What is the Additional Contribution of Mining Electronic Medical Records for Signal Detection? The Experience of EU-ADR Project

Background: In the last ten years, data mining methodologies have been developed using spontaneous reporting system (SRS) databases for drug safety signal detection. Due to challenges associated with SRS, more recently, the possible contribution of mining electronic health records (EHR) for signal detection has been explored. Objective: To compare retrospectively sensitivity and specificity as well as agreement rate of signal detection concerning 10 events in the EUADR database network vs FDA-AERS database. Methods: EU-ADR network comprises of seven databases from four European Countries covering a population of almost 30 million persons. A set of 10 events (e.g. acute myocardial infarction and acute liver injury) warranting priority for monitoring in pharmacovigilance were inspected for their association with drugs captured in EU-ADR and FDA-AERS. In EU-ADR, drugs with statistically significantly (p < 0.05) increased RR (‡2, and ‡1.5 in a sensitivity analysis) were identified as potential signals. In SRS, threshold of EB05>2 was used. To evaluate sensitivity/specificity of the systems, a reference standard was created by selecting for each event 5 drugs being positively and 5 being negatively associated. Moreover, agreement rate between the two system concerning signal detection related to 541 drugs was evaluated. Results: The detection of known drug-event associations in the EUADR system varied based on the nature of the events (e.g. 20%for AMI and 80% for UGIB) and increased substantially if the analyses was restricted to the period preceding first regulatory action (from 20% to 80% forAMI). Across all ten events, EU-ADRand SRS reported respectively a sensitivity = 45% and 77% and a specificity = 96.0% and 97.1%. Sensitivity and specificity of EU-ADR were respectively 72.1% and 82.0% when considering RR‡1.5 as threshold. Looking at the list of 541 drugs, the agreement rate was 66.6% but changed across the events. EU-ADR detected more signals concerning events with high frequency in general population (e.g. AMI, UGIB). Discussion: Mining of electronic medical records is an alternative option for drug safety signal detection and may contribute to complement/strengthen traditional pharmacovigilance methods, especially concerning events with high frequency in general population and not commonly reported as drug-induced. Conclusions: Detection of signals in EHR may change across different types of adverse events and is influenced by the effect of regulatory actions aimed at risk minimization, once the signal is discovered.