Semantic Similarity-Informed Bayesian Borrowing for Quantitative Signal Detection of Adverse Events

We present a Bayesian dynamic borrowing approach to enhance quantitative identification of adverse events in spontaneous reporting systems. The method embeds a robust meta-analytic predictive prior and incorporates semantic similarity measures (SSMs) to enable weighted information sharing from clinically similar MedDRA preferred terms (PTs) to the target PT. This continuous similarity-based borrowing overcomes limitations of rigid hierarchical grouping in disproportionality analysis (DPA). Using data from the FDA Adverse Event Reporting System and a time-stamped reference set, we evaluate our approach-termed IC SSM-against the current standard approach of "traditional" information component (IC) analysis and borrowing at the MedDRA HLGT level (IC HLGT). The primary IC SSM approach at the standard threshold (2.5% posterior quantile $>0$ ) detected true signals over five months earlier than traditional IC. It showed higher sensitivity (57.0%) but lower precision (14.5%) than traditional IC (50.1%, 16.1%), while outperforming IC HLGT (55.7%, 13.9%). In subsequent analyses, even when tightening the alerting quantile of IC SSM to 1% to improve precision (16.4%), superior sensitivity (51.4%) was retained compared to traditional IC. Similarly, relaxing traditional IC's quantile to 5% yielded only 55.6% sensitivity with 14.2% precision, failing to match IC SSM's performance or timeliness and confirming its inherent advantage. IC SSM's superior timeliness and sensitivity stem from integrating clinically meaningful information, making it a scalable and context-aware enhancement to traditional DPA. The method has potential for exploratory analysis in sparse datasets, and refinements to the borrowing strategies and similarity metrics could yield further performance improvements.