Category: AI in Healthcare · Medical Diagnosis · Digital Health

Artificial intelligence is moving beyond simply identifying diseases from medical images or analyzing laboratory reports. Researchers are now exploring AI systems that approach diagnosis as a step-by-step process: gathering patient information, considering possible conditions, evaluating examination findings, and deciding which tests could help narrow the possibilities.

This approach could make AI more useful in clinical settings because real medical diagnosis rarely depends on a single symptom or test. Doctors often need to ask additional questions, assess competing explanations, and revise their initial judgments as new evidence becomes available.

The central question is whether AI can reproduce parts of this clinical reasoning process while maintaining patient safety, avoiding unnecessary tests, and helping doctors make better decisions.

What Is Microsoft’s AI Diagnostic Orchestrator?

The system is called the Microsoft AI Diagnostic Orchestrator, or MAI-DxO. Its underlying idea is to coordinate several AI-based clinical reasoning roles so that the system can investigate a medical case progressively rather than producing an immediate answer from a fixed set of symptoms.

Microsoft researchers described this work in a study published in Science on October 22, 2026. They also developed the Sequential Diagnosis Benchmark, known as SDBench, to evaluate how AI handles the process of diagnosis, including the questions it asks, the tests it requests, and the accuracy of its final conclusions.

Unlike a system that simply receives a list of symptoms and predicts a disease, MAI-DxO is designed to decide what information it needs next. It can request more details about a patient’s history, evaluate examination findings, order simulated investigations, and reconsider its working diagnosis when new evidence appears.

This is important because many difficult medical conditions share similar symptoms. Fatigue, for example, can arise from anemia, thyroid disorders, infections, nutritional deficiencies, or other conditions. An effective diagnostic process must distinguish between these possibilities instead of automatically choosing the most familiar explanation.

How Does Step-by-Step AI Diagnosis Work?

The system follows a sequence of decisions similar to parts of a doctor’s diagnostic workflow.

Patient information

The system begins with a short description of the case and identifies what information is missing.

Targeted questions

It requests relevant details about symptoms, medical history, their duration, and other clinical findings.

Selective testing

It chooses investigations that may distinguish between competing diagnoses rather than treating every possible test as equally useful.

Reassessment

The system interprets new results, revises its hypotheses, and decides whether further investigation is justified.

Diagnostic conclusion

It provides its best-supported diagnosis after considering the available evidence and the cost of additional investigations.

Simplified explanation of the research approach; not a validated clinical protocol for treating patients.

The defining feature is that each decision affects the next one. If the first test provides a convincing explanation, the system may not need to pursue an extensive investigation. If the results are unexpected, it can explore alternative explanations.

The researchers’ approach also uses a simulated panel of clinical experts coordinated by an orchestrator. This design is intended to encourage consideration of multiple possibilities instead of relying on a single initial prediction.

What Did the Research Discover?

The reported results attracted attention because the AI was evaluated on more realistic diagnostic tasks than conventional medical multiple-choice questions.

According to Microsoft’s published findings, the researchers evaluated MAI-DxO on two types of cases: challenging clinical case reports and more typical emergency-department presentations. The results were promising, although they do not yet establish that the system is superior to doctors in routine clinical practice.

REPORTED RESULTS FROM THE STUDY

83%

Complex clinical cases

MAI-DxO diagnostic accuracy

36%

Complex clinical cases

Generalist physicians using reference tools

80%

Emergency-department cases

MAI-DxO diagnostic accuracy

47%

Emergency-department cases

Physician comparison group

Rounded figures from Microsoft’s reported evaluation; the comparisons concern specific test datasets, not all patients or all doctors.

The complex-case evaluation included 304 cases drawn from clinical proceedings published in The New England Journal of Medicine. The researchers also tested the system on 325 emergency-department cases from the MIMIC-IV-ED database.

These results suggest that sequential reasoning can help AI manage difficult diagnostic problems. However, the cases were evaluated in a controlled environment. Real patients may provide incomplete histories, have several conditions simultaneously, or develop new symptoms while being assessed.

It is also important to distinguish diagnostic accuracy from overall quality of care. A correct diagnosis is essential, but clinicians must additionally recognize emergencies, communicate uncertainty, choose appropriate treatment, and respond to a patient’s individual circumstances.

Can AI Improve Diagnostic Accuracy Without Ordering Unnecessary Tests?

This is one of the most important questions raised by the research. Medical testing can save lives, but investigations also have financial costs, practical limitations, and potential risks. A CT scan may expose a patient to radiation, for example, while invasive procedures can cause complications.

A diagnostic AI system must therefore do more than identify diseases correctly. It must also determine which investigations are necessary, when additional evidence is worth obtaining, and when the available information is sufficient.

The problem of unnecessary medical tests

Doctors sometimes order additional investigations because symptoms are ambiguous or because they need to rule out dangerous conditions. This caution can be appropriate, but unnecessary testing can lead to higher costs, incidental findings, anxiety, and follow-up procedures that may not benefit the patient.

Sequential diagnostic AI offers a possible way to address this challenge. Instead of recommending a large collection of tests at the beginning, it can assess the value of each next investigation in the context of what is already known.

