Data Mining in Pharmacovigilance -Detecting the Unexpected

Background: One of the most important aims of pharmacovigilance is to detect signals of adverse drug reactions (ADRs) as early as possible. However, some ADRs are difficult to detect, one example being so called ‘type C’ reactions. These are effects that present as seemingly ‘spontaneous’ diseases occurring during treatment with a drug, such as the occurrence of a cardiovascular event while the patient is taking rofecoxib. As this type of ADR is often mistaken for a spontaneous disease, the causative agent may appear as an innocent bystander.

Objective: The primary aim of this study was to investigate the possibility of using data mining approaches to detect signals of ‘type C’ reactions. We hypothesized that by including concomitant, and not only suspected medications in the calculations of disproportionality analyses, we would be able to identify such reactions.

Study design: We used data from the Swedish Drug Information System, SWEDIS, which contains spontaneous reports submitted by Swedish physicians to the Swedish Medical Products Agency, and applied Bayesian confidence propagation neural network (BCPNN) methodology to calculate the information component (IC) value for drug-event combinations for drugs belonging to the Anatomic Therapeutic Chemical (ATC) classes cardiovascular system, musculoskeletal system and nervous system (number of reports = 51 270) where only the suspected drug was considered, and also where both concomitant and suspected drugs were considered. We then classified drug-event combinations that were signalled by a statistically significantly raised IC value as labelled or unlabelled based on the approved summary of product characteristics (SPC) in Sweden as of November 2007, and further classified them as ‘type C’ reactions or not ‘type C’.

Main outcome measure: The proportion of ‘type C’ reactions signalled when considering both concomitant and suspected drugs compared with suspected drugs only.

Results: The proportion of labelled drug-event combinations when considering suspected drugs was 78.6%. Drug-event combinations classified as ‘type C’ reactions were more likely to be found when considering both concomitant and suspected drugs compared with suspected drugs only; 18/449 versus 0/248 when considering drug-event combinations that were signalled exclusively by one of the approaches. Such drug-event combinations included, for example, sudden death and celecoxib, myocardial infarction and diclofenac, suicide-related events and several antidepressants.

Conclusion: Including both concomitant and suspected drugs in data mining practices may be a way of detecting ‘type C’ reactions earlier. This could constitute an advance in data mining for pharmacovigilance practices.