SNO-PI improves linking clinical terms to SNOMED CT concepts
SNO-PI: Label-Scarce SNOMED CT Entity Linking with Pre-trained Encoders and Curated Terminology Knowledge
Computation and Language
Summary
Linking clinical terms in medical records to the correct medical codes is hard because there are hundreds of thousands of possible codes, but only a few examples of each. The authors introduce SNO-PI, a system that uses advanced AI models trained on both medical notes and medical terminology to better match terms to the right codes, even when there are very few or no labeled examples. Their approach improves the accuracy of this linking by 18% compared to previous methods. However, the system still struggles with codes that have never been seen during training.
What this means in practice
- •For hospital data teams: Automatically link clinical text from discharge summaries to SNOMED CT codes to improve electronic health record data quality and interoperability.
- •For medical coding services: Enhance coding accuracy and efficiency by integrating SNO-PI’s model to identify clinical concepts with limited labeled data.$Commercial implications: Enables development of advanced coding tools that reduce manual work and errors, marketable to hospitals and insurance providers.
Authors
Piyawoot Songsiritat
Abstract
We present SNO-PI, the top-performing submission on SNOMED International's SNOMED CT Entity Linking Benchmark as of 18 October 2026. SNO-PI achieves a macro character IoU of 0.5387, outperforming the organisers' reference system (0.4582; based on the 2024 challenge winner's approach) by 18% relative. Linking clinical mentions to SNOMED CT is a zero- and few-shot problem over a very large label space: the 272 annotated MIMIC-IV discharge summaries cover only 6,595 of the roughly 219,000 eligible concepts, and a third of those occur only once. This motivates a design built on transfer from pre-trained models and human-curated knowledge. SNO-PI combines high-recall span proposal, SapBERT candidate retrieval and listwise ensemble reranking with three BERT-family encoders, each further pre-trained with masked language modelling on unlabelled MIMIC-IV notes and then on SNOMED CT synonyms and hierarchy before task fine-tuning. We find that curated knowledge helps most on concepts never seen in training, yet these concepts remain the main source of error.