Deep learning for pharmacovigilance: recurrent neural network architectures for labeling adverse drug reactions in Twitter posts

Abstract

ADR detection performance in social media is significantly improved by using a contextually aware model and word embeddings formed from large, unlabeled datasets. The approach reduces manual data-labeling requirements and is scalable to large social media datasets.

#Science: The Potential and the Challenges of Utilizing Social Media and Other Electronic Communication Platforms in Health Care

“…All the published studies between 2010 and 2014 took a lexicon-based approach to finding adverse event expressions, a method that is inherently limited given that the variety of expressions present in colloquial text are not typically present in lexicons. More flexible approaches for extraction (machine-learning or pattern based), such as ADRMine 21 and Recurrent Neural Network, 22 are able to capture expressions not present in a lexicon, but require an additional task to be addressed in the aftermath of extraction of such expressions: normalization. Mapping them to standard terms so "my head is being crushed by an elephant" is adequately identified as a report of "migraine.…”

Co-training for Extraction of Adverse Drug Reaction Mentions from Tweets

“…The problem of ADR extraction can be defined as follows. Given a social media post in the form of a word sequence x = x 1 ...., x n , where n is the maximum sequence length, predict an output sequence y 1 , ...., y n , where each y i is encoded using standard sequence labeling encoding scheme such as the IO encoding similar to that used in [4].…”

“…Typically, postmarketing drug safety surveillance (also called as pharmacovigilance) is conducted to identify ADRs after a drug's release. Such surveys rely on formal reporting systems such as Federal Drug Administration's Adverse Event Reporting System (FAERS) 4 . However, often a large fraction (∼94%) of the actual ADR instances are under-reported in such systems [9].…”

“…Earlier work in this direction focused on feature based pipeline followed by a sequence classifier [12]. More recent works are based on Deep Neural Networks [4,13]. Deep learning based methods [5,11] typically rely on the presence of a large annotated corpora, due to their large number of free parameters.…”

L’intelligence artificielle à l’aide des myologues

“…), un très grand nombre de données épidé-miologiques peuvent être recueillies. Leur analyse pourrait permettre d'améliorer les connaissances en santé publique et de surveiller et anticiper le développement de certaines maladies [21], l'efficacité des thérapies administrées [22] ou même d'identifier des réactions indési-rables dues à l'utilisation de certains médicaments [23]. La recherche médicamenteuse pourrait ainsi, elle aussi, profiter de ces outils.…”