#

Key takeaways

• Microsoft developed SDBench to evaluate sequential medical diagnosis.

• The benchmark uses two case sets: NEJM case proceedings and MIMIC-IV-ED presentations.

• MAI-DxO works like a panel of virtual physicians across several rounds.

• The system is measured on accuracy, evidence requested, and estimated test costs.

• Microsoft reports strong accuracy and cost-aware performance in both case sets.

What Microsoft says SDBench measures

Microsoft says it developed SDBench to evaluate performance across two complementary sets of cases. One set includes 304 complex cases taken from The New England Journal of Medicine case proceedings. The other set includes more typical emergency-department presentations from MIMIC-IV-ED. Each case becomes interactive, so AI models and physicians start with limited information and choose what to request next.

The benchmark does not only measure the final answer. It also measures the evidence requested and the estimated test costs. That means it assesses both the diagnosis and the path taken to reach it. This is important because the system must decide what information is worth gathering before it commits to a conclusion.

How MAI-DxO works inside the benchmark

Microsoft says it built MAI-DxO and tested it within SDBench. The system does not rely on one model for an immediate answer. Instead, it operates like a panel of virtual physicians. Each participant proposes diagnoses, selects useful questions or tests, and challenges the emerging view.

MAI-DxO coordinates this exchange over several rounds of debate. It updates its assessment as evidence arrives. It also weighs each test's likely value against its cost. That design makes the process sequential rather than one-step. It also makes the system's reasoning more visible through the evidence it asks for.

Reported results on complex cases

Microsoft reports that on the complex NEJM cases, MAI-DxO achieved 83% diagnostic accuracy. Physicians achieved 22% without external reference resources and 36% with them. Microsoft says this made MAI-DxO about 3.7 times more accurate than physicians without external resources, and about 2.3 times more accurate with them.

The company also says the system can be tuned for cost. When optimised more heavily for cost, MAI-DxO still performed strongly while choosing more cost-effective workups. On NEJM cases, one configuration achieved 81% accuracy at less than half the cost of comparative systems. That suggests the benchmark is not only about accuracy. It also rewards systems that can balance performance with resource use.

Reported results on more typical emergency cases

Microsoft also tested MAI-DxO on more typical emergency-department cases from MIMIC-IV-ED. On that set, MAI-DxO achieved 80% accuracy. Physicians achieved 47% accuracy. Microsoft says that made MAI-DxO more than 1.7 times higher in accuracy.

This second case set matters because it is different from the complex NEJM proceedings. The benchmark therefore covers both difficult published cases and more routine emergency presentations. That gives a broader view of how a diagnostic system behaves when the starting information is limited and the next step must be chosen carefully.

Why the benchmark design matters

SDBench is built around the sequence of diagnosis, not just the endpoint. That is a useful distinction. A system can arrive at the right answer while using poor evidence, or it can choose efficient tests and still perform well. Microsoft's benchmark tries to capture both outcomes.

The debate-style setup in MAI-DxO also reflects that idea. Different virtual physicians can disagree, request different evidence, and refine the shared view over time. The result is a process that is closer to iterative clinical reasoning than to a single prompt-and-answer exchange.

Morocco relevance

The source reports no Morocco-specific facts. For readers in Morocco, the global lesson is conditional: if a team evaluates diagnostic AI, it should measure both accuracy and the cost of the evidence it requests.

Bottom line

Microsoft's report presents SDBench as a benchmark for sequential diagnosis, not just final classification. It combines complex NEJM cases and more typical emergency-department cases, then measures accuracy, evidence use, and estimated test costs. Within that setup, MAI-DxO showed strong reported performance and cost-aware behavior across both case sets.