Key points are not available for this paper at this time.
The large number of adverse-event reports generated by marketed drugs and devices argues for the application of validated computerized algorithms to supplement traditional methods of detecting adverse-event signals. Difficulties in accurately estimating patient exposure and background rates for a given event in a specific population hinder risk estimation in spontaneous adverse-event databases. The United States Food and Drug Administration (FDA) is evaluating a Bayesian data mining system called Multi-item Gamma Poisson Shrinker (MGPS) to enhance the FDA's ability to monitor the safety of drugs, biologics, and vaccines after they have been approved for use. The MGPS computes adjusted higher-than-expected reporting relationships between drugs and adverse events across 35 years of data relative to internal background rates. The MGPS can also adjust for random noise by using a model derived from the data, and corrects for temporal trends and confounding related to age, sex, and other variables by stratifying over 900 categories. Signals can then be compared with or used in conjunction with other sources (e.g. clinical trials, general practice databases) to further study the adverse-event risk. The example of pancreatitis risk with atypical antipsychotics, valproic acid, and valproate is used to discuss the strengths and limitations of MGPS versus traditional methods. Validated data mining techniques offer great promise to enhance pharmacovigilance practices.
Szarfman et al. (Wed,) studied this question.