(P-618) Using Structural Topic Modelling to Reveal Patterns in Reports on Opioid Drugs in a Pharmacovigilance Database

Background: Adverse drug reaction (ADR) reports in pharmacovigilance databases usually contain coded information and a large amount of unstructured or semi-structured information in plain text format. The quantity and format of these data often render them neglected, however structural topic modelling (STM) represents a potentially insightful way of harnessing these rich data and to detect grouping or themes in spontaneous reports to aid signal detection.

Objectives: An explorative study of the potential for structural topic modelling to identify useful patterns in ADR reports involving opioid drugs in a pharmacovigilance database.

Methods: A dataset of ADR reports on opioid drugs reported to the Netherlands Pharmacovigilance Centre Lareb from 1991 to December 2020 was used, comprising a total of 3069 unique reports. Qualitative text analysis was combined with structural topic modelling, an automated text analysis method, to examine this existing but untapped data source. STM was employed to a) identify various topics in the reports and b) investigate topic prevalence as a function of covariates such as date of receipt, source of report (industry/Lareb) and type of drug. A detailed qualitative analysis of each topic resulting from the topic model allowed an evaluation of the quality of the model and the detection of more subtle nuances and sentiments that might otherwise have been overlooked.

Results: 11 meaningful topics were identified in reports submitted directly via the Lareb website and 12 topics in reports received from industry. Reports relating to “patches and skin”, “pain and various symptoms” and the “combination of tramadol and paracetamol”, for example, were clearly present in both. In contrast, reports addressing “pregnancy and birth” and “overdose and death” featured only in industry reports, with many more insights into patient experiences in the Lareb reports. STM also illuminated the relationship between the prevalence of a topic and specific opioids.:E.g. reports dealing with respiratory depression have a significantly (p < .001) higher prevalence for morphine. Finally, many of the standard text formulations in Lareb reports were a hindrance to grasping the underlying content.

Conclusions: STM, especially in iterative combination with qualitative text analysis, represents an innovative and promising approach to analysing and utilizing large amounts of unstructured ADR text data in a comprehensive and nuanced way. The identified topics reflect findings in the literature on opioids, helping validate our findings and use of using STM for the analysis of pharmacovigilance data.