Detect and Classify – Joint Span Detection and Classification for Health Outcomes

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8709–8721 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8709 Detect and Classify – Joint Span Detection and Classification for Health Outcomes Micheal Abaho1 Danushka Bollegala1,2∗ 1University of Liverpool, United Kingdom 2Amazon {m.abaho,danushka,prw,shinds}@liverpool.ac.uk Paula Williamson1 Susanna Dodd1 Abstract A health outcome is a measurement or an ob- servation used to capture and assess the effect of a treatment. Automatic detection of health outcomes from text would undoubtedly speed up access to evidence necessary in healthcare decision making. Prior work on outcome de- tection has modelled this task as either (a) a sequence labelling task, where the goal is to detect which text spans describe health out- comes, or (b) a classification task, where the goal is to classify a text into a pre-defined set of categories depending on an outcome that is mentioned somewhere in that text. However, this decoupling of span detection and classi- fication is problematic from a modelling per- spective and ignores global structural corre- spondences between sentence-level and word- level information present in a given text. To ad- dress this, we propose a method that uses both word-level and sentence-level information to simultaneously perform outcome span detec- tion and outcome type classification. In addi- tion to injecting contextual information to hid- den vectors, we use label attention to appro- priately weight both word and sentence level information. Experimental results on several benchmark datasets for health outcome detec- tion show that our proposed method consis- tently outperforms decoupled methods, report- ing competitive results. 1 Introduction Access to the best available evidence in context of patient’s individual conditions enables health- care professionals to administer optimal patient care (Demner-Fushman et al., 2006). Healthcare professionals identify outcomes as a fundamen- tal part of the evidence they require to make de- cisions (van Aken et al., 2021). Williamson et al. ∗Danushka Bollegala holds concurrent appointments as a Professor at University of Liverpool and as an Amazon Scholar. This paper describes work performed at the Univer- sity of Liverpool and is not associated with Amazon. sentence There were no significance between- group differences in the incidence of wheezing or shortness of breath OSD Outcomes: wheezing, shortness of Breath OC Outcome type: Physiological Joint OSD & OC Outcomes-Outcome type wheezing- Physiological Shortness of Breath- Physiological sentence Cumulative incidence and relative risks with 95% confidence intervals for death from any cause, death from prostate cancer, and metastasis were estimated in intention-to-treat and per-protocol analyses. OSD Outcomes: death from any cause, death from prostate cancer OC Outcome type: Mortality Joint OSD & OC Outcomes-Outcome type death from any cause- Mortality death from prostate cancer- Mortality Table 1: Comparing the output of the three separate HOD tasks given two sample sentences. OSD retrieves the outcome spans, OC classifies the text span into a set of outcome types, and Joint OSD & OC retrieves outcomes and classifies them into outcome types. (2017) define an outcome as a measurement or an observation used to capture …

Content truncated to 500 words