Hierarchy-aware Adverse Reaction Embeddings for Signal Detection
Post-market drug surveillance monitors new and evolving treatments for their effectiveness and safety in real-world conditions. A large amount of drug safety surveillance data is captured by spontaneous reporting systems such as the FAERS. Developing automated methods to identify actionable safety signals from these databases is an active area of research. In this paper, we propose two novel network representation learning methods (HARE and T-HARE) for signal detection that jointly utilize association information between drugs and medical outcomes from the FAERS and ancestral information in medical ontologies. We evaluate these methods using two publicly available reference datasets, EU-ADR and OMOP corpus. Experimental results showed that the proposed methods significantly outperformed standard methodologies based on disproportionality metrics and the existing state-of-the-art aer2vec method with statistically significant improvements on both EU-ADR and OMOP datasets. Through quantitative and qualitative analysis, we demonstrate the potential of the proposed methods for effective signal detection.