Evaluating Knowledge Fusion Models on Detecting Adverse Drug Events in Text

Background Detecting adverse drug events (ADE) of drugs that are already available on the market is an essential part of the pharmacovigilance work conducted by both medical regulatory bodies and the pharmaceutical industry. Concerns regarding drug safety and economic interests serve as motivating factors for the efforts to identify ADEs. Hereby, social media platforms play an important role as a valuable source of reports on ADEs, particularly through collecting posts discussing adverse events associated with specific drugs.Methodology We aim with our study to assess the effectiveness of knowledge fusion approaches in combination with transformer-based NLP models to extract ADE mentions from diverse datasets, for instance, texts from Twitter, websites like askapatient.com, and drug labels. The extraction task is formulated as a named entity recognition (NER) problem. The proposed methodology involves applying fusion learning methods to enhance the performance of transformer-based language models with additional contextual knowledge from ontologies or knowledge graphs. Additionally, the study introduces a multi-modal architecture that combines transformer-based language models with graph attention networks (GAT) to identify ADE spans in textual data.Results A multi-modality model consisting of the ERNIE model with knowledge on drugs reached an F1-score of 71.84% on CADEC corpus. Additionally, a combination of a graph attention network with BERT resulted in an F1-score of 65.16% on SMM4H corpus. Impressively, the same model achieved an F1-score of 72.50% on the PSYTAR corpus, 79.54% on the ADE corpus, and 94.15% on the TAC corpus. Except for the CADEC corpus, the knowledge fusion models consistently outperformed the baseline model, BERT.Conclusion Our study demonstrates the significance of context knowledge in improving the performance of knowledge fusion models for detecting ADEs from various types of textual data.Author Summary Adverse Drug Events (ADEs) are one of the main aspects of drug safety and play an important role during all phases of drug development, including post-marketing pharmacovigilance. Negative experiences with medications are frequently reported in textual form by individuals themselves through official reporting systems or social media posts, as well as by doctors in their medical notes. Automated extraction of ADEs allows us to identify these in large amounts of text as they are produced every day on various platforms. The text sources vary highly in structure and the type of language included which imposes certain challenges on extraction systems. This work investigates to which extent knowledge fusion models may overcome these challenges by fusing structured knowledge coming from ontologies with language models such as BERT. This is of great interest since the scientific community provides highly curated resources in the form of ontologies that can be utilized for tasks such as extracting ADEs from texts.### Competing Interest StatementThe authors have declared no competing interest.### Funding StatementThe author(s) received no specific funding for this work.### Author DeclarationsI confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.YesThe details of the IRB/oversight body that provided approval or exemption for the research described are given below:N/AI confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.YesI understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).YesI have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.YesThe data underlying the results presented in the study are available from various study owners. These studies have been linked in the main manuscript.