Qualitative Chain Graphs and their Use in Medicine

For modelling diseases in medicine, chain graphs are more attractive than directed graphs, i.e., Bayesian networks, as they support representing interactions between diseases that have no natural direction. In particular, representation by chain graphs is preferred over Bayesian networks as they have the ability to capture equilibrium models. Using qualitative abstractions of probabilistic interactions is also of interest in this context, as these would allow focusing on patterns in the interactions rather than looking at the numerical detail, which for medical purposes is of paramount importance. So far, qualitative abstractions of probabilistic interactions have been developed only for Bayesian networks in the form of the framework of qualitative probabilistic networks. In this paper, qualitative abstractions are developed for chain graphs with the practical purpose of using these as constraints on the hyperspace of probability distributions. The usefulness of this approach is explored for disease modelling.