Background: Safety surveillance relies on mining of large pharmacovigilance (PV) databases to generate insights regarding the safe use of pharmaceutical products. The predominant approach to PV data mining involves computation of disproportionality scores for drug-adverse event (drug-AE) pairs. However, this approach requires a database to be sufficiently large, sufficiently diverse for the analysis to be reliably sensitive and specific, and fails to consider the particular safety profile of a product.
Objective: The present study proposes and tests a novel, frequency-based approach to PV data mining that (1) leverages product knowledge and historical drug-AE trends and (2) imposes no requirement for the size and diversity of the database to which it is applied.
Method: A focus group of physicians and scientists was convened to identify quantitative characteristics of data trends that they consider informative when reviewing counts of adverse events for products under surveillance. Feedback was transferred into a series of decision rules that, when applied to adverse event counts, identifies adverse event trends that are classified as Continuing Trend, Emerging Trend, or No Trend. Regression analyses are completed to verify the presence of a linear trend; and categorical measures of association completed to compare this frequency-based approach to disproportionality scores in a simulated database.
Results: A significant, positive linear trend is present for the Continuing Trend and Emerging Trend categories ( P < .0001). There is a significant association between trend categorizations and disproportionality scores ( P < .0001).
Conclusion: The proposed alternative frequency-based method for PV data mining would be useful where disproportionalities scores are not appropriate. Additionally, this method may be useful in conjunction with disproportionality scores, where appropriate, highlighting adverse events that are both reported disproportionately and have increasing trends.
Keywords: data mining; disproportionality; pharmacovigilance; safety statistics; signal detection.