Local clinical text helps recommend medical terms despite rare concepts
SNOMED CT Concept Recommendation from Masked Clinical Context
Artificial Intelligence
Summary
Medical records use special codes called SNOMED CT to standardize information, but computers struggle to suggest the right codes when examples are very rare or missing. The authors tested different methods to guess these codes by hiding them in notes and seeing which method found the right answer best. They found that looking closely at the exact words around the hidden code worked better than fancy techniques that tried to add more context. However, the system still missed many rare codes, showing that knowing the exact wording and having enough examples matter a lot.
What this means in practice
- •For hospital data teams: Improve automatic coding of patient discharge summaries to standard medical terms even when few examples exist.
- •For medical transcript editors: Assist in finding appropriate standardized codes from clinical text by using local context within notes.
Authors
Ali Noori
Abstract
Standardizing clinical language to SNOMED CT supports interoperability, analytics, and reusable phenotyping, but concept recommendation remains difficult when relevant concepts are rare or absent from training data. We present a masked-concept recommendation benchmark using the SNOMED CT Entity Linking Challenge v1.2.1 data derived from MIMIC-IV-Note. The dataset contains 75,491 annotations across 272 discharge summaries, with 204 notes used for training and 68 for historical testing. For each unique note-concept pair, the target mention is masked from a local clinical context and the system ranks SNOMED CT concepts observed during training. We compare a popularity baseline, 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, achieving Recall@1 of 14.81%, Recall@10 of 33.43%, MRR of 0.2114, and nDCG@10 of 0.2297. Retrieval augmentation does not improve this baseline, with Recall@10 of 31.99% and MRR of 0.1937. Performance is strongly affected by concept frequency: Recall@10 is 7.74% for concepts appearing in only one or two training notes versus 43.90% for concepts appearing in more than ten. In addition, 9.66% of test note-concept pairs contain concepts unseen during training. These findings show that local lexical context and terminology coverage are major determinants of recommendation quality in low-resource settings and provide a reproducible baseline for future ontology-grounded and biomedical-encoder retrieval systems.