SNOMED CT Entity Linking Benchmark

A benchmark for linking text in medical notes to entities in SNOMED Clinical Taxonomy. #health

Benchmark
Open
153 joined

Model: Silver-Distilled Concept Linker
Published

TeamSTLabs

Abstract

We present a detector distilled from LLM annotation of 20K notes, whose extents are then corrected toward human annotation convention. It proposes candidate extents, paired with a retriever trained under a full-index softmax over the whole terminology that reads each span in its surrounding note text. Every mention a span proposes is pooled with every other mention of that concept across the entire block. There is no re-ranking stage; instead, a gate attending over all mentions decides once whether the concept earns a denominator slot and which of its mentions realise it. The dictionary is emitted in full, and the model is additive on top of it. On the 272 released notes the detector proposes a candidate covering 88.4% of gold mentions; on a held-out fold of 70 of those notes, with the detector out-of-sample and the calibrator fold-blind, it reads detection F1 0.829 and extent F1 0.717. Retrieval R@5 on the 68-note evaluation block is 0.930 overall and 0.762 on the residual stratum.

Submissions (3)

Submissions
Name Macro char IoUSupport-weighted char IoU
Silver-Distilled Concept Linker
1d 19h ago
0.4849 0.6374
Silver-Distilled Concept Linker
3w ago
0.4810 0.6380
Silver-Distilled Concept Linker
3w 2d ago
0.4694 0.6256