Drug-Drug Interaction Signal Detection from Drug Safety Reports

Hundreds of thousands of patients report adverse reactions from using one or more drugs each year. It is impossible to test every drug combination in clinical trials before a drug reaches the market. Thus, the FDA performs post-marketing surveillance to identify drug combinations that produce harmful reactions. To facilitate this surveillance, consumers and health care professionals submit drug safety reports that include the drugs a patient takes and the reactions they experience. In this research, we propose techniques to support high-fidelity rule mining of interesting drug combinations from safety reports by developing drug name matching, reaction name standardization, and known-rule matching strategies. For evaluation, we design a sensibility metric for drug name matching. We demonstrate that our technique achieves a sensibility score of 0.855, corresponding to a 91% accuracy. We compare methods for reaction name standardization and their effects on known-rule matching, identifying 427 known rules from 4652 generated signals when using our techniques as opposed to 61 known rules from 3276 generated signals without the application of our techniques.