Can Disproportionality Analysis of Post-marketing Case Reports be Used for Comparison of Drug Safety Profiles?

Clinical trials usually do not have the power to detect rare adverse drug reactions. Spontaneous adverse reaction reports as for example available in post-marketing safety databases such as the FDA Adverse Event Reporting System (FAERS) are therefore a valuable source of information to detect new safety signals early. To screen such large data-volumes for safety signals, data-mining algorithms based on the concept of disproportionality have been developed. Because disproportionality analysis is based on spontaneous reports submitted for a large number of drugs and adverse event types, one might consider using these data to compare safety profiles across drugs. In fact, recent publications have promoted this practice, claiming to provide guidance on treatment decisions to healthcare decision makers. In this article we investigate the validity of this approach. We argue that disproportionality cannot be used for comparative drug safety analysis beyond basic hypothesis generation because measures of disproportionality are: (1) missing the incidence denominators, (2) subject to severe reporting bias, and (3) not adjusted for confounding. Hypotheses generated by disproportionality analyses must be investigated by more robust methods before they can be allowed to influence clinical decisions.

Are all quantitative postmarketing signal detection methods equal? Performance characteristics of logistic regression and Multi-item Gamma Poisson Shrinker.

PURPOSE: The detection of safety signals with medicines is an essential activity to protect public health. Despite widespread acceptance, it is unclear whether recently applied statistical algorithms provide enhanced performance characteristics when compared with traditional systems. Novartis has adopted a novel system for automated signal detection on the basis of disproportionality methods within a safety data mining application (Empirica™ Signal System [ESS]). ESS uses two algorithms for routine analyses: empirical Bayes Multi-item Gamma Poisson Shrinker and logistic regression (LR). METHODS: A model was developed comprising 14 medicines, categorized as "new" or "established." A standard was prepared on the basis of safety findings selected from traditional sources. ESS results were compared with the standard to calculate the positive predictive value (PPV), specificity, and sensitivity. PPVs of the lower one-sided 5% and 0.05% confidence limits of the Bayes geometric mean (EB05) and of the LR odds ratio (LR0005) almost coincided for all the drug-event combinations studied. RESULTS: There was no obvious difference comparing the PPV of the leading Medical Dictionary for Regulatory Activities (MedDRA) terms to the PPV for all terms. The PPV of narrow MedDRA query searches was higher than that for broad searches. The widely used threshold value of EB05 = 2.0 or LR0005 = 2.0 together with more than three spontaneous reports of the drug-event combination produced balanced results for PPV, sensitivity, and specificity. CONCLUSIONS: Consequently, performance characteristics were best for leading terms with narrow MedDRA query searches irrespective of applying Multi-item Gamma Poisson Shrinker or LR at a threshold value of 2.0. This research formed the basis for the configuration of ESS for signal detection at Novartis.