Attend to Medical Ontologies: Content Selection for Clinical Abstractive Summarization

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1899–1905 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1899 Attend to Medical Ontologies: Content Selection for Clinical Abstractive Summarization Sajad Sotudeh1, Nazli Goharian1, and Ross W. Filice2 1IR Lab, Georgetown University, Washington DC 20057, USA {sajad, nazli}@ir.cs.georgetown.edu 2MedStar Georgetown University Hospital, Washington DC 20007, USA ross.w.filice@medstar.net Abstract Sequence-to-sequence (seq2seq) network is a well-established model for text summarization task. It can learn to produce readable con- tent; however, it falls short in effectively iden- tifying key regions of the source. In this pa- per, we approach the content selection prob- lem for clinical abstractive summarization by augmenting salient ontological terms into the summarizer. Our experiments on two pub- licly available clinical data sets (107,372 re- ports of MIMIC-CXR, and 3,366 reports of OpenI) show that our model statistically signif- icantly boosts state-of-the-art results in terms of ROUGE metrics (with improvements: 2.9% RG-1, 2.5% RG-2, 1.9% RG-L), in the health- care domain where any range of improvement impacts patients’ welfare. 1 Introduction Radiology reports convey the detailed observations along with the significant findings about a medical encounter. Each radiology report contains two im- portant sections:1 FINDINGS that encompasses ra- diologist’s detailed observations and interpretation of imaging study, and IMPRESSION summarizing the most critical findings. IMPRESSION (usually couple of lines and thrice smaller than finding) is considered as the most integral part of report (Ware et al., 2017) as it plays a key role in communicating critical findings to referring clinicians. Previous studies have reported that clinicians mostly read the IMPRESSION as they have less time to review find- ings, particularly those that are lengthy or intricate (Flanders and Lakhani, 2012; Xie et al., 2019). In clinical setting, generating IMPRESSION from FINDINGS can be subject to errors (Gershanik et al., 2011; Brady, 2016). This fact is especially crucial when it comes to healthcare domain where even 1Depending on institution, radiology reports may or may not include other fields such as BACKGROUND. the smallest improvement in generating IMPRES- SION can improve patients’ well-being. Automat- ing the process of impression generation in radi- ology reporting would save clinicians’ read time and decrease fatigue (Flanders and Lakhani, 2012; Kovacs et al., 2018) as clinicians would only need to proofread summaries or make minor edits. Previously, MacAvaney et al. (2019) showed that augmenting the summarizer with entire on- tology (i.e., clinical) terms within the FINDINGS can improve the content selection and summary generation to some noticeable extent. Our findings, further, suggest that radiologists select significant ontology terms, but not all such terms, to write the IMPRESSION. Following this paradigm, we hy- pothesize that selecting the most significant clinical terms occurring in the FINDINGS and then incorpo- rating them into the summarization would improve the final IMPRESSION generation. We further ex- amine if refining FINDINGS word representations according to the identified clinical terms would result in improved IMPRESSION generation. Overall, the contributions of this work are twofold: (i) We propose a novel …

Content truncated to 500 words