Application of new signal detection methods to EUROmediCAT data: Bayesian Confidence Propagation Neural Network-Biclustering (BCPNN-Bic) and BCPNN-Bicluster-Bayesian Hierarchical Modelling (BBBHM)

Background: Signal detection statistical methods are commonly used in spontaneous reporting datasets for detection of potentially harmful drug-reaction associations. New methods tailored to smaller datasets have been shown to be effective in simulation studies, however these have yet to be applied to real world data. EUROmediCAT is a network of population-based case-malformed registries that provide a post-marketing surveillance system for medications taken during pregnancy. The performance of these new methods will be evaluated in the EUROmediCAT data.

Objectives: To compare signals from the new Bayesian Confidence Propagation Neural Network-Biclustering (BCPNN-Bic) and Bayesian Confidence Propagation Neural Network-Bicluster-Bayesian Hierarchical Modelling (BBBHM) methods, with those of other signal detection methods when applied to data from a network of population-based congenital anomaly registries with information on medications taken during pregnancy.

Methods: Data were from 14 European EUROmediCAT member registries covering births from 2005-2019, with a total birth population of 2.7 million. Potential harmful drug-anomaly associations were investigated for 84 EUROCAT anomaly subgroups to WHO-ATC medication groups. This was repeated with increasing specificity of the medication from 3, 4, 5 and 7-digits. Six methods were compared: case-malformed incidence ratios; Sequential probability ratio testing (SPRT); Bayesian Confidence Propagation Neural Networks (BCPNN); and Du Mouchel’s Gamma Poisson Shrinker (GPS); and the new approaches, BCPNN-Bic and BBBHM.

Results: Preliminary results showed the BCPNN-Bicluster, incidence ratios and BCPNN identified the most signals, and GPS and SPRT the fewest. Most signals were identified by at least two methods, however for the 4, 5 and 7-digit analyses no signals were identified by all methods. Numbers of signals identified by the SPRT and GPS decreased as the specificity of the medication increased. This is likely due to decreasing numbers of cases and controls per drug-anomaly pair, which corresponds to the reduction in power with smaller sample sizes observed in simulations. For the incidence ratios, BCPNN, BCPNN-Bicluster and BBBHM the number of signals increased as the specificity of the medication increased.

Conclusions: The GPS and SPRT are not capable of identifying signals of potential harm at the 5 and 7-digit level of the WHO-ATC in this dataset. As such, the GPS and SPRT should not be used for identification of harmful medications in the EUROmediCAT data, or similar settings. As the BBBHM can detect signals at the 5 and 7-digit level, and simulations suggest it has the lowest false positive rate, it may be the preferred method for identifying harmful drugs within the EUROmediCAT data.