Identifying signals of potentially harmful medications in pregnancy: use of the double false discovery rate method to adjust for multiple testing

Abstract

Aims Surveillance of medication use in pregnancy is essential to identify associations between first trimester medications and congenital anomalies (CAs). Medications in the same Anatomical Therapeutic Chemical classes may have similar effects. We aimed to use this information to improve the detection of potential teratogens in CA surveillance data. Methods Data on 15 058 malformed fetuses with first trimester medication exposures from 1995–2011 were available from EUROmediCAT, a network of European CA registries. For each medication‐CA combination, the proportion of the CA in fetuses with the medication was compared to the proportion of the CA in all other fetuses in the dataset. The Australian classification system was used to identify high‐risk medications in order to compare two methods of controlling the false discovery rate (FDR): a single FDR applied across all combinations, and a double FDR incorporating groupings of medications. Results There were 28 765 potential combinations (523 medications × 55 CAs) for analysis. An FDR cut‐off of 50% resulted in a reasonable effective workload, for which single FDR gave rise to eight medication signals (three high‐risk medications) and double FDR 50% identified 16 signals (six high‐risk). Over a range of FDR cut‐offs, double FDR identified more high‐risk medications as signals, for comparable effective workloads. Conclusions The double FDR method appears to improve the detection of potential teratogens in comparison to the single FDR, while maintaining a low risk of false positives. Use of double FDR is recommended in routine signal detection analyses of CA data.

Bayesian hierarchical methods in the detection of potentially teratogenic first‐trimester medications

“…Foetuses were excluded if exposed only to topical medications, codes with less than five digits (detail below ATC4 level; n = 1219), second/third trimester medications (n = 1490), or with unknown timing (n = 12 073). Further description of these exclusions by registry have been described in more detail in our previous study . After these exclusions, a total of 15 058 malformed foetuses were included in the analysis dataset.…”

“…Five hundred twenty‐three medications with at least three exposed foetuses were investigated. As in the previous EUROmediCAT analysis, foetuses exposed only to medications with fewer than three exposures in total were included in the dataset as controls (since medications with fewer than three exposures are not analysed for signal detection) …”

“…Further description of these exclusions by registry have been described in more detail in our previous study. 3 After these exclusions, a total of 15 058 malformed foetuses were included in the analysis dataset. ATC5 codes were analysed, but where information was only available to ATC4 level, this was also included.…”

“…Routine screening of congenital anomaly (CA) data for potentially teratogenic medications taken during the first trimester of pregnancy is performed by EUROmediCAT . EUROmediCAT compares the odds of exposure to a specific medication for a specific CA with the odds of exposure to the same medication for all other medication‐exposed CAs using a double false discovery rate (FDR) procedure to incorporate groupings of similar medications when adjusting for multiple testing …”

“…1 EUROmediCAT compares the odds of exposure to a specific medication for a specific CA with the odds of exposure to the same medication for all other medication-exposed CAs 2 using a double false discovery rate (FDR) procedure to incorporate groupings of similar medications when adjusting for multiple testing. 3 In the field of pharmacovigilance, large spontaneous reporting databases detect signals of adverse events (AEs) using disproportionality analyses to identify drug-AE combinations with more observed reports than expected under the assumption of independence. 4 Bayesian hierarchical models (BHMs) have been applied to these spontaneous reporting datasets, 5 with the aim of improving estimated associations for any drug-AE combination by incorporating information from other similar drugs (or AEs) using specified groupings.…”