Abstract by Mia Aakjær

Active surveillance systems based on electronic healthcare data are being established to complement current pharmacovigilance, which comprises primarily of SRS after launch. Corresponding to the four studies, the objectives of the thesis were to I) develop and test a near-real-time epidemiological surveillance system with repeated cohort studies, II) screen for signals with tree-based scan statistics and assess the agreement with the method in Study I, III) investigate the risk of acute pancreatitis among new users of fluoxetine, and IV) investigate the risk of serious arrhythmia in initiators of citalopram and escitalopram.

Data from the Danish national registries from January 1, 1996, to December 31, 2016, were used. Four cohort studies using intention-to-treat approaches were conducted. Patients were followed for a maximum of 6 months from filing a prescription until the event of interest or censoring. In Study I, the exposures were selective serotonin reuptake inhibitors (SSRIs) and serotonin-norepinephrine reuptake inhibitors (SNRIs). The comparators were citalopram or other SSRIs and SNRIs. In total, 51 serious medical events were screened. A general set of confounders were applied in multivariate regression models to adjust for confounding. We detected 31 drug-medical event combinations positively associated with the outcome of interest in the exposed group as signals in a prospective setting. Of these, 67.7% were not listed in the current Danish Summaries of Product Characteristics (SmPCs). Study II had the same exposures and comparators as those in Study I. A wide range of ICD-10 diagnosis codes was screened using tree-based scan statistics. The confounders from Study I were used but in propensity score-matched cohorts. In total, 300 drug-medical event combinations were detected as signals. Of these, 82% were not present in the SmPCs. Of the 31 combinations detected as signals from Study I, 19 were investigated. Of these, seven (36.8%) were re-detected. Cohen’s kappa was used to assess the agreement. We found a fair agreement between the methods, κ = 0.36 (95% confidence interval 0.15-0.58). Study III focused on fluoxetine having citalopram and other SSRIs as comparators. The outcome was acute pancreatitis. We adjusted for confounding using tailored risk factors for the outcome applied with propensity score methods. Overall, no increased risk of acute pancreatitis was identified, but analyses including recurrent pancreatitis events showed higher point estimates. Study IV emulated a target trial, in which the exposures were citalopram and escitalopram, and the outcome was serious arrhythmia. Propensity score methods using tailored risk factors were applied. No increased risk of serious arrhythmia was observed. However, after the issue of drug safety warnings of QT prolonging risks, lower point estimates were observed.

In conclusion, signal detection with epidemiological approaches in electronic healthcare data could complement current safety surveillance. After an in-depth analysis of fluoxetine use and acute pancreatitis and citalopram and escitalopram initiation and the risk of serious arrhythmia, we found that neither poses an increased risk. However, signal evaluation could be conducted quickly and efficiently with pre-specified study designs requiring limited tuning related to the specific drug-associated risks.