A data-driven method to detect adverse drug events from prescription data

Abstract

Drug safety issues such as Adverse Drug Events (ADEs) can cause serious consequences for the public. The clinical trials that are undertaken to assess medicine efficacy and safety prior to marketing, generally, may provide sufficient samples for discovering common ADEs. However, more samples are needed to detect infrequent and rare events. Additionally, clinical trials may not include all subgroups of patients. For these reasons, post-marketing surveillance of medicines is necessary for identifying drug safety issues. Most regulatory agencies use the Spontaneous Reporting Systems to identify associations between medicines and suspected ADEs. Data mining with effective analytical frameworks and large-scale medical data is potentially an alternative method to discover and monitor ADEs. In the present paper, we aim to detect potential ADEs from prescription data by discovering ADE associated prescription sequences. In an ADE associated prescription sequence 〈Dp→Ds〉, the prior medicine Dp leads to an ADE for which the succeeding medicine Ds is dispensed to treat. We propose a data-driven method which integrates (1) a constrained sequential pattern mining to uncover prescription sequences as potential signals of ADEs, (2) domain constraints to eliminate interference signals and (3) an adapted Self-Controlled Case Series model to evaluate the potential signals of ADEs. Despite ample prior works using Electronic Health Records (EHRs), our method utilises pure prescription data which does not contain additional information, e.g. symptoms or diagnoses as included in EHRs. To assess the performance of the proposed method, we apply it to a real-world dataset from the Pharmaceutical Benefits Scheme of Australia. The dataset contains over 50 million records covering approximately 2 million patients. The results demonstrate the effectiveness of our method in identifying both known ADEs and unknown yet suspicious ADEs with limited detection of false positive signals. Comparing to a recognised gold standard, our method successfully detects 67.4% of the positive adverse events while only 8.78% false positives exist.