Learning bayesian-network topologies in realistic medical domains

Abstract

In recent years, a number of algorithms have been developed for learning the structure of Bayesian networks from data. In this paper we apply some of these algorithms to a realistic medical domain—stroke. Basically, the domain of stroke is taken as a typical example of a medical domain where much data are available concerning a few hundred patients. Learning the structure of a Bayesian network is known to be hard under these conditions. In this paper, two different structure learning algorithms are compared to each other. A causal model which was constructed with the help of an expert clinician is adopted as the gold standard. The advantages and limitations of various structure-learning algorithms are discussed in the context of the experimental results obtained.