Data Mining and Statistically Guided Clinical Review of Adverse Event Data in Clinical Trials

Some approaches to the analysis of adverse event data arising from clinical trials are presented. These include (a) an inside-out data mining method where the adverse events are used as explanatory variables, classifying the treatment allocation, (b) a support method where we fit separate regression models to each adverse event with and without a treatment effect, and (c) a three-level hierarchical Bayesian mixture model for analysis of adverse event counts. The problem of understanding treatment-emergence of the adverse events is formulated as one of data mining rather than hypothesis testing. Our approaches provide an ordering of the adverse events by the strength of evidence of a treatment effect, rather than p values for prespecified hypotheses. The three methods produce intuitive graphical summaries showing the treatment effect on adverse event incidence. These graphs can be readily linked to relevant supportive information such as reports summarizing predicted risks for (demographic) subpopulations of interest and patient-level data such as laboratory information, concomitant medications, and medical history. This results in a statistically guided and thorough review of drug safety in the clinical trial.