Adverse Drug Reaction Mining in Pharmacovigilance Data Using Formal Concept Analysis

In this paper we discuss the problem of extracting and evaluating associations between drugs and adverse effects in pharmacovigilance data. Approaches proposed by the medical informatics community for mining one drug - one effect pairs perform an exhaustive search strategy that precludes from mining high-order associations. Some specificities of pharmacovigilance data prevent from applying pattern mining approaches proposed by the data mining community for similar problems dealing with epidemiological studies. We argue that Formal Concept Analysis (FCA) and concept lattices constitute a suitable framework for both identifying relevant associations, and assisting experts in their evaluation task. Demographic attributes are handled so that the disproportionality of an association is computed w.r.t. the relevant population stratum to prevent confounding. We put the focus on the understandability of the results and provide evaluation facilities for experts. A real case study on a subset of the French spontaneous reporting system shows that the method identifies known adverse drug reactions and some unknown associations that has to be further investigated.