ABSTRACT
Different structures and coding schemes may limit rapid evaluation of a large pool of potential drug safety signals using multiple longitudinal healthcare databases. To overcome this restriction, a semi-automated approach utilising common data model (CDM) and robust pharmacoepidemiologic methods was developed; however, its performance needed to be evaluated. Twenty-three established drug-safety associations from publications were reproduced in a healthcare claims database and four of these were also repeated in electronic health records. Concordance and discrepancy of pairwise estimates were assessed between the results derived from the publication and results from this approach. For all 27 pairs, an observed agreement between the published results and the results from the semi-automated approach was greater than 85% and Kappa coefficient was 0.61, 95% CI: 0.19-1.00. Ln(IRR) differed by less than 50% for 13/27 pairs, and the IRR varied less than 2-fold for 19/27 pairs. Reproducibility based on the intra-class correlation coefficient was 0.54. Most covariates (>90%) in the publications were available for inclusion in the models. Once the study populations and inclusion/exclusion criteria were obtained from the literature, the analysis was able to be completed in 2-8 h. The semi-automated methodology using a CDM produced consistent risk estimates compared to the published findings for most selected drug-outcome associations, regardless of original study designs, databases, medications and outcomes. Further assessment of this approach is useful to understand its roles, strengths and limitations in rapidly evaluating safety signals.
