(194) Ontology-based Semantic Similarity Measures in Drug Safety: A Literature Review

Background: In pharmacovigilance (PV), adverse event reports are often coded using MedDRA Preferred Terms (PT). However, individual PTs may not adequately capture related events within the same medical context. While customized groupings by medical experts like MedDRA SMQs exist, they oversimplify the relationship between PTs and offer only a binary inclusion or not. Ontology-based semantic similarity measures (SSMs) offer a potential solution to the challenge of defining outcomes, especially in the context of retrieving information from multiple safety data sources. Despite this potential, literature summarizing the use of SSMs in PV remains scarce.

Objectives: This review aims to explore how ontology-based SSMs have been applied in PV and the broader biomedical field, with a focus on identifying opportunities for their use in PV.

Methods: We conducted a literature review using EMBASE, with end date February 1, 2024. The initial search employed keywords "semantic similarity" OR "semantic distance," yielding 1,503 hits. Subsequent refinements included terms related to (1) “MedDRA” or “WHO-ART”, (2) “safety” or “pharmacovigilance”, and (3) “ontology” and “medical concepts” while excluding gene-related studies.

Results: Fifty papers met the inclusion criteria for this review. Ontology-based SSMs were categorized into four main types: path-based, corpus information content (IC)-based, intrinsic IC-based, and vector-based. Corpus IC and vector-based SSMs require a corpus of medical texts for computation, while intrinsic IC-based SSMs rely solely on ontologies, making them easily transferable across different data sources. These intrinsic IC-based SSMs emerged in literature after 2011 (Sanchez, 2011). Refinements based on specific terms yielded relevant publications, although some focused on gene ontology or were unrelated to PV. The most used ontologies were MeSH, SNOMED-CT, and UMLS. Evaluations typically utilized the reference set of similar medical terms proposed by Pedersen (2007), although some studies employed larger sets. Few evaluations of SSM types beyond intrinsic IC-based SSMs were found, and while some showed promising performance, they were not conclusive, especially within the realm of drug safety.

Conclusions: The literature on ontology-based SSMs in drug safety is limited, with only a few applications published, primarily focused on clustering medical concepts. Although various ontology-based SSMs have been described, their evaluation remains sparse. However, these measures hold promise for applications in drug safety, including outcome definition and information retrieval from diverse data sources.