Disproportionality Analysis Signal Detection Using the New ‘Information Component–Bicluster–Bayesian Hierarchical Model’

INTRODUCTION: Disproportionality signal detection methods can help identify potential adverse outcomes of medication exposure. Such methods scan all medication-outcome pairs within a dataset to identify possible associations requiring further investigation. Many signal detection methods use separate statistical tests for each pair, however, information sharing between groups of 'similar' medications or outcomes may produce more sensitive and specific detection of harmful effects. Data-defined groups identified through clustering may provide a better reflection of dependencies within the data. OBJECTIVE: We present a new signal detection method that combines the information component, biclustering and Bayesian hierarchical modelling (IC-Bic-BHM) to identify groups within the data and use these groups within the analysis. METHODS: We use simulations to demonstrate the importance of using groupings that accurately reflect dependency within the data; and compare sensitivity, specificity, positive predictive value and negative predictive value of the IC-Bic-BHM with other methods. Results of application to 536 medications and 61 types of congenital anomaly in babies/foetuses from 87,513 EUROmediCAT pregnancies were also compared. RESULTS: Sensitivity and specificity of Bayesian hierarchical models are dependent upon how accurately groupings reflect dependency within the data. The IC-Bic-BHM is the first method to use data-defined dependencies between medications and outcomes in signal detection. Simulations suggest it performs well in datasets with large numbers of medication-outcome pairs but requires adequate power to detect groups. The IC-Bic-BHM identified signals involving several known teratogenic medications within EUROmediCAT data. CONCLUSIONS: This study presents a novel approach to signal detection but requires further investigation before adoption.