Bayesian neural networks with confidence estimations applied to data mining. (Q1583507): Difference between revisions

An international database of case reports, each one describing a possible case of adverse drug reactions (ADRs), is maintained by the Uppsala Monitoring Centre (UMC), for the WHO international program on drug safety monitoring. Each report can be seen as a row in a data matrix and consists of a number of variables, like drugs used, ADRs, and other patient data. The problem is to examine the database and find significant dependencies which might be signals of potentially important ADRs, to be investigated by clinical experts. We propose a method by which estimated frequencies of combinations of variables are compared with the frequencies that would be predicted assuming there were no dependencies. The estimates of significance are obtained with a Bayesian approach via the variance of posterior probability distributions. The posterior is obtained by fusing a prior distribution (Dirichlet of dimension \(2^{n-1})\) with a batch of data, which is also the prior used when the next batch of data arrives. To decide whether the joint probabilities of events are different from what would follow from the independence assumption, the ``information component'' \(\log(P_{ij}/(P_{i}P_{j}))\) plays a crucial role, and one main technical contribution reported here is an efficient method to estimate this measure, as well as the variance of its posterior distribution, for large data matrices. The method we present is fundamentally an artificial neural network denoted Bayesian confidence propagation neural network (BCPNN). We also demonstrate an efficient way of finding complex dependencies. The method is now (autumn 1998) being routinely used to produce warning signals on new unexpected ADR associations.