Bayesian Networks for the Analysis of Drug Safety

Abstract

We present a method for predicting the rate of adverse reactions to a drug. The approach employs a graphical model, as previously used for assessing causality in individual cases, but instead of attempting to interpret what did happen in an individual case we use it to predict what will happen in the next case or series of cases treated. The approach is illustrated on the adverse reactions of pseudomembranous colitis due to antibiotics. Based on a representation of the process by which this adverse event is recognized and reported, and expert opinion on the probabilities of key signs given diagnosis, we demonstrate how the model predicts the rate of true and reported adverse events. We also demonstrate how an elaboration of this approach can allow the estimates to be modified continuously and updated as new cases are reported. Although further work needs to be done to make the system practically usable, we see it as a demonstration of a new and radically different approach to the main concern of pharmaco‐epidemiology—the assessment of drug safety. The crucial shift is from a primary concern with estimating incidence rates in patients treated in the past to predicting incidence rates in future patients to be treated with a drug.

Bayesian statistics in medicine: a 25 year review

“…A third Institute of Statisticians Conference on Applied Bayesian Statistics was held in 1992, with the proceedings running to three issues of The Statistician. Medical application included graphical elicitation of priors for clinical trials [115], monitoring of clinical trials [116], drug safety [117], case-control studies in cancer epidemiology [118], back-calculation of the numbers infected in the HIV epidemic [119], dosage regimens in population pharmaco-kinetics [120], analysis of binary cross-over data [121] and modelling heterogeneity in environmental epidemiology [122].…”

Selecting treatment strategies with dynamic limited-memory influence diagrams

“…that represents the distribution over C when the decision-maker has set D equal to d [19]. Hence, C is not conditioned on D, but rather parameterized by D. 1 Fig.…”

“…As the figure suggests, one way to do probabilistic inference is to unroll ðB 0 ; B t Þ into one big static network and to use a standard inference algorithm, such as the junction tree algorithm [19]. However, although the complexity of inference is determined by the size of the largest clique that is obtained after triangularization of the graph underlying the static network [37], space complexity also grows linearly in the horizon N, and therefore this approach is unsuitable for large horizons.…”

Multi-objective optimization in graphical models

“…Graphical models [35,111,148,71,29] provide a common formalism to describe a wide range of systems. Graphical models have been adopted in a wide variety of application areas, including genetics [46,93], error-correcting codes [102], speech processing [116], image analysis [49,9,142], computational biology [130,117], scheduling [12] and electronic commerce [78].…”