This indicates that the competitive dimension of medical AI is shifting from single-parameter algorithm accuracy to a comprehensive comparison of “clinical reasoning orchestration capability + cost efficiency.”

This shift aligns with the industry’s overall pace. According to data from Grand View Research, the global medical AI market is projected to reach $42.62 billion in 2025, $56.2 billion in 2026, and is expected to expand to $1.05 trillion by 2036, with a compound annual growth rate of about 34% from 2026 to 2036.

On the supply side, since acquiring Nuance for nearly $20 billion in 2022, Microsoft has embedded Dragon Copilot into clinical documentation and voice workflows, covering a large number of hospitals and doctors in the United States. MAI-DxO forms a natural extension from “document automation” to “diagnostic decision support.”

The next key verification milestone is emerging. Microsoft plans to collaborate with Beth Israel Deaconess Medical Center in Boston in the coming months to launch prospective tests in real clinical environments, focusing on evaluating performance in emergency triage scenarios. At the same time, coupled with its partnership with the Mayo Clinic on foundational medical models, Microsoft’s complete healthcare commercialization puzzle is closing.

For the secondary market, short-term sentiment has already been catalyzed under the endorsement of Science, but the valuation center for medical AI will ultimately depend on whether real-world blind studies can reproduce the advantages observed in simulated environments, as well as how payers and regulatory frameworks define the pricing and responsibilities of “AI-assisted diagnosis.”