Key Notes
- Microsoft’s MAI-DxO achieved 83.2% accuracy on complex diagnostic cases and 80.3% on emergency-admission cases in controlled simulations, exceeding the participating physicians’ averages.
- Budget-constrained configurations lowered estimated workup costs while retaining an accuracy advantage, although the headline AI settings cost more than the physician averages.
- The study did not test bedside care, specialist teams, treatment decisions or patient outcomes, which remain essential questions for clinical evaluation.
Microsoft’s diagnostic AI achieved substantially higher accuracy than participating generalist physicians in a study published in Science, correctly identifying 83.2% of complex medical cases compared with 36.1% for doctors using reference resources. On cases drawn from emergency-department admissions, the system scored 80.3%, against a physician average of 46.8%.
The results come from controlled, text-only simulations rather than consultations with live patients. The October 22 paper, Sequential diagnosis with language models, presents a system that asks questions, requests tests and revises its diagnosis as evidence arrives. It also shows that accuracy and cost depend on how the system is configured.
Asking Better Questions Before Ordering Tests
The system, called the Microsoft AI Diagnostic Orchestrator, or MAI-DxO, coordinates a language model through a simulated panel of clinical experts. Different reasoning roles develop possible diagnoses, select informative questions or tests and challenge the emerging conclusion. The aim is to keep a plausible initial explanation from becoming a diagnosis accepted too quickly.
This is a different task from answering a medical examination question containing every relevant detail. In the Sequential Diagnosis Benchmark, or SDBench, the diagnostician starts with a brief case summary and decides what information to seek next. A separate Gatekeeper supplies information from the record, while a Judge scores the final diagnosis against the recorded answer.
The researchers observed a striking difference in approach. On the complex cases, DxO using OpenAI’s GPT-5 asked an average of 25.5 questions before its first test request, compared with 5.6 for physicians. More questioning can expose a useful detail before a broad set of investigations is ordered, although a longer questioning sequence would also need to be practical in a real consultation.
What the Accuracy Scores Actually Measure
SDBench was built from 304 clinicopathological conference cases published in the New England Journal of Medicine and 325 emergency-department-to-admission encounters from MIMIC-IV-ED. These were converted into interactive diagnostic challenges. The headline results describe performance within that environment, rather than the percentage of ordinary patients an AI would diagnose correctly.
The 36.1% physician result came from the condition allowing search and reference tools. Without those resources, the average was 22.3%. The physician comparison on complex cases used a subset of 56 cases held out during development, with multiple physician attempts across those cases; the study’s full benchmark corpus and the human test subset should not be confused.
Participants worked individually. The authors explicitly say the comparison does not establish superiority over specialists matched to each case, multidisciplinary teams or real-world clinical practice. That matters because the difficult published cases are selected for their teaching value and often involve problems that would prompt specialist consultation.
Lower Costs Required a Different Operating Setting
The headline accuracy result was not also the cheapest setting. On complex cases, the 83.2% AI result carried an average estimated workup cost of $5,431, compared with $4,725 for physicians using references. Under tighter budget constraints, the AI reached 80.9% at $2,241 and 79.6% at $1,303, retaining an accuracy advantage while lowering estimated costs.
The emergency-case results showed the same trade-off. The 80.3% AI configuration cost an estimated $4,979 per case, compared with $3,400 for physicians. A budget-constrained version achieved 69.5% accuracy at $2,037. It is therefore accurate to say that some settings were both more accurate and less expensive than the physician average, but not that every AI configuration delivered the maximum accuracy and the lowest cost together.
These are standardized cost estimates based on U.S. hospital pricing data, not observed reductions in patients’ bills. The calculation also does not fully capture factors such as waiting for a result, the burden of an invasive procedure or whether a test is available locally. Those considerations can change which diagnostic path is appropriate even when a model assigns it an attractive price.
A Diagnosis Benchmark Leaves Much of Medicine Untested
The study excludes healthy or benign encounters, so its disease mix does not mirror routine practice. Neither the AI nor the physicians examined patients directly or interpreted the full range of raw images, waveforms and pathology slides used in clinical work. Missing information was mediated through a text-based system, which can generate diagnostically neutral synthetic findings when a requested result is absent from the source record.
The assessment focused on identifying disease, rather than treatment decisions, patient safety or outcomes over time. Our coverage of AI prognosis research highlighted a related distinction: predicting risk, identifying a diagnosis and improving care are different claims that require different evidence.
The paper also discloses the researchers’ commercial interests. All authors are Microsoft employees and hold company equity; eight are inventors on a related patent application. The study received no funding from outside Microsoft. Those disclosures provide context for evaluating the findings and the value of independent replication.
The Next Test Is Working Alongside Clinicians
Microsoft’s announcement says it plans prospective research with Beth Israel Deaconess Medical Center in Boston. Patients would interact directly with the system, with its performance compared against physicians’ independent assessments. That planned work could reveal problems that a carefully controlled case simulation cannot reproduce.
The company has also released benchmark resources so other researchers can examine the evaluation and test additional models, subject to the access conditions of the clinical datasets. Our earlier reporting on Microsoft’s lung-cancer collaboration concerned a more specific imaging workflow; MAI-DxO explores the broader sequence of gathering evidence and revising a diagnostic hypothesis.
The promising result is that a structured AI workflow can improve diagnostic reasoning and its estimated cost trade-off in a demanding test. Establishing whether it helps clinicians deliver safer, faster and more effective care requires the next stage of evidence. The Science study supplies a research foundation for that work, rather than a verdict that AI has replaced clinical judgment.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
