P220: Bayesian Confidence Propagation Neural Network for Signal Detection: Comparing Traditional Disproportionality Methods
Poster Presenter
Alice Hsu
TA Safety
Amgen Inc. United States
Objectives
To evaluate the performance of Bayesian Confidence Propagation Neural Network (BCPNN) against established pharmacovigilance signal detection methods (ROR, EBGM, IC) using FAERS data, focusing on precision and false positive reduction.
Method
Analyzed 1814 drug-event combinations from FAERS through September 2023 using four methods: ROR05 (threshold >2), EBGM (threshold >2), IC (threshold >1), and BCPNN (IC>1), all with N=3. Used R's PhVid package with comprehensive SDR adjudication.
Results
ROR05 identified 149 SDRs, serving as the reference. EBGM detected 127 SDRs, reducing false positivity by 15%. IC05 identified 143 SDRs with a 4% false detection rate. BCPNN identified 93 SDRs, reducing the false detection rate by 38% compared to ROR05. All SDRs detected by EBGM, IC, and BCPNN were subsets of the 149 SDRs identified by ROR05. Analysis of these 1814 drug-event combinations revealed that alternative methods (EBGM, IC, and BCPNN) improved both precision and specificity compared to the reference method. The adjudication process categorized findings into two main groups: known adverse events already documented in product labeling and false signals attributed to indication and clinical confounding. Importantly, comprehensive evaluation confirmed that no SDRs represented new potentially causal associations requiring further investigation, validating the improved precision of newer methods, particularly BCPNN.
Conclusion
BCPNN demonstrated benefits in reducing false positives while maintaining detection of clinically relevant associations. Its advantages include efficient processing of large datasets through neural network architecture, certainty measures for detected associations, and adaptability to new data using Bayesian statistics. The method's ability to analyze all drug-event pairs simultaneously enables detection of complex patterns that might be missed by simpler approaches. However, challenges remain, including the need for high-quality data, computational demands for complex analyses, and careful threshold setting to balance sensitivity and specificity. While promising, BCPNN's adoption should be approached thoughtfully, with organizations conducting thorough testing and tailoring the method to their specific needs. Further research is needed to refine BCPNN, enhance interpretability, explore its potential across various pharmacovigilance settings, and address challenges in detecting signals for rare events and newly introduced drugs.
NOTE: The views expressed herein represent those of the author and do not necessarily represent the views or practices of the author's employer or any other party.
