Identification of risk factors for adverse drug reactions in a pharmacovigilance database

INTRODUCTION: In addition to identifying new safety signals, pharmacovigilance databases could be used to identify potential risk factors for adverse drug reactions (ADRs). OBJECTIVE: To evaluate whether data mining in a pharmacovigilance database can be used to identify known and possible novel risk factors for ADRs, for use in pharmacovigilance practice. METHOD: Exploratory data mining was performed within the Swedish national database of spontaneously reported ADRs. Bleeding associated with direct oral anticoagulants (DOACs)-rivaroxaban, apixaban, edoxaban, and dabigatran-was used as a test model. We compared demographics, drug treatment, and clinical features between cases with bleeding (N = 965) and controls who had experienced other serious ADRs to DOACs (N = 511). Statistical analysis was performed by unadjusted and age adjusted logistic regression models, and the random forest based machine-learning method Boruta. RESULTS: In the logistic regression, 13 factors were significantly more common among cases of bleeding compared with controls. Eleven were labelled or previously proposed risk factors. Cardiac arrhythmia (e.g., atrial fibrillation), hypertension, mental impairment disorders (e.g., dementia), renal and urinary tract procedures, gastrointestinal ulceration and perforation, and interacting drugs remained significant after adjustment for age. In the Boruta analysis, high age, arrhythmia, hypertension, cardiac failure, thromboembolism, and pharmacodynamically interacting drugs had a larger than random association with the outcome. High age, cardiac arrhythmia, hypertension, cardiac failure, and pharmacodynamically interacting drugs had odds ratios for bleeding above one, while thromboembolism had an odds ratio below one. CONCLUSIONS: We demonstrated that data mining within a pharmacovigilance database identifies known risk factors for DOAC bleeding, and potential risk factors such as dementia and atrial fibrillation. We propose that the method could be used in pharmacovigilance for identification of potential ADR risk factors that merit further evaluation.