SNOMED CT Entity Linking Benchmark

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

Benchmark
Open
180 joined

Model: STLabs Silver-Distilled Linker
Published

TeamSTLabs

Abstract

We present a combination of ideas from STLabs and Team Pi, building on the SNO-PI system (Songsiritat, 2026, arXiv), with thanks to Team Pi. The STLabs base, with high-recall detection distilled from LLM labels and a strong fused retriever (top-1 0.864 overall, 0.686 on concepts unseen in training), is combined with SNO-PI’s span-proposal union, SNOMED CT synonym-to-concept pre-training of the encoders, and listwise reranker with an explicit NONE option that learns mention-level abstention (top-1 0.925 overall, 0.686 unseen, single reranker, cross-validated). The STLabs gate ensemble is retired: with abstention decided at the mention, no concept-level pricing or dictionary floor is needed. The reranker is an ensemble of six members (three encoders, each with a short and a wide context). The LLM-distilled reranking student from STLabs’ previous model, trained on LLM labels over unannotated MIMIC-IV notes, is not yet in this ensemble; it is being rebuilt with a gold-tuned LLM teacher and is pending evaluation as an additional member.

Submissions (10)

Submissions
Name Macro char IoUSupport-weighted char IoU
STLabs Silver-Distilled Linker
2w 4d ago
0.5541 0.6584
STLabs Silver-Distilled Linker
2w 5d ago
0.5462 0.6476
Silver-Distilled Concept Linker
3w ago
0.4972 0.6394
Silver-Distilled Concept Linker
3w 2d ago
0.4985 0.6423
Silver-Distilled Concept Linker
3w 6d ago
0.4970 0.6376
Silver-Distilled Concept Linker
1mo ago
0.4970 0.6424
Silver-Distilled Concept Linker
1mo ago
0.4969 0.6418
Silver-Distilled Concept Linker
1mo 3w ago
0.4849 0.6374
Silver-Distilled Concept Linker
2mo 1w ago
0.4810 0.6380
Silver-Distilled Concept Linker
2mo 1w ago
0.4694 0.6256