Microsoft's artificial intelligence diagnoses with over twice the accuracy of doctors in complex medical cases

The new system MAI-DxO achieves 83.2% diagnostic accuracy in cases from The New England Journal of Medicine. For doctors with access to reference materials, the result is 36.1%. The study raises new questions about the future of medicine, the role of the physician, and the capabilities of artificial intelligence.

Artificial intelligence takes another serious step in a field where the accuracy of decisions can have a direct impact on human life. Researchers from Microsoft presented a system that not only analyzes symptoms and suggests a probable diagnosis but also goes through a sequential diagnostic process – asking questions, assessing what tests are needed, analyzing the results obtained, and adjusting its hypotheses based on new information.

The system is called Microsoft AI Diagnostic Orchestrator, or MAI-DxO. The results of the new study were published on October 22, 2026, in the scientific journal Science. According to the authors, the approach shows significantly higher accuracy than that of doctors in two sets of test medical cases.

This is an important difference from known demonstrations of medical artificial intelligence. Instead of receiving all the information about the patient at once and choosing an answer, the system must decide what to ask or investigate next. This process is a crucial part of real medical diagnosis.

What is MAI-DxO?

MAI-DxO is not just a standalone model with medical knowledge. It is a system for coordinating diagnostic steps that uses a language model and organizes the interaction between several virtual roles resembling medical specialists.

Instead of immediately committing to one diagnosis, the system can consider different possibilities, check whether the initial assumption is confirmed, and assess which next piece of information would be most useful.

The process involves several key steps: initial case assessment, formulation of possible diagnoses, selection of additional questions and tests, interpretation of results, and updating the diagnostic hypothesis.

In the main results presented in the publication, MAI-DxO works with GPT-5. However, the essential idea is that the diagnostic strategy is organized by a separate coordinating mechanism, rather than relying solely on a single response from the language model.

This matters because medical diagnosis is rarely the result of merely recognizing a symptom. It often requires careful history-taking, distinguishing between similar diseases, and selecting appropriate tests, without missing important possibilities or ordering unnecessary procedures.

On what basis was the study conducted?

Researchers developed the Sequential Diagnosis Benchmark (SDBench) – an evaluation system that checks not only the final diagnostic answer but also the way in which it is reached.

Two sets of medical cases were used.

The first includes 304 complex clinical cases published in The New England Journal of Medicine. These are cases selected for their diagnostic complexity and educational value. They do not represent a random sample of all patients visiting a doctor.

The second set includes 325 cases of patients admitted to emergency departments from the MIMIC-IV-ED database. It allows for checking how the system performs in more typical situations related to emergency medical care.

In the tests, the diagnostician starts with a limited description of the case. Additional information is provided when requested. Thus, the system must choose what questions to ask, what tests to order, and when it has enough data to formulate a diagnosis.

This makes the assessment closer to the sequential nature of clinical reasoning, although it continues to be conducted in a controlled environment with text descriptions, rather than in actual patient examinations.

Result #1: 83.2% vs. 36.1% in complex cases

In the 304 complex cases from The New England Journal of Medicine, MAI-DxO achieves a diagnostic accuracy of 83.2% when working with GPT-5 and without a budget constraint for tests.

Doctors participating in the comparison achieve an average accuracy of 36.1% when they have access to the internet and reference resources. Without such resources, their result is 22.3%.

The difference between the system and doctors with access to reference materials is 47.1 percentage points. In relative terms, the accuracy of the AI system is approximately 2.3 times higher in this test.

This is an impressive result, but the context is crucial. The cases are particularly complex, and the comparison is with individual general practitioners and hospital specialists, not with teams of experts discussing a specific patient in a real clinical environment.

Moreover, 83.2% is the result under a specific configuration of the system and a specific set of cases. It should not be interpreted as the overall probability of MAI-DxO making a correct diagnosis for any random patient.

Result #2: 80.3% in emergency department cases

In the second test, the system was evaluated on 325 cases from the MIMIC-IV-ED database.

