Employing Maximum Mutual Information for Bayesian Classification

In order to employ machine learning in realistic clinical settings we are in need of algorithms which show robust performance, producing results that are intelligible to the physician. In this article, we present a new Bayesian-network learning algorithm which can be deployed as a tool for learning Bayesian networks, aimed at supporting the processes of prognosis or diagnosis. It is based on a maximum (conditional) mutual information criterion. The algorithm is evaluated using a high-quality clinical dataset concerning disorders of the liver and biliary tract, showing a performance which exceeds that of state-of-the-art Bayesian classifiers. Furthermore, the algorithm places less restrictions on classifying Bayesian network structures and therefore allows easier clinical interpretation.