Microsoft’s MAI-DxO Beats Doctors 83% to 36% in Medical Diagnosis Study

Microsoft has reported a major milestone in medical artificial intelligence: its AI diagnostic system, MAI-DxO, achieved 83.2% accuracy when solving complex medical cases, compared with 36.1% for generalist physicians using reference materials. The findings were published in Science on October 22, 2026, and suggest that AI systems designed to investigate cases step by step could improve diagnostic reasoning.

The system also outperformed physicians in a separate evaluation using emergency-department cases. However, the results come from controlled benchmarks based on existing medical records, not a clinical trial demonstrating better patient outcomes. Microsoft is now working with Beth Israel Deaconess Medical Center to evaluate the technology in a real hospital setting, an important next step toward determining whether these gains translate into safer and more effective care.

Key Takeaways

• 83.2% accuracy: MAI-DxO correctly diagnosed 83.2% of complex cases from The New England Journal of Medicine case proceedings.
• 36.1% for doctors: Generalist physicians using reference materials achieved 36.1% accuracy on the same complex-case benchmark.
• 80.3% in emergency cases: MAI-DxO also outperformed physicians on a separate emergency-department dataset, scoring 80.3% versus 46.8%.
• Step-by-step diagnosis: The system selects questions and tests, reviews new evidence and updates its diagnostic hypotheses.
• Potential cost savings: Microsoft’s research suggests the system can achieve high accuracy while using fewer or less costly diagnostic tests.
• Clinical testing still needed: The results do not establish that MAI-DxO is ready to diagnose or treat patients independently in routine healthcare.

What Is Microsoft’s MAI-DxO?

MAI-DxO stands for Microsoft AI Diagnostic Orchestrator. It is an AI system designed to coordinate the medical diagnostic process rather than simply generate a diagnosis from a fixed collection of symptoms.

In conventional AI-assisted diagnosis, a model may receive a patient’s symptoms, laboratory results and medical history all at once, then produce its best assessment. MAI-DxO follows a more interactive approach.

The system begins with limited information and determines what it needs to know next. It can request additional details about a patient’s history, examine findings, consider potential tests and revise its working diagnosis as new evidence becomes available.

Microsoft describes the approach as resembling a panel of virtual physicians. Different diagnostic perspectives are considered, and an orchestrator coordinates the process, weighs competing explanations and evaluates the usefulness of additional information.

The objective is not simply to identify the correct disease. It is to make the diagnostic process more systematic, including deciding which questions and tests are most valuable before reaching a conclusion.

How MAI-DxO Performed Against Doctors

Microsoft evaluated the system using SDBench, a benchmark built around two complementary datasets: complex clinical cases published in The New England Journal of Medicine and more typical emergency-department presentations drawn from the MIMIC-IV-ED database.

The published results show a substantial performance gap in both evaluations.

Diagnostic evaluation	MAI-DxO accuracy	Physician accuracy
Complex NEJM medical cases	83.2%	36.1% with references
Emergency-department cases	80.3%	46.8%

Source: Microsoft AI’s research announcement.

On the complex NEJM cases, the system was approximately 2.3 times as accurate as physicians who could consult reference materials. Compared with physicians working without those resources, whose accuracy was 22%, the difference was larger.

The emergency-department evaluation is important because it tests a different type of case. The NEJM dataset emphasizes difficult, unusual diagnostic problems, while the emergency dataset represents a broader range of clinical presentations.

Nevertheless, these comparisons should not be interpreted as proof that AI is universally better than doctors. Benchmark results depend on case selection, available information, evaluation methods and the conditions under which participants work.

Why the AI Uses a Step-by-Step Diagnostic Process

Medical diagnosis rarely involves identifying a disease from one symptom alone. Doctors gather information, consider several possible explanations and decide which investigation is likely to be most informative.

MAI-DxO attempts to reproduce that sequence using AI models and an orchestration layer.

Its process can be summarized in four stages:

1. Gather information: Start with the available case description and identify important missing details.
2. Choose questions or tests: Determine which additional evidence could help distinguish between competing diagnoses.
3. Update the assessment: Incorporate the new information and revise the likely explanations.
4. Reach a diagnosis: Select the best-supported conclusion after evaluating the available evidence.

This approach can help avoid a common problem with automated reasoning: committing too early to an explanation and then interpreting all subsequent information through that assumption.

It also creates an opportunity to evaluate the diagnostic process itself. A system that reaches the correct answer but orders many unnecessary tests may be less useful than one that reaches the same answer efficiently.

Can MAI-DxO Reduce Medical Testing Costs?

Microsoft’s research evaluates diagnostic accuracy alongside estimated costs associated with investigations and tests.

The reasoning is straightforward. Medical tests can provide valuable evidence, but they also involve financial costs, waiting times and, in some cases, discomfort or clinical risk. A diagnostic system that chooses investigations strategically could potentially reduce unnecessary testing.

