Hierarchical Models for Multiple, Rare Outcomes Using Massive Observational Healthcare Databases.

Area of Science:

• Pharmacovigilance and pharmacoepidemiology
• Statistical modeling in healthcare
• Computational biology and bioinformatics

Background:

• Clinical trials have limited power to detect rare adverse drug events (ADEs).
• Existing Bayesian self-controlled case series (BSCCS) models do not account for shared pathology across multiple ADEs.
• Large patient claims and electronic health record databases offer potential for post-approval ADE detection.

Purpose of the Study:

• To develop an improved BSCCS model that integrates a pathology hierarchy to account for shared underlying biological mechanisms across multiple ADEs.
• To address the challenge of increased model dimensionality introduced by considering shared pathology.
• To demonstrate the feasibility and benefits of analyzing multiple ADE outcomes simultaneously at scale.

Main Methods:

• Developed a novel informative hierarchical prior to link outcome-specific effects within the BSCCS framework.
• Created an efficient method to reduce model dimensionality, making the hierarchical model compatible with existing computational tools.
• Validated the approach using synthetic data with varying risk and prevalence, and real-world data from the MarketScan Lab Results dataset.

Main Results:

• Demonstrated decreased bias in drug risk estimates when accounting for shared pathology compared to traditional methods.
• Exposed bias resulting from outcome aggregation in previous analyses of warfarin and dabigatran for intracranial hemorrhage and gastrointestinal bleeding.
• Confirmed the approach's effectiveness even when analyzing extremely rare conditions.

Conclusions:

• Simultaneous analysis of multiple ADE outcomes using a hierarchical BSCCS model is feasible and beneficial for improving drug safety surveillance.
• Integrating pathology hierarchies enhances the accuracy of ADE risk estimation, particularly for rare events.
• The developed method offers a scalable solution for leveraging large healthcare databases to detect and understand adverse drug events.