Comparison of Bayesian Networks for Diabetes Prediction

Abstract

A Bayesian network (BN) can be used to predict the prevalence of diabetes from the cause–effect relationship among risk factors. By applying a BN model, we can capture the interdependencies between direct and indirect risks hierarchically. In this study, we propose to investigate and compare the predictive performances of BN models with non-hierarchical (BNNH), and non-hierarchical and reduced variables (BNNHR) structures, hierarchical structure by expert judgment (BNHE), and hierarchical learning structure (BNHL) with type-2 diabetes. ROC curves, AUC, percentage error, and F1 score were applied to compare performances of those classification techniques. The results of the model comparison from both datasets (training and testing) obtained from the Thai National Health Examination Survey IV ensured that BNHE can predict the prevalence of diabetes most effectively with the highest AUC values of 0.7670 and 0.7760 from the training and the testing dataset, respectively.