MAI-DxO achieves an accuracy of 80.3%, while the average result of doctors is 46.8%. This is a difference of 33.5 percentage points, or approximately 1.7 times higher accuracy for the AI system.

This result is important because it tests the technology in a different type of medical situation. However, this also concerns retrospective cases presented through information in a testing environment, rather than testing the system on patients currently being treated in an emergency department.

The next step will be to check whether a similar result can be achieved in real medical practice, where information is often incomplete, time is limited, and decisions must also consider the individual condition of the patient.

Not only accuracy but also the cost of diagnosis

One of the interesting aspects of the study is that the evaluation is not limited to whether the system reaches the correct diagnosis. Researchers also consider the approximate cost of the diagnostic process.

In the complex cases from NEJM, the configuration with the highest accuracy reaches 84.5%, with an estimated average cost of $8,945 per case. In a configuration without a budget constraint, the accuracy is 83.2%, and the estimated cost is $5,431. With stricter cost constraints, the system maintains an accuracy of 80.9% at a cost of $2,241 per case.

In comparison, the average estimated cost of the diagnostic process for doctors with access to reference resources is $4,725 per case.

These amounts are not actual bills paid by patients or hospitals. They are estimates calculated within the study based on prices for medical services and tests. Therefore, they should not automatically be taken as a forecast for future savings in the healthcare system.

Nevertheless, the results indicate something significant: in diagnosis, it matters not only how much information a system can gather but also whether it selects the right information without unnecessarily increasing the number of tests.

What does this mean for the future of medicine?

Microsoft's research indicates a potential shift in how medical artificial intelligence is evaluated. Instead of measuring solely its ability to answer medical questions, its skills in gathering information, testing hypotheses, selecting tests, and assessing when to continue the diagnostic process are becoming increasingly important.

If such systems prove effective in real clinical conditions, they could assist doctors in complex cases, draw attention to missed opportunities, and contribute to more sensible test selection.

However, this does not mean that the doctor becomes redundant. Clinical work involves physical examination, patient observation, interpretation of imaging and laboratory data, risk assessment, conversation with the patient, and decision-making under uncertainty. It also requires responsibility for the consequences of treatment.

That is why the next stage of development is so important. According to Reuters, Microsoft plans to collaborate with Beth Israel Deaconess Medical Center in Boston to test the technology's capabilities for initial triage and patient sorting in emergency department waiting rooms.

This will be a more significant test than the comparison with archival medical cases: it will show whether the technology can support real clinical decisions under the actual constraints of healthcare.

What should we know before declaring AI a better diagnostician than a doctor?

There are several important limitations. First, the study was conducted in a controlled, text-based environment. Second, the complex cases from NEJM were selected precisely because they represent diagnostic challenges and do not reflect the frequency of various diseases in everyday practice. Third, the comparison with individual doctors does not prove that the AI system will outperform specialists or multidisciplinary teams when working with real patients.

It is also necessary to check how the system handles rare diseases, incomplete or contradictory data, possible errors in tests, and situations where the correct decision is to seek immediate human intervention.

Therefore, the most accurate conclusion from the study is not that artificial intelligence is already a better doctor. More importantly, an AI system shows impressive results in consistently solving complex diagnostic tasks, and this approach deserves verification in real medical conditions.

The future of medicine is likely not to be determined solely by who can make a more accurate diagnosis – the human or the machine. It will also depend on how doctors can use technologies without losing the human judgment, responsibility, and trust that underpin treatment.

The material is for informational purposes only and does not constitute medical advice. The results of the study are not a basis for self-diagnosis or changes in treatment.

Sources

1. Microsoft AI – Building Medical Superintelligence, October 22, 2026.

2. Reuters – Microsoft develops AI tool that can diagnose 'sequentially' like a doctor, October 23, 2026.

3. Medical Economics – AI system outperforms physicians on diagnostic accuracy in Microsoft study, October 22, 2026.

4. Scientific publication in Science – DOI: 10.1126/science.aeh7930.