For How Many and What Types of Drugs Can Longitudinal Healthcare Databases Detect Safety Signals? A View from the EU-ADR Project

Background: Withdrawal of drugs from the market due to safety issues continues to underscore the need for more effective postmarketing safety evaluation.Longitudinal electronic healthcare record (EHR) databases containing administrative claims and medical records have emerged as an important complementary source of information for drug safety signal detection. Objectives: To evaluate how many and what types of drugs may be investigated for signal detection in EHR databases as a function of actual drug use,minimal detectable relative risk (RR),and background rates of adverse events. Methods: We extracted and analyzed data on prescribed/ dispensed drugs among individuals registered within the EU-ADR database network for the period 1996–2009.Drug use was assessed as number of personyears(PYs) of exposure by ATC code.Using empirically-derived background incidence rates(IR) of frequent and rare events in EUADR we estimated how much drug exposure would have to be present in the databases to allow for signal detection with certain power (b¼0.2,¼0.05) across varying magnitudes of RR. Results: A total of 2289 drugs were used by 19647452 individuals during 59594132 PYs of follow-up. OTC drugs, drugs used in-hospital, and nonreimbursable drugs were not captured in the databases.The number of drugs that can be investigated to detect a weak association (RR¼ 2.0), if present, for a frequent event like acute myocardial infarction (AMI) and for a rare event like rhabdomyolysis (RHABD) using currently available data are 531(23%) and 19(0.8%), respectively. As strength of association increased,number of drugs that may be investigated also increased. If current size of EU-ADR would be multiplied by 10, assuming same drug use patterns, the maximum percentage of drugs that can be investigated even for a strong association (RR¼6) is only 70% for AMI and 31% for RHABD. Conclusions: Mining of EHR databases may allow for signal detection,although power to do so may be low for infrequently used drugs,even if size of the network is increased.The greatest potential of these systems is in the detection of events that have a high or moderate background incidence in the general population.