A Semi-Causal Bayesian Network Approach to Prognosis

Various machine learning techniques have been proposed for the development of prognostic models, including those based on Bayesian networks. An advantage of a Bayesian network compared to many other classifiers is that the model can provide insight by representing the temporal structure of the domain. While it has been shown that Bayesian networks can perform well in terms of classification accuracy, we show in this paper that constraining the learning of a Bayesian network with temporal domain knowledge can harm the classification performance. Therefore, we propose to combine elements of naive classifiers with temporal domain knowledge, resulting in semi-causal Bayesian networks. We evaluate this approach in the development of a prognostic model for epithelial ovarian cancer, and argue that the model is understandable for domain experts and comparable to the performance of traditional prognostic models and tree-augmented naive classifiers.