Genetic-evolved Bayesian networks in a biomedical application

Abstract

This study presents genetic algorithm (GA) for discovering Bayesian network structure. The algorithm is applied to a medical datasets for vertebral column. Data set containing values for six biomechanical features is used to classify patients into three categories: disk hernia (DH), spondylolisthesis (SL), and normal (NO) or two categories: abnormal (AB), and NO. On ten-fold cross-validation run, the average AUC (the area under the ROC curve) measures of 0.874 and 0.923 for two and three categories are obtained, respectively. Results indicate that GA is relatively effective algorithm. Consequently, the GA-evolved BN is powerful tool for knowledge representation and inference because the causality relationship can be observed.