SNOMED CT Entity Linking Benchmark

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

Benchmark
Open
142 joined

Model: Dictionary Backbone with Abstention-Trained Neural Residual
Published

Abstract

A three-stage system that anchors on lexical matching and uses learned components only to fill gaps the dictionary cannot reach.

Stage one is a section-keyed dictionary mapping (section header, normalised mention) pairs to SNOMED CT concept IDs, built from the 272 annotated MIMIC-IV discharge summaries and from the May 2023 International RF2 release restricted to the seven in-scope semantic tags (218,467 concepts, 376,867 terms). Terminology names are expanded with two linguistic rules — "X of Y" inversion and conjunction/preposition substitution — and keys that misfire on training notes are removed by a true/false-positive ratio test. Training-derived keys always take priority; terminology entries only fill gaps, which is what gives the dictionary zero-shot reach over concepts with no annotated examples.

Stage two is a BiomedBERT token classifier tagging BIO spans over the three top-level hierarchies (finding, procedure, body structure), run in 512-token windows with 128-token overlap so mentions crossing a boundary are not lost. It proposes mentions whose surface forms appear nowhere in SNOMED — 97% of the concepts the dictionary misses are phrased in ways no SNOMED name matches, such as "ensures" for Administration of nutritional supplement or "s3" for Third heart sound.

Stage three links the remainder. A SapBERT bi-encoder retrieves the top 50 concepts from a mean-pooled index over the full in-scope terminology, restricted by the tagger's predicted hierarchy; a cross-encoder then reranks them using the mention, its section header, surrounding context, and each candidate's name, synonyms and IS-A parents. Retrieval recall is effectively identical for concepts seen and unseen in training (82.7% versus 83.2% at rank 50), which is the property the dictionary structurally lacks.

The reranker is trained with abstention queries drawn from spans the dictionary fires on that annotators did not label, so it learns to reject an entire candidate set rather than always selecting a best option. Without this it recovered genuinely missed concepts but introduced roughly twice as many spurious ones; with it, the optimal emission threshold falls and the score rises. Emission is deliberately recall-biased, because the macro-averaged metric divides by the union of gold and predicted concept sets, making a plausible new concept positive in expectation while suppression clears a high bar.

Submissions (1)

Submissions
Name Macro char IoUSupport-weighted char IoU
Dictionary Backbone with Abstention-Trained Neural Residual
3w 1d ago
0.4657 0.6165