Early detection of pharmacovigilance signals with automated methods based on false discovery rates: A comparative study

摘要

Background: Improving the detection of drug safety signals has led several pharmacovigilance regulatory agencies to incorporate automated quantitative methods into their spontaneous reporting management systems. The three largest worldwide pharmacovigilance databases are routinely screened by the lower bound of the 95% confidence interval of proportional reporting ratio (PRR 02.5), the 2.5% quantile of the Information Component (IC 02.5) or the 5% quantile of the Gamma Poisson Shrinker (GPS 05). More recently, Bayesian and non-Bayesian False Discovery Rate (FDR)-based methods were proposed that address the arbitrariness of thresholds and allow for a built-in estimate of the FDR. These methods were also shown through simulation studies to be interesting alternatives to the currently used methods. Objective: The objective of this work was twofold. Based on an extensive retrospective study, we compared PRR 02.5, GPS 05 and IC 02.5 with two FDRbased methods derived from the Fisher's exact test and the GPS model (GPS pH0 [posterior probability of the null hypothesis H 0 calculated from the Gamma Poisson Shrinker model]). Secondly, restricting the analysis to GPS pH0, we aimed to evaluate the added value of using automated signal detection tools compared with 'traditional' methods, i.e. non-automated surveillance operated by pharmacovigilance experts. Methods: The analysis was performed sequentially, i.e. every month, and retrospectively on the whole French pharmacovigilance database over the period 1 January 1996-1 July 2002. Evaluation was based on a list of 243 reference signals (RSs) corresponding to investigations launched by the French Pharmacovigilance Technical Committee (PhVTC) during the same period. The comparison of detection methods was made on the basis of the number of RSs detected as well as the time to detection. Results: Results comparing the five automated quantitative methods were in favour of GPS pH0 in terms of both number of detections of true signals and time to detection. Additionally, based on an FDR threshold of 5%, GPS pH0 detected 87% of the RSs associated with more than three reports, anticipating the date of investigation by the PhVTC by 15.8 months on average. Conclusions: Our results show that as soon as there is reasonable support for the data, automated signal detection tools are powerful tools to explore large spontaneous reporting system databases and detect relevant signals quickly compared with traditional pharmacovigilance methods.