MedConstructor: de-identified benchmark and blinded cardiology expert-validation dataset
收藏资源简介:
De-identified benchmark and expert-validation data supporting the study of MedConstructor, a protocol-grounded platform that assesses clinical reasoning represented as knowledge graphs. The deposit includes anonymised blinded cardiologist ratings (two-stage design: a 3-expert pilot with 36 variants and 108 ratings, and a 5-expert primary cohort with 73 variants and 365 ratings), the deterministic clinical-weighted model scores and sub-metrics, agreement and inter-rater reliability summaries, controlled graph-evaluation results (evaluator v4.3), RAG retrieval and ablation results, reference-graph quality, and the protocol grounding of the 12 cardiology tasks. All expert-validation values were computed directly from the platform database, which is the single source of truth. Only anonymised, aggregated data are released (experts labelled E1–E5); no patient data, identifiable clinical cases, or expert names are included.



