SNOMED CT Concept Recommendation from Masked Clinical Context

Summary

The paper presents a masked-concept recommendation benchmark for mapping clinical context to SNOMED CT concepts, using data from the SNOMED CT Entity Linking Challenge v1.2.1 and MIMIC-IV-Note. The dataset contains 75,491 annotations across 272 discharge summaries, split into 204 training notes and 68 historical test notes. For each unique note-concept pair, the target mention is masked from a local context, and the system ranks concepts seen during training. The comparison covers popularity, sparse TF-IDF concept prototypes, dense latent semantic analysis embeddings, sparse-dense fusion, retrieved-note evidence, and a retrieval-augmented hybrid. Sparse TF-IDF performs best, reaching 14.81% Recall@1, 33.43% Recall@10, an MRR of 0.2114, and nDCG@10 of 0.2297. Retrieval augmentation performs worse than this baseline, with 31.99% Recall@10 and an MRR of 0.1937. Frequency has a major effect: Recall@10 is 7.74% for concepts appearing in one or two training notes, compared with 43.90% for concepts appearing in more than ten. A further 9.66% of test pairs contain concepts absent from training. The results establish a reproducible baseline and indicate that local lexical context and terminology coverage are central challenges for low-resource clinical concept recommendation.