Microsoft’s research explicitly evaluated estimated diagnostic costs alongside accuracy. The findings suggest that carefully coordinated AI reasoning can achieve strong diagnostic performance while using less costly diagnostic work-ups in the tested scenarios.

However, a lower estimated testing cost does not automatically mean better or safer care. An AI system might reduce spending by appropriately avoiding a low-value test, or it might make a dangerous mistake by failing to recommend an essential investigation. Clinical studies must distinguish between these outcomes.

How AI could select more useful tests

Imagine a patient visits a clinic with persistent fatigue and weakness. Several conditions could explain these symptoms, and immediately ordering every conceivable investigation would be inefficient.

An AI-assisted diagnostic process could help the clinician identify the most relevant missing information, select appropriate initial tests, and reconsider the possibilities once the results arrive.

For instance, an initial investigation might suggest a nutritional deficiency. The clinician could then determine whether that explanation fits the patient’s symptoms and history or whether further investigation is needed.

This is an illustrative example, not a claim that MAI-DxO has been validated to manage this particular patient scenario. In actual practice, test selection must depend on clinical assessment, the patient’s risk factors, and applicable medical guidelines.

Potential Benefits for Hospitals and Patients

Earlier recognition of difficult conditions

AI could help clinicians consider less obvious diagnoses when a patient’s symptoms do not fit a common pattern. This may be especially useful for complicated cases that require extensive reasoning.

More efficient diagnostic testing

By prioritizing investigations with greater potential value, AI could help reduce avoidable expenditure and make better use of limited laboratory and imaging resources.

Support for busy medical teams

An AI assistant could organize relevant clinical information and suggest alternative diagnoses, helping doctors review complicated cases more systematically.

Potential support in underserved areas

If validated and made accessible, diagnostic assistance could help healthcare professionals in facilities with limited specialist support. Its effectiveness would depend on local infrastructure, language coverage, and clinical oversight.

These are potential applications rather than established benefits of routine deployment. Research must establish whether the system improves patient outcomes and workflow in actual healthcare settings.

What Are the Risks and Limitations?

Despite the promising results, several barriers must be addressed before this technology can be trusted with real patients.

Incomplete or misleading information: AI systems depend on the information they receive. Missing symptoms, inaccurate medical histories, or poorly recorded test results may lead to incorrect conclusions.

Bias in medical data: If training and evaluation datasets underrepresent particular populations, performance may vary by age, sex, ethnicity, socioeconomic circumstances, or the healthcare setting. Independent evaluation across diverse populations is essential.

Delayed treatment: Sequential reasoning involves gathering information over multiple steps. If a patient has a life-threatening emergency, prolonged investigation must never delay urgent treatment.

False confidence: A fluent AI explanation can sound convincing even when the conclusion is wrong. Doctors need access to relevant evidence, uncertainty estimates, and the ability to challenge the system’s recommendations.

Limited real-world evidence: The reported results come from benchmark evaluations rather than a demonstration that routine use improves patient outcomes. Microsoft has announced plans to work with Beth Israel Deaconess Medical Center in Boston to evaluate emergency-department triage applications.

Another important limitation is that the comparison was not against every possible form of medical expertise. Performance against individual generalist physicians on selected cases does not establish superiority over specialist teams or experienced clinicians working with the full range of real-world resources.

What Happens Next?

The next phase of research should move beyond retrospective case simulations and evaluate the system in real clinical workflows.

Hospitals need to establish whether AI recommendations remain accurate when information is incomplete, patients have multiple health conditions, and clinicians face time pressure. They must also measure how often the system recommends unnecessary tests, misses dangerous conditions, or creates delays.

Researchers should assess outcomes across several dimensions:

• Diagnostic safety: How frequently does the system miss a serious disease or suggest an incorrect diagnosis?
• Testing efficiency: Does it reduce low-value investigations without missing essential ones?
• Time to treatment: Does AI assistance help patients receive appropriate care sooner?
• Cost-effectiveness: Do any savings remain after implementation, supervision, and additional testing are considered?
• Fairness: Does performance remain reliable across different patient populations and healthcare facilities?

These questions will determine whether sequential diagnostic AI becomes a practical clinical tool rather than simply an impressive research result.

Conclusion

Microsoft’s MAI-DxO research represents an important development in AI-assisted medical diagnosis. By gathering information progressively, selecting investigations, and revising its hypotheses, the system attempts to reflect the dynamic nature of clinical reasoning more closely than conventional approaches that generate an answer from a fixed set of information.

The reported accuracy and cost findings are encouraging, particularly for complex diagnostic cases. Nevertheless, they do not establish that the technology can safely replace doctors or improve outcomes in everyday healthcare.

The most promising future may be a partnership in which AI helps clinicians evaluate possibilities, identify useful tests, and recognize overlooked evidence while doctors remain responsible for clinical judgment and patient care.

Ultimately, the measure of success will not be how convincingly an AI can name a disease. It will be whether it helps patients receive the right diagnosis, the right tests, and the right treatment at the right time.