Abstract

Adverse Drug Reaction (ADR) detection from social reviews refers to the task of exploring medical online stores and social reviews for extracting any mention of abnormal reactions that occur after consuming a particular medical product by the consumers themselves. A variety of approaches have been used for extracting ADR from social/medical reviews. These approaches include machine learning, dictionary-based and statistical approaches. Yet, these approaches showed either a high dependency on using an external knowledge source for ADR detection or relying on domain-dependent mechanisms that might lose contextual information. This study aims to propose word sequencing with Long Short-Term Memory (LSTM) architecture. A benchmark dataset of MedSyn has been used in the experiments. Then, a word indexing, mapping, and padding method have been used to represent the words within the reviews as fixed sequences. Such sequences have been fed into the LSTM consequentially. Experimental results showed that the proposed LSTM could achieve an F1 score of up to 92%. Comparing such a finding to the baseline studies reveals the superiority of LSTM. The demonstration of the efficacy of the proposed method has taken different forms including the examination of word indexing with different classifiers, the examination of different features with LSTM, and through the comparison against the baseline studies.