A generative model for unsupervised discovery of relations and argument classes from clinical texts

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 519–528, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics A generative model for unsupervised discovery of relations and argument classes from clinical texts Bryan Rink and Sanda Harabagiu Human Language Technology Research Institute University of Texas at Dallas Richardson, TX, USA {bryan,sanda}@hlt.utdallas.edu Abstract This paper presents a generative model for the automatic discovery of relations between entities in electronic medical records. The model discovers relation instances and their types by determining which context tokens ex- press the relation. Additionally, the valid se- mantic classes for each type of relation are de- termined. We show that the model produces clusters of relation trigger words which bet- ter correspond with manually annotated re- lations than several existing clustering tech- niques. The discovered relations reveal some of the implicit semantic structure present in patient records. 1 Introduction Semantic relations in electronic medical records (EMRs) capture important meaning about the as- sociations between medical concepts. Knowledge about how concepts such as medical problems, treat- ments, and tests are related can be used to improve medical care by speeding up the retrieval of relevant patient information or alerting doctors to critical in- formation that may have been overlooked. When doctors write progress notes and discharge sum- maries they include information about how treat- ments (e.g., aspirin, stent) were administered for problems (e.g. pain, lesion) along with the out- come, such as an improvement or deterioration. Ad- ditionally, a doctor will describe the tests (e.g., x- ray, blood sugar level) performed on a patient and whether the tests were conducted to investigate a known problem or revealed a new one. These textual descriptions written in a patient’s record encode im- portant information about the relationships between the problems a patients has, the treatments taken for the problems, and the tests which reveal and investi- gate the problems. The ability to accurately detect semantic rela- tions in EMRs, such as Treatment-Administered-for- Problem, can aid in querying medical records. Af- ter a preprocessing phase in which the relations are detected in all records they can be indexed and re- trieved later as needed. A doctor could search for all the times that a certain treatment has been used on a particular problem, or determine all the treat- ments used for a specific problem. An additional application is the use of the relational information to flag situations that merit further review. If a pa- tient’s medical record indicates a test that was found to reveal a critical problem but no subsequent treat- ment was performed for the problem, the patient’s record could be flagged for review. Similarly, if a Treatment-Worsens-Problem relation is detected previously in a patient’s record, that information can be brought to the attention of a doctor who advises such a treatment in the future. By considering all of the relations present in a corpus, better medical ontologies could be built automatically or existing ones can be improved by adding additional …

Content truncated to 500 words