Microsoft has unveiled the Microsoft AI Diagnostic Orchestrator (MAI-DxO), an AI diagnosis system that works through patient cases step by step like a human doctor. The findings were published in the journal Science on 22 October 2026 after testing on both complex and routine emergency cases. The system scored far higher than generalist physicians while aiming to lower test costs, a result Microsoft describes as a step toward medical superintelligence.
What Is AI Diagnosis and the Microsoft Diagnostic Orchestrator?
AI diagnosis is the use of artificial intelligence to study symptoms, patient history and test reports to suggest the likely disease and next tests. The Microsoft Diagnostic Orchestrator (MAI-DxO) is Microsoft’s research system for this task that acts like a virtual panel of doctors debating and revising the diagnosis.
Diagnostic AI refers to software that supports the diagnostic process. It does not simply match one symptom to one disease. It builds a list of possible conditions, called a differential diagnosis, asks follow up questions, orders tests and updates its view as new evidence arrives. Clinical reasoning is the name doctors give to this step by step thinking.
Microsoft developed MAI-DxO inside Microsoft AI, the company’s AI division led by Chief Executive Officer Mustafa Suleyman (as of October 2026). Microsoft was founded in 1975 and is headquartered in Redmond, Washington in the United States. The company is led by Chairman and Chief Executive Officer Satya Nadella (as of October 2026). MAI-DxO fits into Microsoft’s wider medical AI work, which includes its 2022 acquisition of clinical speech firm Nuance and its partnership with the Mayo Clinic to build a specialised health AI model.
The larger goal behind MAI-DxO is what Microsoft calls medical superintelligence. The idea is an AI that can reason across complex medical information and help extend expert level care to places and patients that lack specialists. Microsoft stresses that MAI-DxO is still a research demonstration. It is not approved for clinical use and cannot be downloaded or used as a patient app.
How MAI-DxO Mimics Clinical Reasoning of Doctors
MAI-DxO starts with only a short case summary, such as the main complaint and basic details. It then decides what to ask next, which test to order and when it has enough proof to give a final diagnosis. This loop mirrors how doctors build the picture gradually instead of receiving all reports at once.
The system works by dividing the job among five virtual roles played by the same underlying language model. Each role handles one part of the reasoning and then debates with the others. Microsoft calls this structured discussion a Chain of Debate. After each round, the panel agrees on one action: ask more questions, order tests or commit to a diagnosis.
| Virtual Role | Job in the Panel |
|---|---|
| Dr. Hypothesis | Keeps a ranked list of the top three likely diagnoses and updates their chances after every new finding |
| Dr. Test-Chooser | Picks up to three tests per round that will best separate the leading possibilities |
| Dr. Challenger | Acts as a critic, points to conflicting evidence and suggests tests that could prove the leading view wrong |
| Dr. Stewardship | Controls cost by favouring cheaper but equally useful tests and rejecting low value expensive tests |
| Dr. Checklist | Checks that test names are valid and that the panel’s reasoning stays consistent |
A separate cost tracker can estimate the running bill and flag costly orders before they go through. MAI-DxO is model agnostic, which means it can sit on top of many different large language models. Microsoft tested it with models from OpenAI, Google, Anthropic, Meta and DeepSeek. The strongest published results used OpenAI’s GPT-5 and GPT-5.5, while an earlier preprint used OpenAI’s o3. In practice the panel asked far more history questions before its first test than physicians did, a sign of detailed upfront reasoning.
Is AI Diagnosis Better Than Doctors? Accuracy Results Compared
In Microsoft’s controlled tests, MAI-DxO’s AI diagnosis was more accurate than individual generalist physicians. It scored 83.2 percent on complex NEJM cases against 36.1 percent for doctors with references, and 80.3 percent on emergency admissions against 46.8 percent for doctors.
