Data mining for prospective early detection of safety signals in the Vaccine Adverse Event Reporting System (VAERS): a case study of febrile seizures after a 2010-2011 seasonal influenza virus vaccine.

BACKGROUND: Reports of data mining results as an initial indication of a prospectively detected safety signal in the US Vaccine Adverse Event Reporting System (VAERS) have been limited. In April 2010 a vaccine safety signal for febrile seizures after Fluvax(®) and Fluvax(®) Junior was identified in Australia without the aid of data mining. In order to refine Northern Hemisphere influenza vaccine safety surveillance, VAERS data mining analyses based on vaccine brand name were initiated during the 2010-2011 influenza season.
OBJECTIVE: We describe the strategies that led to the finding of a novel safety signal using empirical Bayesian data mining.
METHODS: The primary US VAERS analysis calculated an empirical Bayesian geometric mean (EBGM), which was adjusted for age group, sex and year received. A secondary age-stratified analysis calculated a separate EBGM for 11 pre-defined age subsets. These bi-weekly analyses were generated with database restrictions that separated live and inactivated vaccines as well as with the US VAERS database. A cutoff of 2.0 at the fifth percentile of the confidence interval (CI) for the EBGM, the EB05, was used to identify vaccine adverse event combinations for further evaluation. Examination of potential interactions among concomitantly administered vaccines is based on the Interaction Signal Score (INTSS), which is a relative measure of how much excess disproportionality is present in the three-dimensional combination of two vaccines and one adverse event term. An INTSS >1 indicates that the CI for the three-dimensional analysis is larger than and does not overlap with the CI from the highest two-dimensional analysis. We subsequently examined the possibility of masking by removing all 2,095 Fluzone(®) 2010-2011 reports from the 10 December 2010 version of the VAERS database. In addition, we calculated relative reporting ratios to observe the relative contribution of adjustment and the Multi-Item Gamma Poisson Shrinker (MGPS) algorithm to EBGM values.
RESULTS: On 10 December 2010, US VAERS analyses we found an EB05 >2 for Fluzone(®) 2010-2011 and the Medical Dictionary for Regulatory Activities (MedDRA(®)) term "febrile seizure". MedDRA(®) terminology is the medical terminology developed under the auspices of the International Conference on Harmonization of technical requirements for Registration of Pharmaceuticals for Human Use (ICH). No other vaccine products had independent vaccine-febrile seizure combinations with an EB05 >2. Three-dimensional analyses to examine possible interactions among vaccine products concomitantly administered with Fluzone(®) 2010-2011 yielded Interaction Signal Score values <1. Removal of all Fluzone(®) 2010-2011 reports from the VAERS database failed to demonstrate a previously masked vaccine adverse event pair with an EB05 >2. The inactivated vaccine database restriction resulted in a 41 % reduction in background VAERS reports and a 24 % reduction in foreground VAERS reports.
CONCLUSION: Empirical Bayesian data mining in VAERS prospectively detected the safety signal for febrile seizures after Fluzone(®) 2010-2011 in young children. The EB05 threshold, database restrictions, adjustment and baseline data mining were strategies adopted a priori to enhance the specificity of the 2010-2011 influenza vaccine data mining analyses. A database restriction used to separate live vaccines resulted in a reduced EB05. Adjustment of data mining analyses had a larger effect on estimates of disproportionality than the MGPS algorithm. Masking did not appear to influence our findings. This case study illustrates the value of VAERS data mining for vaccine safety monitoring.