Identification of Drug-Related Medical Conditions in Social Media

Identification of Drug-Related Medical Conditions in Social Media Franc¸ois Morlane-Hond`ere Cyril Grouin Pierre Zweigenbaum LIMSI, CNRS, Universit´e Paris-Saclay, F-91405 Orsay {francois.morlane-hondere,cyril.grouin,pz}@limsi.fr Abstract Monitoring social media has been shown to be an interesting approach for the early detection of drug adverse effects. In this paper, we describe a system which extracts medical entities in French drug reviews written by users. We focus on the identification of medical conditions, which is based on the concept of post-coordination: we first extract minimal medical-related entities (pain, stomach) then we combine them to identify complex ones (It was the worst [pain I ever felt in my stomach]). These two steps are respectively performed by two classifiers, the first being based on Conditional Random Fields and the second one on Support Vector Machines. The overall results of the minimal entity classifier are the following: P=0.926; R=0.849; F1=0.886. A thourough analysis of the feature set shows that, when combined with word lemmas, clusters generated by word2vec are the most valuable features. When trained on the output of the first classifier, the second classifier’s performances are the following: p=0.683;r=0.956;f1=0.797. The addition of post-processing rules did not add any significant global improvement but was found to modify the precision/recall ratio. Keywords: Pharmacovigilance; Natural Language Processing; Named Entity Recognition 1. Introduction Most modern drugs and medicines are known to cause ad- verse reactions. Some of them are detected during pre- marketing clinical trial and documented while others are not. Thus, it is crucial to monitor the post-marketing phase of a drug to determine whether its risk/benefit ratio is still positive or not. This activity is called pharmacovigilance. Today, adverse drug reactions are reported spontaneously by patients and professionals via online forms or phone calls to regulating authorities like the Food and Drug Ad- ministration (FDA) in the United States or the Agence Na- tionale de S´ecurit´e du M´edicament et des produits de sant´e (ANSM) in France. However, studies and reports have shown the relative inefficiency of this reporting method, which is mainly due to the lack of awareness of patients and professionals (Alshakka et al., 2013; Acad´emie Nationale de Pharmacie, 2014). Improving the efficiency of the way drug-related problems are identified is critical given the human and economic costs of adverse drug reactions. Recent years have seen the birth of what can be called computer-assisted pharmacovig- ilance or e-pharmacovigilance (Neubert et al., 2013). This consists in using text mining to extract pharmacovigilance- related information from sources such as biomedical arti- cles, electronic health records or drug labels (Harpaz et al., 2014; Shang et al., 2014). Social media have also been used for that purpose. Characteristics of user-generated content are valuable for computer-assisted pharmacovigilance in that it is massive, continuously generated and easy to access. It is also chal- lenging in that the informal writing style that defines user- generated content has been shown to deteriorate the per- formance of traditional NLP tools (Gimpel et al., 2011). A recent review describes twenty-two studies that aim at extracting adverse drug reactions from …

Content truncated to 500 words