(138) An Application of Bayesian Modelling for Information Sharing in Safety Signal Detection

Background: The multi-item gamma Poisson shrinker (MGPS) is a quantitative approach used routinely for safety signal detection in spontaneous reporting systems. It highlights statistical alerts for clinical review as potential signals based on Bayesian shrinkage of an O/E ratio, H0 being independence between drug and adverse events (AE). MGPS treats AE MedDRA preferred terms (PT) as independent not allowing for clinical similarity of AEs leading to potential false negative alerts (dilution) and false positive spurious findings. Berry and Berry[1] addressed this challenge in clinical trial data with a three-level hierarchical mixture model, borrowing strength of association across MedDRA system organ classes (SOC).

Objectives: A first application of Bayesian hierarchical mixture model (HMM) on FDA adverse event reporting system (FAERS) with comparison to MGPS.

Methods: We tested the HMM on FAERS, cumulative from 2004Q1 to 2009Q4, to detect AEs, defined at the PT level, associated with a test case, Zolpidem. Due to computational complexity, the data was restricted to this period which still provides a representative, historical lookback of data. Several implementations of the HMM were examined. Herein, we present borrowing at the MedDRA high level group term, with normal prior on the logarithm of odds ratio (OR). An alert is defined as P5>1 of OR posterior distribution. We then evaluated time to detection of true positives (TP) with quarterly cumulative assessment between 2005Q1 and 2009Q4. TPs were labelled AEs from the latest published version of SIDER (2015).
All analyses were performed using JAGS[2].

Results: Of the 886 PTs tested in the analysis, in 2009Q4 HMM had 360 alerts and MGPS had 313 with 303 in common, with 13.3% and 15.7% of PPV, respectively. In the quarterly analyses MGPS generated 398 alerts while HMM generated 495 alerts, some appearing and disappearing during the period. In total 68 TPs were detected by any method: 17 first by HMM, 1 first by MGPS, 37 in same quarter by both methods, 13 solely by HMM.

Conclusions: In our experiment, the Bayesian borrowing allows for better, and faster, detection of some labelled events. The borrowing methodology and MedDRA levels chosen have a dramatic impact on the results and demonstrate the importance of using a sensible ontology with respect to the analysis. Further work is required to finetune the HMM models and employ a suitable ontology.

[1] S. M. Berry and D. A. Berry, Biometrics, 60, 2004, 418-26.
[2] Y.-S. Su and Y. Masanao. R package version 0.03-08, URL https://CRAN.R-project.org/package=R2jags (2012).