A qualitative approach to signal mining in pharmacovigilance using formal concept analysis.

"Pharmacovigilance is the process and science of monitoring the safety of medicines, consisting in (i) collecting and managing data on the safety of medicines (ii) looking at the data to detect 'signals' (any new or changing safety issue)" [1]. Pharmacovigilance is mainly based on spontaneous reports: when suspecting an adverse drug reaction, health care practitioners send a report to a spontaneous reporting system (SRS). This produces huge databases containing numerous reports and their manual exploration is both cost and time prohibitive. Existing techniques that automatically extract relevant signals rely on statistics or Bayesian models but do not provide information to the experts about possible biases lying in the data, nor about the specificity of a signal to a particular patient profile. Our extraction method combines numerical methods from the state of the art with a qualitative approach that helps interpretation. We build a synthetic representation of the database that is used to (i) identify unexpected patterns and biases (ii) extract potentially relevant signals w.r.t. patient profiles (iii) provide traceability facilities between extracted signals and raw data.