Modelling Patterns of Evidence in Bayesian Networks: a Case-study in Classical Swine Fever Linda C. van der Gaag 1 , Janneke Bolt 1 , Willie Loeffen 2 , and Armin Elbers 2 1 Department of Information and Computing Sciences, Utrecht University, P.O. Box 80.089, 3508 TB Utrecht, The Netherlands {linda,janneke}@cs.uu.nl 2 Central Veterinary Institute, Wageningen UR, P.O. Box 65, 8200 AB Lelystad, The Netherlands {armin.elbers,willie.loeffen}@wur.nl Abstract. Upon engineering a Bayesian network for the early detection of Classical Swine Fever in pigs, we found that the commonly used ap- proach of separately modelling the relevant observable variables would not suffice to arrive at satisfactory performance of the network: explicit modelling of combinations of observations was required to allow identify- ing and reasoning about patterns of evidence. In this paper, we outline a general approach to modelling relevant patterns of evidence in a Bayesian network. We demonstrate its application for our problem domain and show that it served to significantly improve our network’s performance. 1 Introduction Over the last decades, researchers developed Bayesian networks to support med- ical and veterinary practitioners in their diagnostic reasoning processes for a variety of biomedical domains. Examples from our own engineering experiences include a Bayesian network for establishing the stage of oesophageal cancer in patients who have been diagnosed with the disease [1], naive Bayesian networks for deciding upon the most likely causal pathogen of clinical mastitis in dairy cows [2], and a dynamic Bayesian network for diagnosing ventilator-associated pneumonia in critically ill patients in an intensive care unit [3]. Our most recent engineering efforts concern a network for the early detection of an infection with the Classical Swine Fever (CSF) virus in individual pigs. Upon constructing our Bayesian network for the early detection of Classical Swine Fever, we found that the commonly used engineering approach of sepa- rately modelling the clinical signs found with the disease, would not suffice to arrive at satisfactory performance of the network. In-depth interviews with re- searchers and veterinary practitioners across the European Union showed that the aspecificity of especially the early signs of the disease makes a clinical di- agnosis highly uncertain and that satisfactory diagnostic performance can only be reached by reasoning about the presence or absence of specific combinations 1