Drug Safety Signal Detection from Online Health Communities

Abstract

The development of Web 2.0 and Health 2.0 technologies leads the booming of online health communities (OHCs) such as MedHelp, WebMD and so on. Such platforms are not only empowering individuals to play a substantial role in their own health, but also generating informative consumer-contributed-contents that can be utilized to mine timely and useful knowledge, thus providing automated insights and discovery. Since pharmacovigilance, namely adverse drug reactions (ADRs) and drug-drug interactions (DDIs), represents a serious health problem all over the world, how to detect drug safety signals has drawn many researchers' attention and efforts. Currently, methods proposed to detect ADR and DDI signals are mainly based on traditional data sources such as spontaneous reporting data, electronic health records, pharmaceutical databases, and biomedical literature. However, these data sources are either limited by under-reporting ratio, privacy issues, high cost, or long publication cycle. In this dissertation, we propose to harness consumer-contributed-contents extracted from online health communities for drug safety signal detection and developed various techniques to serve our purpose. Specifically, for ADR detection, we propose to utilize association rule mining to extract associations between drugs and ADRs, and use such measures as confidence, leverage, lift, etc. to capture the association strength. Experiment results demonstrate that our methods are capable of detecting Food and Drug Administration (FDA) alerted ADR signals. Furthermore, we also develop matrix-based and tensor-based techniques to detect ADR signals in a timely manner. The experiment results showed that both approaches are able to detect ADR signals much earlier than FDA's official alert or labeling revision time. Especially, tensor-based method outperforms matrix-based techniques because they are capable of handling missing data that is quite common in the area of social media, and can better capture and exploit temporal patterns. For DDI detection, we also use association rule mining and propose another measure, namely Interaction Ratio, to capture association strength between two drugs and an ADR. Experiment results demonstrated that our method is able to effectively detect DDIs reported by DrugBank. Other than directly mining consumer-contributed-contents, we also propose to construct a heterogeneous healthcare network based on those contents and to utilize link mining techniques for drug safety signal detection. Concretely, for both ADR and DDI detection problems, we extract topological features from the network, quantify the features using different weighting schemas, and use supervised learning techniques to predict drug safety signals. Further, for DDI detection, we propose to detection drug-drug interactions and association ADRs at the same time with triad prediction. Experiment results show that our techniques are effective in discovering both ADR and DDI signals.%%%%Ph.D., Information Studies – Drexel University, 2016