Reuse of Bayesian Networks: A Case-study in Classical and African Swine Fever

Abstract

Developing a Bayesian Network has a high workload, also for domain experts, when not enough data is available to learn the model. We aim to reduce this workload by reusing an existing Bayesian Network when developing a new network. We study this by developing an initial model for African Swine Fever (ASF) by reusing the already existing Classical Swine Fever (CSF) model. African Swine Fever is a highly contagious disease, which is currently present in Poland and the Czech Republic. The risk of contamination in the Netherlands is substantial, and especially because no vaccine is available, a quick diagnosis is essential. Therefore, we developed a Bayesian Network to support early detection of the disease without having to wait for lab results. The existing model for CSF consists of fi?ve phases, each representing a part of the body affected. These phases are used as a base, on which to build the reused model. The initial structure of the ASF model is determined, using only literature, very limited expert interviews and data of inoculation studies. When learning the parameters of the model, the probabilities of the CSF model where reused where possible. The remaining conditional probability tables are determined by using a variant of the EM algorithm. The resulting network displays how a good initial model can be made in signifi?cant less time compared to developing a new one.