Mining the WHO Drug Safety Database Using Lasso Logistic Regression

For reasons such as low incidence, occurrence in groups frequently excluded from clinical trials and long onset times, some adverse drug reactions (ADRs) of a new medicinal product stay unnoticed until after market launch. The World Health Organization (WHO) in collaboration with the Uppsala Monitoring Centre (UMC) continuously collect spontaneous ADR reports from the entire world and use data mining approaches to detect which drugs are most likely to cause which previously unanticipated ADRs. This WHO drug safety database, being the largest of its kind, comprises about 3.8 million accumulated reports. The currently used data mining methods are based on two-dimensional projections of the data with respect to a given drug-ADR combination. This combination is then given an association score based on the discrepancy between the observed and expected number of reports on it. In this thesis these disproportionality-based methods are represented by the information component (IC) measure of the UMC, a shrunk Bayesian measure. A limitation with the IC is its incapability to deal with confounding by co-medication and masking. Confounding by co-medication means that the association between a drug and a certain ADR might seem stronger than it really is because that drug is used together with another drug, which in turn is truly associated with the ADR. Masking, on the other hand, is a phenomenon whereby a very strong association between an ADR and some drug might weaken the associations between that ADR and other drugs. Here a novel method to mine the WHO drug safety database is proposed to address these issues, the lasso logistic regression (LLR). Instead of studying each combination separately, in the LLR model the ADR under study is fixed and its presence on a report is predicted by the presence of all occurring drugs in the database, thus yielding a logistic regression framework. Further, independent prior Laplace distributions are put on the parameters, resulting in a lasso-type shrinkage where a subset of the parameters are shrunk to exactly zero. The LLR was confirmed to correct for confounding by co-medication and masking in simulated scenarios and specific clinical examples. Further, with a specific degree of shrinkage the LLR had 10 % higher recall and maintained precision in comparison to the IC with respect to a test database. Although its transparency is limited, the LLR has an important role to play in the future of ADR monitoring.