Adverse Drug Reaction Classification With Deep Neural Networks

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 877–887, Osaka, Japan, December 11-17 2016. Adverse Drug Reaction Classification With Deep Neural Networks Trung Huynh1,3, Yulan He2, Alistair Willis1 and Stefan R¨uger1 1Knowledge Media Institute, Open University, UK 2Systems Analytics Research Institute, Aston University, UK 3Google UK trunghlt@gmail.com, y.he@cantab.net {alistair.willis, stefan.rueger}@open.ac.uk Abstract We study the problem of detecting sentences describing adverse drug reactions (ADRs) and frame the problem as binary classification. We investigate different neural network (NN) architectures for ADR classification. In particular, we propose two new neural network models, Convolu- tional Recurrent Neural Network (CRNN) by concatenating convolutional neural networks with recurrent neural networks, and Convolutional Neural Network with Attention (CNNA) by adding attention weights into convolutional neural networks. We evaluate various NN architectures on a Twitter dataset containing informal language and an Adverse Drug Effects (ADE) dataset con- structed by sampling from MEDLINE case reports. Experimental results show that all the NN architectures outperform the traditional maximum entropy classifiers trained from n-grams with different weighting strategies considerably on both datasets. On the Twitter dataset, all the NN architectures perform similarly. But on the ADE dataset, CNN performs better than other more complex CNN variants. Nevertheless, CNNA allows the visualisation of attention weights of words when making classification decisions and hence is more appropriate for the extraction of word subsequences describing ADRs. 1 Introduction Adverse Drug Reactions (ADRs) are potentially very dangerous to patients and are amongst the top causes of morbidity and mortality (Pirmohamed et al., 2004). Many ADRs are hard to discover as they happen to certain groups of people in certain conditions and they may take a long time to expose. Healthcare providers conduct clinical trials to discover ADRs before selling the products but normally are limited in numbers. Thus, post-market drug safety monitoring is required to help discover ADRs after the drugs are sold on the market. In the United States, Spontaneous Reporting Systems (SRSs) is the official channel supported by the Food and Drug Administration. However these system are typically under-reported and many ADRs are not recorded in the systems. Recently unstructured data such as medical reports (Gurulingappa et al., 2012b; Gurulingappa et al., 2012a) or social network data (Ginn et al., 2014; Nikfarjam et al., 2015) have been used to detect content that contains ADRs. Case reports published in the scientific biomedical literature are abundant and generated rapidly. Social networks are another source of redundant data with unstructured format. While an individual tweet or Facebook status that contains ADRs may not be clinically useful, a large volume of these data can expose serious or unknown consequences. Common approaches to detect content with ADRs used Support Vector Machines (SVMs), Random Forest, Maximum Entropy classifiers with heavily hand-engineered features (Rastergar-Mojarad et al., 2016; Sarker et al., 2016; Zhang et al., 2016). These features normally include n-grams with different weighting schemes. When used with unigrams, these approaches suffer from the fact that their models do not take in account the interaction …

Content truncated to 500 words