Automated Adverse Drug Reaction detection in social media datasets, using deep learning methods (HAN, FastText and CNN)

Zahra Rezaei *, Hossein Ebrahimpour-Komleh *, Behnaz Eslami , Ramyar Chavoshinejad , Mehdi Totonchi ,


Apparently, the health-related studies have been recently in the media spotlight. Social media, like Twitter, has become a valuable online tool to describe the early detection of various adverse drug reactions (ADR). Different medications have adverse effects on various cells and tissues, sometimes more than one cell components would be adversely affected. These types of side effect are occasionally associated with direct or indirect influence of prescribed drug dosage but do not have general unfavorable genetic-based consequences on patients. This study aimed to demonstrate a quick and accurate method to collect and classify information related to the distribution of approved data on Twitter.


We selected "ask a patient" dataset and combination of Twitter "Ask a Patient" datasets including 6,623, 26,934 and 11,623 reviews. We used deep learning methods with the word2vec to classify ADR comments posted by users and present an architecture by HAN, FastText and CNN.


Natural language processing (NLP) deep learning is able to address more advanced peculiarity in learning information compared to other types of machine learning. Additionally, the current study highlighted the advantage of incorporating various semantic features including topics and concepts.


Our approach predicts drugs safety with the accuracy of 93% (combination of Twitter and "Ask a Patient" datasets) in a binary manner. Despite the obvious benefit of various conventional classifiers, deep learning-based text classification methods seems to be precise and influential tools to detect ADR.