Abstract

Adverse drug reactions remain difficult to accurately predict during drug development. Some adverse events are not detected before a drug is widely marketed, causing harm to patients. The effects of drug combinations are also difficult to predict, potentially resulting in adverse drug reactions due to drug-drug interactions. Therefore, there is a need for computational methods to aid the prediction of adverse events. In contrast to pre-clinical and clinical data, post-marketing surveillance of adverse events is based on large numbers of reports from a diverse patient population, for example in terms of concomitant medications. While this makes spontaneous reporting data particularly interesting to the study of adverse events, the data also suffers from a range of biases such as potential confounding factors, complicating the discovery of drug-adverse event associations. The application of propensity score matching, a statistical technique that addresses potential confounders by selecting subsets of patients with similar baseline characteristics for comparison, previously showed promise for analysing spontaneous reporting data. Here, we describe the implementation of a propensity score model on the largest publicly available standardised version of the Food and Drug Administration Adverse Event Reporting System to date. The results show that the approach reduces bias related to drug prescription between patient groups, improving conditions for causal inference about adverse drug reactions. Currently, adverse events associated with individual drugs have been calculated using the method, and the adverse events associated with risperidone and cerivastatin are presented as examples. Future work will include extending the method to detect drug-drug interactions. Another important part of future work is the integration of drug-adverse event associations with other chemical and biological data to help explain and predict adverse drug effects and drug interactions.