Training Data Augmentation for Detecting Adverse Drug Reactions in User-Generated Content

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2349–2359, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2349 Training Data Augmentation for Detecting Adverse Drug Reactions in User-Generated Content Sepideh Mesbah1, Jie Yang2, Robert-Jan Sips3, Manuel Valle Torre1, Christoph Lofi1, Alessandro Bozzon1, and Geert-Jan Houben1 1Delft University of Technology, Van Mourik Broekmanweg 62628 XE Delft , the Netherlands 2Amazon∗, 440 Terry Ave N, Seattle, WA 98109, USA 3MyTomorrows, Anthony Fokkerweg 61, 1059 CP Amsterdam, the Netherlands s.mesbah@tudelft.nl, jiy@amazon.com, r.sips@mytomorrows.com {m.valletorre, c.lofi, a.bozzon, g.j.p.m.houben}@tudelft.nl Abstract Social media provides a timely yet challenging data source for adverse drug reaction (ADR) detection. Existing dictionary-based, semi- supervised learning approaches are intrinsi- cally limited by the coverage and maintain- ability of laymen health vocabularies. In this paper, we introduce a data augmentation ap- proach that leverages variational autoencoders to learn high-quality data distributions from a large unlabeled dataset, and subsequently, to automatically generate a large labeled training set from a small set of labeled samples. This allows for efficient social-media ADR detec- tion with low training and re-training costs to adapt to the changes and emergence of infor- mal medical laymen terms. An extensive eval- uation performed on Twitter and Reddit data shows that our approach matches the perfor- mance of fully-supervised approaches while requiring only 25% of training data. 1 Introduction Adverse Drug Reactions (ADRs) is the fourth leading cause of death in the United States (Chee et al., 2011). ADR detection is, therefore, a crit- ical element of drug safety. Studies have shown that clinical trials are not able to fully character- ize drugs’ adverse effects (Harpaz et al., 2012; Chee et al., 2011; Ahmad, 2003). Traditional tech- niques of post-market ADR mainly rely on volun- tary and mandatory reporting of ADRs by patients and health providers, but they suffer from delays in reporting, under-reporting, or data incomplete- ness (Sarker et al., 2015). Social media is becoming a preferred chan- nel for millions of users and patients to share, discuss, and seek health information (He et al., 2017); such user-generated content can, therefore, provide valuable insights for monitoring Adverse ∗Work performed before joining Amazon. Drug Reactions (ADRs) from an additional point of view (Lee et al., 2017; Sarker and Gonzalez, 2015; Aramaki et al., 2011). ADR detection from social media is, however challenging, as online users report ADRs using a different language style and terminology that largely depend on the user’s medical proficiency, as well as the type of online medium (e.g., health forums vs micro-post social networks). In particular, laymen often use diverse dialects (Karisani and Agichtein, 2018) when de- scribing medical concepts, and make abundant use of informal terminology. Existing approaches for detecting terms in in- formal medical language mainly rely on semi- manually generated dictionaries (e.g. laymen health dictionaries) (Zeng and Tse, 2006), or su- pervised machine-learning-based sequence classi- fiers (Huynh et al., 2016; Chowdhury et al., 2018). Due to the language …

Content truncated to 500 words