The head to head comparison used two very different case sets. The complex set had 304 clinicopathological conference cases from the New England Journal of Medicine (NEJM), a weekly medical journal published since 1812 that is known for difficult teaching cases. The routine set had 325 emergency department to admission cases drawn from MIMIC-IV-ED, a large United States hospital database from the Beth Israel Deaconess Medical Center in Boston. Doctors in the test were active primary care, hospital and emergency physicians from the United States and United Kingdom with 5 to 20 years of experience.
| Test Group | MAI-DxO with GPT-5 | Generalist Physicians |
|---|---|---|
| Complex NEJM cases (304 cases) | 83.2% correct | 36.1% correct with books and internet, 22.3% without them |
| Emergency admissions (325 cases) | 80.3% correct | 46.8% correct on average |
The gap was widest on the hardest cases. Against unaided doctors, MAI-DxO was nearly four times as accurate. Even the best individual doctor in the study reached 51.8% on NEJM cases and 62.5% on emergency cases, still well below the AI system on both accuracy and cost in that setup. When MAI-DxO made mistakes, they were usually close misses at the wrong level of detail. Physician errors tended to land further from the correct answer.
Raw model power explains part of the gain. GPT-5 alone scored about 78.9% on complex cases. Adding DxO lifted that by about four points while cutting estimated cost by about $3,000 per case. With the newer GPT-5.5, DxO raised accuracy from about 72% to nearly 87% on complex cases. On routine emergency cases, both GPT-5.5 with and without DxO scored near 85%, which shows the orchestrator helps most on hard cases and on weaker base models. For example, Google’s Gemini 2.5 Flash rose from about 46.4% alone to nearly 69% with DxO.
What Is SDBench and How Was MAI-DxO Tested?
SDBench is a test that forces AI and doctors to diagnose step by step from limited information. SDBench stands for Sequential Diagnosis Benchmark. It was built by Microsoft researchers to replace static multiple choice tests with interactive cases that measure both accuracy and cost.
Older benchmarks give the full case write up at the start. That helps AI systems that are good at recall but hides how doctors actually work. SDBench starts with only a brief vignette. The participant must ask for history, request exams and order tests one round at a time. A Gatekeeper model holds the full case file and reveals a finding only when it is asked for in a clear way. Vague requests are refused. Imaging and key confirmatory results stay hidden until the exact test is ordered. A separate Judge model checks whether the final diagnosis matches the reference answer. A Cost Estimator adds up test prices using United States hospital standard charges from Kaiser Permanente Washington.
The study behind these results is titled Sequential Diagnosis with Language Models. It was led by Microsoft researchers including Harsha Nori and colleagues and published in Science on 22 October 2026. Science is a peer reviewed journal published by the American Association for the Advancement of Science since 1880. The paper was received in April 2026 and accepted in August 2026. An earlier preprint had appeared on arXiv in June 2025 with results based on the o3 model.
How Can Researchers Access the Paper and SDBench Code?
Researchers can read the peer reviewed paper in Science and the preprint on arXiv, and can test the benchmark through Microsoft’s open code. The SDBench code including the Gatekeeper, Judge and cost tools is shared under an MIT licence on GitHub, with data access subject to NEJM and PhysioNet rules.
MAI-DxO itself is not offered as a public app, download or online diagnosis page. There is no patient facing MAI-DxO login or free AI diagnosis tool from Microsoft at present. Microsoft has open sourced SDBench for selected NEJM cases so outside teams can repeat the tests and challenge the claims. Full NEJM case text needs a licence from the NEJM Group for non commercial research. The emergency cases come from the curated MIMIC-IV-ED set, which needs PhysioNet approval and agreement to the data use terms.
Cost of Diagnosis and the Push for Affordable Care
Accuracy alone does not make a diagnosis useful. Every extra scan adds money, waiting time and discomfort. Microsoft therefore scored MAI-DxO on two fronts at once: whether the final answer was right and how much the ordered tests would cost in a United States hospital price list.
Without a budget limit, MAI-DxO ordered thorough workups that cost more on average than physician workups. Its mean estimated cost was $5,431 per NEJM case against $4,725 for physicians, and $4,979 per emergency case against $3,400 for physicians. When Microsoft switched on budget controls, the system kept most of its accuracy while spending much less. The budgeted versions cancelled low value expensive tests and leaned on careful questioning first.
| Budget Setting | Accuracy | Mean Estimated Test Cost |
|---|---|---|
| NEJM cases, higher cost setting | 80.9% | $2,241 |
| NEJM cases, lower cost setting | 79.6% | $1,303 |
| Emergency cases, higher cost setting | 73.2% | $2,634 |
| Emergency cases, lower cost setting | 69.5% | $2,037 |
Each of these cheaper settings still beat the average physician score while costing less than the average physician workup. Microsoft reports that no human or AI method tested was both more accurate and less costly than MAI-DxO across the two datasets. The cost figures are estimates from standard charges. They do not capture pain, risk, delays or local Indian price levels, so they show direction rather than exact hospital bills.
