In this section: Signals • Collaborations • Methods
The detection of medicines safety signals is a core activity of UMC, and over the year 20 signals were published in VigiLyze for use by national pharmacovigilance centres. These included several earlier signals that had been kept under review for additional data, as well as one of bradycardia with fluorouracil from the Latin America signal detection workshop in 2019. Signal assessments from selected national centres, including Netherlands and Eritrea, were also shared via VigiLyze.
Birgitta Grundmark gave a presentation on 'Strengthening the capacity for signal detection and management for pharmacovigilance national centres' in Bogotá, Colombia at the annual ISoP meeting
UMC produced four scientific publications informing the practical signal detection work. One evaluated the robustness of disproportionality analysis in smaller databases, which indicated that disproportionality analysis is computationally robust for countries with as few as 500 reports, whereas for data subsets in general, a minimum of 3,000 to 5,000 reports may be recommended. This was followed by the publication of a guidebook, “Signal detection for national pharmacovigilance centres with small data sets”.
Another paper described the development and evaluation of methods to detect signals of risk groups for adverse drug reactions in VigiBase. Seven signals describing previously unrecognised potential risk groups were communicated, related to elderly, male, female, underweight, obese and patients from Asia.
A manuscript which presents the theoretical basis for the method and includes an assessment of the stability and the clinical coherence of the identified patterns was submitted for publication, and subsequently published in Drug Safety in July 2020.
A paper was published about an initiative with the Netherlands Pharmacovigilance Centre Lareb and the Dutch thyroid patient organisation Schildklier Organisatie Nederland (SON) to communicate a signal of panic attacks with levothyroxine directly to patients. This generated considerable engagement resulting in additional patient experiences being received, further strengthening the original signal. Finally, a paper was accepted for publication that described development and evaluation of methods to detect signals of drug-drug interactions in VigiBase. Method development is a central part of UMC’s research work: questioning how signal detection may best be performed, testing the algorithms we apply, pushing the boundaries of how data is used. A systematic evaluation of UMC’s adverse event cluster analysis algorithm was completed. Previously derived vector representations for drugs and adverse events were made available for internal use and evaluation. They may offer an approach to identify semantically similar drugs and adverse events based on reporting patterns in VigiBase.
During the year an algorithm was evolved for identifying reports with outlying doses, which uses the overall reported doses of the same drug in VigiBase as reference. A signal detection sprint with the WHO Collaborating Centre in Rabat, Morocco, then explored its use to detect reports related to possible medication errors not explicitly coded as such. It found that there was rarely enough information in narratives or the structured information of such reports to support the required assessment – possibly because most reports where the reporter has reflected on an outlying dose related to a medication error have been coded as such.
Signal detection for drug interactions is more challenging than pairwise drug-adverse reaction monitoring. There’s always a chance that the observed adverse reaction is due to one of the drugs and not to a drug interaction, and we need more detailed reports to assess this possibility.
Last August we published 'A feasibility study of drug-drug interaction signal detection in regular pharmacovigilance' in Drug Safety. The article describes a study initiated in 2016, with the aim of finding drug interaction signals in VigiBase by using a previously proposed statistical method by UMC. The study resulted in three signals and it showed that statistical signal detection for drug interactions is possible.
I really hope we can develop our methods even further in the future and increase knowledge of drug combinations that should be avoided.
Sara Hult
Pharmacovigilance Scientist
Pharmacovigilance Scientist
An evaluation of Japanese reporting patterns in VigiBase highlighted a higher proportion of well-documented reports, reports submitted by physicians, and reports with just a single adverse event term as key features. A paper was accepted for publication that assesses the impact of regulatory actions and reporting patterns of interstitial lung disease in Japan. A Letter to the Editor in World Psychiatry cited analyses of VigiBase to better describe and understand the impact of pneumonia as a suspected adverse reaction to clozapine. Another UMC paper described the use of VigiBase combined with longitudinal observational health data in the assessment of a signal for colitis with nintedanib.
UMC’s Research team forge strong collaborations with individuals and groups internationally to further the science of pharmacovigilance. The completed IMI WEB-RADR project, which sought to utilise and evaluate social media and mobile reporting technologies for pharmacovigilance purposes, produced three further papers: the first the evaluation of adverse event recognition in Twitter and Facebook, exposing the difficulties in implementing automatic adverse event recognition in Twitter, and warning of discrepancies between published performance evaluations and observed results for independent data; the second a benchmark performance evaluation reference set for adverse event recognition in Twitter; and the third the project’s final recommendations for use of social media in pharmacovigilance.
The Consortium for Advanced Research Training in Africa (CARTA), of which UMC is a member, pioneered a new concept for stimulating and funding joint research in pharmacovigilance involving both academia and national pharmacovigilance centres in Africa which yielded several applications. UMC also contributed to 13 training initiatives in Africa, Latin America, Asia, and the Mediterranean region, including causality assessment and signal detection, using the tools VigiFlow and VigiLyze.
Jennifer Wall and other UMC staff provided training on VigiFlow and VigiLyze at the national centres meeting in Bogotá
The collaboration with the UK’s Medicines and Healthcare products Regulatory Agency (MHRA) to develop and evaluate methods for de-identifying free text case narratives intensified: a study protocol was approved; data from the Yellow Card Scheme obtained; and algorithms consolidated in preparation for the next phase of the study later in 2020. A pilot study was completed which focused on patients’ expectations on pharmacovigilance and whether information in VigiBase can provide answers to some of their questions.
Four MSc thesis projects were carried out at UMC, by students from universities in Uppsala and Stockholm. These were “Retrospective disproportionality analysis to investigate the usefulness of the WHODrug Standardised Drug Groupings in the pharmacovigilance signal detection process”, “Extracting adverse drug reactions from product labels using deep learning and natural language processing”, “Improving the speed and quality of an adverse event cluster analysis with stepwise expectation maximization and community detection”, and “Normalization of adverse event verbatims to MedDRA”.
UMC portfolio of scientific methods
vigiRank: Predictive model that ranks pharmacovigilance safety signals according to multiple aspects of strength of evidence.
vigiMatch: Probabilistic record-matching method to detect unexpectedly similar pairs of records in a database.
vigiGrade: Multidimensional measure of data quality in pharmacovigilance (completeness, relevance, consistency, etc.).
vigiPoint: Algorithm to pinpoint the key features of a subset of database records in contrast to a broader set.
vigiTrace: Suite of analytics methods for the analysis of longitudinal event history data, including chronographs for statistical graphical overviews and the calibrated self-controlled cohort design for temporal screening.
