Clinical Concept Linking with Contextualized Neural Representations

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8585–8592 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 8585 Clinical Concept Linking with Contextualized Neural Representations Elliot Schumacher1 Andriy Mulyar2∗ Mark Dredze1 1Johns Hopkins University 2Virginia Commonwealth University {eschumac,mdredze}@cs.jhu.edu aymulyar@vcu.edu Abstract In traditional approaches to entity linking, link- ing decisions are based on three sources of in- formation – the similarity of the mention string to an entity’s name, the similarity of the con- text of the document to the entity, and broader information about the knowledge base (KB). In some domains, there is little contextual in- formation present in the KB and thus we rely more heavily on mention string similarity. We consider one example of this, concept link- ing, which seeks to link mentions of medical concepts to a medical concept ontology. We propose an approach to concept linking that leverages recent work in contextualized neu- ral models, such as ELMo (Peters et al., 2018), which create a token representation that inte- grates the surrounding context of the mention and concept name. We find a neural ranking approach paired with contextualized embed- dings provides gains over a competitive base- line (Leaman et al., 2013). Additionally, we find that a pre-training step using synonyms from the ontology offers a useful initialization for the ranker. 1 Introduction Medical concept linking produces structured top- ical content from clinical free text (Aronson and Lang, 2010). Healthcare providers often refer to medical concepts in clinical text notes that are ab- sent from associated health record metadata despite their importance to understanding a patient’s medi- cal status. For example, in The patient reports a his- tory of seizure disorder..., the phrase seizure disor- der refers to the concept epilepsy contained within the Unified Medical Language System (UMLS) on- tology (Bodenreider, 2004). However, this may be absent from metadata as it is not part of the current diagnosis. Concept mentions can use non-standard ∗ Contribution performed during an internship at Johns Hopkins University. terms (e.g. epilepsy), thus concept linking requires non-lexical methods. Additionally, some terms (cancer) are ambiguous and could refer to multiple concepts (breast cancer, colon cancer, etc.) The related task of Entity Linking – linking named entities (people, places, and organizations) to a knowledge base – has been explored in non- medical domains (Dredze et al., 2010; Durrett and Klein, 2014; Gupta et al., 2017). Entity linking systems consider three sources of information: 1) similarity between mention strings and names for the KB entity; 2) comparison of the document con- text to information about the KB entity (e.g. entity description); 3) information contained in the KB, such as entity popularity or inter-entity relations. In contrast to the dense KBs in entity linking, concept linking uses sparse ontologies, which con- tain a unique identifier (CUI), title, and links to synonyms and related concepts, but rarely long- form text. For example, while the concept epilepsy has many synonyms in UMLS, it has no defini- tion or other long description. …

Content truncated to 500 words