Prediction of Drug Interaction and Adverse Reactions, with data from Electronic Health Records, Clinical Reporting, Scientific Literature, and Social Media, using Complexity Science Methods

Abstract

Human health conditions, such as adverse drug reactions (ADR) caused by drug-drug interactions (DDI), are too complex to be tackled effectively by a single domain of expertise. Their associated wide range of data sources, from electronic health record (EHR), social media, to the published scientific literature, requires an interdisciplinary approach common to complexity science and its sub-fields of data and network science. We divide our work in three parts. Using city-wide public health care dispensation records from Blumenau—a mid-size city in southern Brazil—we report primarily on the large number of major DDI being prescribed, with women having a 60% increased risk of DDI when compared to men—the increased risk becomes 90% when only major DDI are considered; this DDI risk also increases with age, with patients age 70-79 having a 34% risk of DDI when they are dispensed two or more drugs concomitantly; and our ability to correctly classify patients with DDI using machine learning techniques. Then we study and predict DDI and ADR from social media data. We focus on different cohorts of interest, for which we build networks from Instagram and Twitter timelines. The network analysis uncovers population-level associations of drugs and symptoms, useful for public health surveillance, as well as affords a means to identify edges to predict putative known and unknown DDI and ADR. Lastly, we present a preliminary study of the timing of DDI observation across different data sources such as social media, clinical reports, and the scientific literature on DDI. We select a set of DDIs and show that social media measurements of DDI and ADR mentions may precede scientific literature when large longitudinal social media data is available. We exemplify with the case for the co-administration of opioids and benzodiazepines. Overall, the results we present in this thesis have important consequences for private and public health policy and regulation, further demonstrating that the methods of complexity science are very useful for studying DDI in particular and public health in general, to the benefit of society.