Learning Bayesian Networks from Classification trees and Expert knowledge: a preliminary study

To address the classification problem when the number of cases is too small to effectively use just a single technique, this paper suggests to use a “hybrid” approach, that combines tree classifiers, expert knowledge and Bayesian networks. The output of the “hybrid” strategy takes the form of a Bayesian network, where the serious drawback of requiring huge amounts of data was overcome by coupling the network with another classifier and using expert knowledge. The technique was applied to two clinical case-studies and its predictive performance was compared with the performances of the single approaches. The results show that the proposed technique benefits from several advantages of the three single approaches and outperforms all of them: it shows a better sensitivity, that plays a crucial role in classifying new patients into the high-risk category, and is competitive with regards to specificity. Therefore, even though additional studies are needed to validate the “hybrid” approach, it seems a promising technique to develop reliable classification systems for small datasets.