Microsoft reported configurations that maintained high accuracy while using more cost-effective diagnostic workups. One cost-optimized configuration achieved 81% accuracy on the complex NEJM cases at less than half the estimated cost of comparative systems.

These figures are based on the study’s cost methodology, not a guarantee of equivalent savings in hospitals. Real-world expenses depend on local prices, staffing, equipment, follow-up care and how clinicians use the results.

The findings nevertheless suggest that the value of diagnostic AI should be assessed through more than accuracy alone. Cost, testing efficiency and the quality of the diagnostic process also matter.

How Microsoft Plans to Test MAI-DxO in Hospitals

Despite the benchmark results, MAI-DxO has not yet demonstrated that it can independently deliver better outcomes for patients in routine clinical practice.

Microsoft has announced a partnership with Beth Israel Deaconess Medical Center in Boston to evaluate the technology in an emergency-department setting. The planned work will help assess how the system performs when integrated into a real clinical environment.

Hospital testing introduces challenges that retrospective benchmarks cannot fully capture. Patients may have incomplete histories, unusual combinations of conditions or symptoms that change over time. Clinicians must also account for patient preferences, urgency, communication and practical limitations on testing.

Researchers will need to establish whether the system can make useful recommendations consistently, whether clinicians can interpret its outputs and whether its use improves care without introducing unacceptable risks.

Regulatory requirements, privacy safeguards and clear responsibility for clinical decisions will also be essential before broader deployment.

What MAI-DxO Could Mean for Patients and Doctors

If further testing confirms its benefits, diagnostic AI could help clinicians work through complicated cases, identify overlooked possibilities and prioritize investigations.

Patients in areas with limited access to specialists could potentially benefit from tools that help local healthcare teams consider a wider range of diagnostic explanations. Hospitals might also use AI to support case reviews, identify situations requiring specialist input or improve the consistency of diagnostic workups.

However, the technology should not be viewed as a replacement for medical professionals based on these results alone. Doctors do more than identify diseases: they examine patients, interpret context, communicate uncertainty, weigh treatment risks and take responsibility for clinical decisions.

The most useful near-term role may be decision support, in which AI helps clinicians analyze evidence while qualified professionals retain oversight and make treatment decisions.

The Bigger Picture

MAI-DxO illustrates how AI research is moving from answering medical questions to coordinating more complex reasoning workflows. Rather than relying on a single response, the system gathers evidence, revisits its hypotheses and considers the value of further investigations.

That development could become important across healthcare, particularly when cases are difficult and the next diagnostic step is not obvious. But impressive benchmark accuracy is only one part of the evidence needed to establish clinical usefulness.

The real test will be whether these systems can improve patient outcomes, operate reliably across diverse populations and support clinicians without creating new risks. Microsoft’s results provide a promising research signal, while prospective clinical evaluation will determine how much of that promise can be translated into practice.

FAQs

1. What is Microsoft’s MAI-DxO?

MAI-DxO is Microsoft’s AI Diagnostic Orchestrator, a system that coordinates language models to investigate medical cases step by step. It selects questions and potential tests, evaluates new evidence and works toward a diagnosis.

2. Did MAI-DxO really score 83% compared with 36% for doctors?

Yes. Microsoft reported 83.2% accuracy for MAI-DxO and 36.1% for generalist physicians using reference materials on a benchmark of complex NEJM cases. These were controlled evaluations, not a comparison of patient outcomes in ordinary clinical practice.

3. Does MAI-DxO outperform doctors in emergency cases too?

In Microsoft’s separate evaluation using MIMIC-IV-ED emergency-department cases, MAI-DxO achieved 80.3% accuracy, compared with 46.8% for physicians. The findings are encouraging but still require validation in real-world clinical settings.

4. Can MAI-DxO replace doctors?

The results do not establish that MAI-DxO can replace doctors. The system needs further clinical testing to determine its safety, reliability and effect on patient outcomes. Doctors remain essential for examination, treatment decisions, communication and clinical oversight.

5. When could patients use MAI-DxO?

Microsoft has announced plans to evaluate the technology with Beth Israel Deaconess Medical Center. Broad patient access will depend on the results of further testing, clinical validation, regulatory requirements and decisions about how the system should be used in healthcare.

Looking Ahead

The next major milestone for MAI-DxO is clinical evaluation. Researchers will need to determine whether the system’s strong performance on historical cases carries over to real patients, where information can be incomplete and decisions may have immediate consequences. Results from hospital testing should help clarify its appropriate role in emergency care and other diagnostic workflows.

For the healthcare industry, the broader opportunity is to use AI to support better-informed decisions without treating benchmark scores as a substitute for clinical evidence. If MAI-DxO can demonstrate reliable performance, meaningful cost benefits and improved patient outcomes, it could become a valuable tool for clinicians tackling complex diagnoses.