Significance of data mining in routine signal detection: Analysis based on the safety signals identified by the FDA

Data mining has been introduced as one of the most useful methods for signal detection by spontaneous reports, but data mining is not always effective in detecting all safety issues. To investigate appropriate situations in which data mining is effective in routine signal detection activities, we analyzed the characteristics of signals that the US Food and Drug Administration (FDA) identified from the FDA Adverse Event Reporting System (FAERS).

Among the signals that the FDA identified from the FAERS between 2008 1Q and 2014 4Q, we selected 233 signals to evaluate in this study. We conducted a disproportionality analysis and classified these signals into two groups according to the presence or absence of statistical significance in the reporting odds ratio (ROR). Then, we compared the two groups based on the characteristics of the suspected drugs and adverse events (AEs).

Safety signals were most frequently identified for new drugs that had been on the market for less than 5 years, but some signals were still identified for old drugs (≥20 years), and most of them were statistically significant. The proportion of the signals for "serious" events was significantly higher in the group of nonsignals by ROR (Fisher's exact test, P = 0.032).

Data mining was shown to be effective in the following situations: (1) early detection of safety issues for newly marketed drugs, (2) continuous monitoring of safety issues for old drugs, and (3) signal detection of nonserious AEs, to which little attention is usually given.