Limitations, Safety and What Comes Next for Medical AI
The Science paper is careful about what the numbers do not prove. The comparison was with individual generalist physicians, not with specialists matched to each case or with teams of doctors working together. The authors state that the gap would likely narrow against case matched specialists. Complex NEJM cases are chosen for teaching value and often carry specialist conclusions, which makes them unusually hard for a single generalist.
The benchmark was text only. It did not test reading of scans, waveforms, pathology slides or bedside examination. It also left out real world uncertainty such as patients who forget details or whose condition changes over time. Every case ended in a firm pathological diagnosis, so healthy visits and benign complaints were excluded. Direct transfers to intensive care were left out of the emergency set during data cleaning. All of this means the test reflects idealised diagnostic reasoning, not full clinical practice. It did not measure treatment choice, patient safety or outcomes.
Funding and independence also matter for readers to know. All study authors were Microsoft employees holding equity in the company, and Microsoft funded the research. Peer review in Science adds weight, but outside replication using the open SDBench code will be important.
The next step is live testing. Microsoft has announced a partnership with Beth Israel Deaconess Medical Center in Boston to study whether the system can help with triage and reasoning in emergency settings. Patients would interact with the system in a supervised study while performance is checked against independent physician views. Separately, Microsoft is working with the Mayo Clinic, a leading United States hospital group, to develop a frontier health AI model for clinical reasoning tasks such as earlier detection and personalised treatment planning. Microsoft says any future clinical use would need rigorous safety testing, regulatory review and clear accountability.
Why MAI-DxO Matters for India and Global Healthcare
For health systems, the promise is wider access to careful reasoning at lower cost. Many patients first meet a general physician or an emergency doctor, not a specialist team. A tool that helps that first doctor ask sharper questions and avoid low value tests could shorten the path to the right treatment. That matters in countries with specialist shortages and high out of pocket spending on diagnostics.
India has built a digital base that makes such tools relevant but also sets clear guardrails. The Ayushman Bharat Digital Mission (ABDM) provides health IDs, interoperable records and consent based sharing that can feed structured data to AI systems. In February 2026, the Health Ministry launched the Strategy for Artificial Intelligence in Healthcare for India (SAHI) along with the Benchmarking Open Data Platform for Health AI (BODH), built with IIT Kanpur, to test and validate health AI before large scale use. India’s framework stresses that AI must support human judgment, stay explainable and equitable across languages and settings, and follow the Digital Personal Data Protection Act and health data rules.
MAI-DxO fits that test first approach. Its SDBench method shows how to score both accuracy and cost, an idea close to BODH’s goal of structured validation. Indian researchers can run the open benchmark, adapt cost sheets to Indian price levels and check performance on local case mixes. Real use in India would still need trials in Indian hospitals, support for Indian languages, checks for bias across groups and approval under medical device and data rules. The Science result is therefore best read as a research signal, not a ready clinic tool.
Key Takeaways
- The Microsoft AI Diagnostic Orchestrator (MAI-DxO) was published in Science on 22 October 2026 as a system that mimics step by step clinical reasoning.
- MAI-DxO scored 83.2% on 304 NEJM complex cases compared with 36.1% for generalist physicians with reference tools.
- MAI-DxO scored 80.3% on 325 MIMIC-IV-ED emergency admissions compared with 46.8% for physicians.
- SDBench (Sequential Diagnosis Benchmark) tests accuracy and cost using a Gatekeeper model, with code shared on GitHub under an MIT licence.
- Budgeted MAI-DxO held 79.6% to 80.9% accuracy on NEJM cases at $1,303 to $2,241, below average physician test costs.
- MAI-DxO is a research demonstration only and is not approved for clinical use, with live validation planned with Beth Israel Deaconess Medical Center.
