Statistical methods for knowledge discovery in adverse drug reaction surveillance Statistical methods for knowledge discovery in adverse drug reaction surveillance

Collections of individual case safety reports are the main resource for early discovery of unknown adverse reactions to drugs once they have been introduced to the general public. The data sets involved are complex and based on voluntary submission of reports, but contain pieces of very important information. The aim of this thesis is to propose computationally feasible statistical methods for large-scale knowledge discovery in these data sets. The main contributions are a duplicate detection method that can reliably identify pairs of unexpectedly similar reports and a new measure for highlighting suspected drug–drug interaction. Specifically, we extend the hit-miss model for database record matching with a hit-miss mixture model for scoring numerical record fields and a new method to compensate for strong record field correlations. The extended hit-miss model is implemented for the WHO database and demonstrated to be useful in real world duplicate detection, despite the noisy and incomplete information on individual case safety reports. The Information Component measure of disproportionality has been in routine use since 1998 to screen the WHO database for excessive adverse drug reaction reporting rates. Here, it is further refined. We introduce improved credibility intervals for rare events, post-stratification adjustment for suspected confounders and an extension to higher order associations that allows for simple but robust screening for potential risk factors. A new approach to identifying reporting patterns indicative of drug–drug interaction is also proposed. Finally, we describe how imprecision estimates specific to each prediction of a Bayes classifier may be obtained with the Bayesian bootstrap. Such case-based imprecision estimates allow for better prediction when different types of errors have different associated loss, with a possible application in combining quantitative and clinical filters to highlight drug–ADR pairs for clinical review.