Performance of Bayesian networks (naive Bayes and tree augmented naive Bayes) in detecting diabetes types 1 and 2

Abstract

This study is aimed at studying diabetes data using Bayesian networks and evaluating their 
performance in detecting types 1 and 2 diabetes. The diabetes dataset was obtained from 
the medical records unit of the Abubakar Tafawa Balewa University Teaching Hospital 
Bauchi, Bauchi State, consisting of 569 cases with 8 different variables. Classification of 
diabetes patients was carried out by naive Bayes and tree augmented naive Bayes (TAN) 
networks; the networks were trained and tested using a 10-fold cross-validation and 
the quality of prediction of these networks in terms of sensitivity, specificity, area under 
the ROC (AUR), kappa statistic, mean absolute error (MAE), and root mean square error 
(RMSE) was evaluated. The results indicate that TAN network proved to be the better 
network because the method correctly classified 530 (93.15%) and misclassified only 39 
(6.85%) patients; it also has higher AUR (0.949) and kappa (0.7869), as well as lower MAE 
and RMSE of 0.1032 and 0.2384, respectively.