Bayesian networks for multivariate data analysis and prognostic modelling in cardiac surgery

Abstract

Prognostic models are tools to predict the outcome of disease and disease treatment. These models are traditionally built with supervised machine learning techniques, and consider prognosis as a static, one-shot activity. This paper presents a new type of prognostic model that builds on the Bayesian network methodology that implements a dynamic, process-oriented view on prognosis. In contrast to traditional prognostic models, prognostic Bayesian networks explicate the scenarios that lead to disease outcomes, and can be used to update predictions when new information becomes available. A recursive data analysis strategy for inducing prognostic Bayesian networks from medical data is presented, and applied to data from the field of cardiac surgery. The resulting model outperformed a model that was constructed with off-the-shelf Bayesian network learning software, and had similar performance as class probability trees.