MedCase-Structured (5-Shot): leaderboard
Diagnostic reasoning over structured EHR data: MedCaseReasoning case reports converted into HL7 FHIR R4 bundles with SNOMED CT, LOINC, RxNorm and CVX codes and the diagnosis hidden (MedCase-Structured, Muti et al., 2026). An LLM judge (Claude Sonnet 4.6) decides whether the named diagnosis is clinically equivalent to the reference; accuracy in percent, mean of three runs at medium reasoning; higher is better. Five training cases are shown as worked examples.
Metric: Diagnostic accuracy (%; the model reads a case as a terminology-coded FHIR R4 bundle with the final diagnosis hidden and names the most likely diagnosis; an LLM judge (Claude Sonnet 4.6) decides clinical equivalence to the reference; mean of three runs; five examples from the training split in the prompt; medium reasoning, temperature 1.0, 800 generated tokens). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 58.18 | #121 (GPT-5.4) |
| 2 | Claude Opus 4.6 (Medium) | 56.3 | #100 (Claude Opus 4.6) |
| 3 | Gemini 3.1 Pro (Preview) (Medium) | 40.91 | #91 (Gemini 3.1 Pro (Preview)) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=medcase-structured-5-shot · How It Works · Data refreshed daily, snapshot 2026-11-12.