Adaptation of Bayesian data mining algorithms to longitudinal claims data: coxib safety as an example

Area of Science:

• Pharmacovigilance
• Health Informatics
• Biostatistics

Background:

• Bayesian data mining methods are established for drug safety signal detection in adverse event systems.
• Adaptation of these methods for longitudinal administrative claims data remains unevaluated.

Purpose of the Study:

• To assess the feasibility of applying Bayesian data mining, specifically the empirical Bayes MGPS algorithm, to longitudinal administrative claims data.
• To validate the data mining approach by comparing its results to traditional epidemiologic methods.

Main Methods:

• Utilized the Medicare Current Beneficiary Survey to form a cohort exposed to coxibs or NS-NSAIDs (1999-2003).
• Applied the empirical Bayes MGPS algorithm to simultaneously analyze 259 outcomes for coxib vs. NS-NSAID users, adjusting for covariates.
• Conducted a parallel analysis using traditional methods for acute myocardial infarction to establish concurrent validity.

Main Results:

• Empirical Bayes MGPS identified an association between celecoxib and acute myocardial infarction (GBM ratio 1.91).
• Rofecoxib use was linked to acute cerebrovascular events (GBM ratio 1.85) and other diagnoses potentially indicating drug use.
• Data mining results aligned with traditional epidemiologic analyses, confirming validity.

Conclusions:

• Bayesian data mining methods show utility for drug safety evaluations using administrative claims data.
• Further research is necessary to generalize these findings to diverse drug exposures and claims databases.