Exploiting Social Media with Tensor Decomposition for Pharmacovigilance

Adverse drug reactions (ADRs) represent a serious health problem all over the world, and how to detect ADRs in an early stage has drawn many researchers´ attention and efforts. There are an increasing number of studies focused on this area and many techniques have been proposed to detect ADRs based on various data sources such as spontaneous reporting data, electronic health record, pharmaceutical databases, and biomedical literature. However, these data sources are limited by high cost, under-reporting ratio, privacy issues, or long cycle that publishing a paper in journals could usually take months or even a year. In this work, we propose to detect early ADR signals from social media data. We collected threads of 20 drugs from MedHelp, extracted 14 adverse reactions, either alerted by FDA or added on drug labels, with their alert releasing or labeling revision time being gold standard, and utilize confidence, leverage and lift to identify ADR signals. We also propose to use tensor decomposition to handle the sparseness and missing data in social media. The experiment results showed that both matrix-based and tensor-based approaches are able to detect ADR signals much earlier than FDA´s official alert or labeling revision time. Especially, tensor-based method outperformed matrix-based techniques and can better capture temporal patterns.