Bayesian Data Mining Techniques: The Evidence Provided by Signals Detected in Single Company Spontaneous Reports Databases

Purpose: To apply and evaluate Bayesian inter-product quantitative methods for signaling an excess of adverse events to specific pharmaceutical products, taking into account sales data as well as other information accessible to a company's drug-monitoring system. Methods: The Bayesian confidence propagation neural network (BCPNN) and the gamma Poisson shrinkage (GPS) were applied to a selected sample of spontaneously reported adverse events following the administration of a Bracco contrast medium. Both the conventional approach and sales data were exploited to represent the patients' population drug exposure. Results: Available data allow the detection of potential safety issues of a drug in comparison to those expected in its pharmaceutical category. No difference in signal detection performance between the BCPNN and GPS methods was found. Instead, adjustment by sales data markedly affected the signals detected, with the desirable property of preserving the risk order, for any given adverse drug reaction, among different drugs. Conclusions: Without comprehensive data on the adverse events reported worldwide for all pharmaceutical products, signaling methods are appropriate to compare the safety of drugs sharing a similar clinical indication. Sales data show a relevant impact on the value of signals, improving the analysis of spontaneous reports collected by a company monitoring system.