Hierarchy-aware Adverse Reaction Embeddings for Signal Detection

Area of Science:

• Pharmacovigilance and Drug Safety
• Artificial Intelligence in Healthcare
• Network Science

Background:

• Post-market drug surveillance is crucial for monitoring treatment safety and effectiveness in real-world settings.
• Spontaneous reporting systems like the FDA Adverse Event Reporting System (FAERS) generate vast amounts of drug safety data.
• Automated methods for identifying actionable safety signals from large databases are an active research area.

Purpose of the Study:

• To propose novel network representation learning methods, HARE and T-HARE, for enhanced drug safety signal detection.
• To leverage association information between drugs and medical outcomes from FAERS.
• To integrate ancestral information from medical ontologies into signal detection models.

Main Methods:

• Developed two novel network representation learning methods: HARE and T-HARE.
• Utilized drug-outcome association data from the FAERS database.
• Incorporated medical ontology data to enrich network representations.
• Evaluated performance on the EU-ADR and OMOP corpus reference datasets.

Main Results:

• The proposed HARE and T-HARE methods significantly outperformed standard disproportionality metrics.
• Statistically significant improvements were observed compared to the state-of-the-art aer2vec method.
• Both methods demonstrated superior performance on the EU-ADR and OMOP datasets.
• Quantitative and qualitative analyses confirmed the effectiveness of the proposed methods.

Conclusions:

• HARE and T-HARE offer a powerful approach for automated drug safety signal detection.
• Network representation learning effectively integrates diverse data sources for improved pharmacovigilance.
• These methods hold significant potential for enhancing the safety monitoring of pharmaceutical products.