Towards a Method of Building Causal Bayesian Networks for Prognostic Decision Support

Abstract

We describe a method of building a decision support system for clinicians deciding between interventions, using Bayesian Networks (BNs). Using a case study of the amputation of traumatically injured extremities, we explain why existing prognostic models used as decision aids have not been successful in practice. A central idea is the importance of modeling causal relationships, both so that the model confiorms to the clinician's way of reasoning and so that we can predict the probable effect of the available interventions. Since we cannot always depend on data from controlled trials, we depend instead on 'clinical knowledge' and it is therefore vital that this elicited rigorously. We propose three stages of knowledge modeling covering the treatment process, the information generated by the process and the causal relationship. These stages lead to a causal Bayesian network, which is used to predict the patient outcome under different treatment options.