P236: Hierarchical Deep Learning for Proactive Safety Signal Detection in Pharmacovigilance
Poster Presenter
Linghui Li
Associate Director
Takeda United States
Objectives
To develop and evaluate a hierarchical deep learning time series framework (deepVARHierarchical) that enhances safety signal detection by modeling temporal dynamics and cross-event correlations across the MedDRA hierarchy and Standardised MedDRA Queries (SMQs).
Method
A DeepVAR-based recurrent neural network models hierarchical AE time series across MedDRA levels (SOC, HLGT, HLT, PT) and SMQs. Using an aggregation matrix, normalized AE counts are jointly forecast, and signals are flagged when observed counts exceed the upper bound of the model-derived 95% predict
Results
Using multi-year global AE data at monthly and quarterly resolutions, the deepVARHierarchical model was benchmarked against traditional disproportionality methods such as Proportional Reporting Ratio (PRR), Reporting Odds Ratio (ROR) and Information Component (IC) and Empirical Bayes Geometric Mean (EBGM).
Key Performance Outcomes
• >20% improvement in early detection sensitivity for emerging safety signals compared with PT-level disproportionality methods
• >30% efficiency gain in signal review workflows through prioritized, model-driven alerting
• >20% reduction in data processing and manual review burden, enabling pharmacovigilance-led, AI-assisted decision support
• Stable and scalable performance across >1,000 concurrent AE time series
The model learned temporal dynamics—including trends, seasonality, and cross-AE correlations—beyond the capability of cross-sectional count-based methods. Importantly, clinically meaningful signals emerged at SMQ and SOC levels, even when individual PT counts did not exceed traditional statistical thresholds.
Hierarchical consistency constraints ensured alignment between PT-, SMQ-, and SOC-level forecasts, improving interpretability and regulatory transparency. In retrospective validation, several safety patterns were identified months earlier than conventional disproportionality approaches.
The architecture scales efficiently to high-dimensional global safety datasets and supports automated re-training, enabling continuous and proactive signal monitoring.
Conclusion
The deepVARHierarchical framework represents a significant advancement in AI-driven pharmacovigilance. By integrating temporal forecasting, cross-event correlation learning, and hierarchical aggregation, this approach overcomes key limitations of traditional disproportionality analyses, particularly in sparse or evolving data environments.
Unlike static, count-based methods, the model provides proactive anomaly detection grounded in expected temporal behavior, enabling earlier and more coherent signal identification across the MedDRA hierarchy.
This scalable, automated deep-learning framework lays the foundation for next-generation signal intelligence systems. Ongoing work includes integration of additional real-world data sources, refinement of anomaly thresholds, and expansion across therapeutic areas to further validate generalizability and regulatory relevance.
