A Science paper published October 22, 2026, reports that Microsoft’s Diagnostic Orchestrator (DxO), also called MAI-DxO, reached 83.2% accuracy on a controlled benchmark of complex medical cases. Participating physicians averaged 36.1% with reference materials. On a separate emergency-department dataset, DxO scored 80.3%, compared with a physician average of 46.8%. These are benchmark results, not measurements of patient outcomes or routine clinical care. The study tested two distinct datasets through SDBench, a benchmark for sequential diagnosis.
What the two benchmarks measured
SDBench included 304 clinicopathological conference cases from the New England Journal of Medicine (NEJM) and 325 emergency-department-to-admission encounters from MIMIC-IV-ED. The physician averages differed by dataset and by whether physicians could consult references.
| SDBench dataset | Cases | DxO accuracy | Physician comparison |
| NEJM clinicopathological conference (CPC) cases | 304 | 83.2% | 36.1% open-book; 22.3% closed-book |
| MIMIC-IV-ED emergency-department encounters | 325 | 80.3% | 46.8% average |
The 83.2% figure is DxO’s accuracy on the NEJM cases, scored against a reference diagnosis. It does not measure whether patients improved or received safer treatment. The emergency-department result comes from a separate set of cases, so it is a second benchmark, not an extension of the CPC score.
How SDBench moves from a case to a diagnosis
Sequential diagnosis means gathering information in stages before settling on an answer. In SDBench, a system starts with a short case presentation and can request history or examination findings, order tests, revise its list of possible diagnoses, and submit a final diagnosis. A Gatekeeper mediates requests for case information, while a Judge evaluates the final answer.
The Judge uses a physician-authored, five-level rubric to compare the answer with a reference diagnosis. Scores of 4 or 5 count as correct; the paper defines those scores as diagnoses for which management would be unchanged. This scoring rule is what the reported accuracy percentages represent.
Microsoft’s English-language demonstration, published in June 2025, follows one curated NEJM case through SDBench. It shows DxO requesting information and tests, updating its possible diagnoses, tracking estimated costs, and arriving at a diagnosis. It illustrates the benchmark’s workflow; it is not a clinical evaluation.
What the study’s cost estimates represent
For the CPC cases, the paper reports different estimated work-up costs for DxO configurations with different accuracy results. These are estimates in U.S. dollars per case, not patient bills.
| CPC configuration or comparison | Accuracy | Estimated work-up cost per case |
| DxO, no budget constraint | 83.2% | $5,431 |
| DxO, budget-focused | 80.9% | $2,241 |
| DxO, budget-focused | 79.6% | $1,303 |
| DxO, diagnostic-excellence setting | 84.5% | $8,945 |
| Physicians, open-book average | 36.1% | $4,725 |
| Physicians, closed-book average | 22.3% | $3,049 |
The estimates use Current Procedural Terminology (CPT) pricing derived from a Kaiser Permanente Washington chargemaster, a hospital standard-charges file accessed in April 2026. The study assigned history and examination questions a fixed physician-visit cost. For the separate MIMIC-IV-ED cases, DxO averaged $4,979 per case and physicians averaged $3,400 under the same estimated-cost approach.
The paper notes that standardized U.S. prices do not capture real-world variation, test turnaround time, invasiveness, discomfort, or logistical availability. The figures describe the benchmark’s cost model, not what a patient would be charged.
Where the benchmark’s conclusions stop
The study used text-based simulated encounters, not full patient visits or electronic health record workflows. Its datasets exclude healthy or benign encounters, and the setup does not include multimodal evidence such as imaging or pathology slides, communication uncertainty, treatment safety, or longitudinal management. The physician comparison was with individual physicians, not specialist teams. The paper therefore does not establish DxO’s performance in routine care, superiority over specialist teams, or effects on patient outcomes.
The paper states that all authors are Microsoft employees and hold Microsoft equity, and that the studies received no funding from outside Microsoft.
On October 22, Microsoft announced a planned prospective evaluation with Beth Israel Deaconess Medical Center. The announced design would have patients receive standard care while interacting with DxO in parallel, then compare the system’s recommendations with physicians’ independent assessments. The announcement described the evaluation plan, not clinical results.
