Exploration Of Statistical Shrinkage Parameters In Adverse Drug Reaction Signal Detection In Spontaneous Reporting System Of China

Background:Adverse Drug Reaction(ADR) Monitoring is an important way of drugs post-marketing evaluation, and the spontaneous reporting system(SRS) is the primary means of Adverse Drug Reaction Monitoring. The number of China’s adverse drug reaction reports has grown rapidly in recent years, has become one of the world’s largest database of adverse drug reactions. Disproportionate analys Is(DPA) method can generate a large number of adverse drug reactions signals from spontaneous reporting system database. However, the traditional disproportionate analysis method has certain defects, namely when the expect frequency or the number of reprots cases is small, which could generate false positive signals. How to properly use data mining methods to discovery of the true signal efficient and accurate related to the health of society and also the focus of pharmacovigilance issues. Hence, it is very important to apply appropriate data mining and statistical methods to implement post-marketing surveillance(PMS) in SRS data. Statistical shrinkage is a potential statistical method to improve the accuracy of signal detection results and avoid spurious associations detected by DPA. In recent years, the WHO Uppsala Monitoring Centre has achieved good results by using statistical shrinkage DPA. China’s spontaneous reporting system is different from other country or agency. So,it is necessary to explorate statistical shrinkage parameters in adverse drug reaction signal detection in China.Aim:This thesis aims to explore appropriate shrinkage parameters enhanced signal stability to reduce the number of false-positive signals for three kinds of DPA. Compared changes in the indicators at the different parameters. Finally, the optimal shrinkage parameters will be chosen and applied in the ADR signal detection to avoid the false signals.Methods:Disproportionality analyses methods, including proportional reporting ratio(PRR), reporting odds Ratio(ROR), information component(IC), were employed. DPA are based on the ratio of observation(O) frequency to expected(E) frequency. When O/E ratio exceeds a critical value, a potential signal is determined. We added the shrinkage parameters in the numerator and denominator, denoted as, respectively, in the formula of disproportionality analysis. We assumed and shrinkage parameters were subjectively set to between 0-5, with interval of 0.1. ADR product label database was deemed as a proxy of golden standard to evaluate the effect of statistical shrinkage. Reports in the years of 2010-2011 were extracted from the national SRS database as the data source for analysis in this study. Statistical analyses were performed using SAS 9.3 software.Results:After data processing, 1823144 reports representing 140506 combinations were kept. The analysis generated 59227, 57427 and 38393 signals by traditional PRR, ROR and IC, respectively. After analysis of the numbers of signals using statistical shrinkage, we found that the number of signals was reduced with the increase of the parameter value. When α was in the range of 1.1-5.0, the number of signals stabilized at more than 10000. We found that the number of potential signals of the three methods was reduced with the increase of the parameter value. And the smaller the expected value was, the easier the potential signals were to disappear. The numbers of signals reduced and expected frequency distribute were in line with expectation. Statistical shrinkage DPA not only reduced false positive signals but also diminished a small part of the true signals. Finally, we determine the most appropriate parameter values based on the Youden index for statistical shrinkage DPA(based on ROR, PRR and IC).Conclusions:The numbers of signals reduced and expected frequency distribute were in line with expectation. Reducing the number of signals and excluding potential signals with very small expected frequency made it earlier to find out true signals and avoided false positives signals. Statistical shrinkage, as a statistical method, is used in the ADR signal detection with its unique advantage. It can enhance the signal stability and less prone to highlight spurious associations detected by DPA. Youden indices can comprehensive response to sensitivity and specificity two indicators. The Youden indices of the three methods were increasing with the growing increase shrinkage parameters. And the Youden indices reached the maximum at rang of 0.5. So, the value of 0.5 of was suggested as the most appropriate statistical shrinkage parameters for SRS of China.