ISSN 2630-0583 (Print)

ISSN 2630-0656 (Online)

JCST

Journal of Current Science and Technology

http://jcst.rsu.ac.th

Rangsit Journal of Arts and Sciences. Vol.3 No.1 , January - June 2013.

Application program for prediction of the type 2 diabetes in Thai people using artificial neural network

Wuttichai Luangruangrong, Annupan Rodtook, and Sanon Chimmanee

Abstract

           Among non-communicable diseases, diabetes kills the most people in Asia and is only becoming more prevalent in this region.  Analyzing Type 2 diabetes risk factors utilizing prediction tools instead of blood testing is a challenge for accurate diabetes diagnosis.  Recently, many researchers have studied the risk factors of diabetes by using Logistic Regression, Radial Basis and Back-propagation Neural Network (BNN) and applying them as a tool for diabetes prediction.  This paper presents the new factors of smoking and alcohol consumption to improve performance in diabetes prediction.  The predictive role of some traditional factors, i.e., body mass index, blood pressure and waist circumference and Family History (FMH) are also improved by adjusting the previously accepted ranges. The newly proposed diabetes prediction method is based on BNN.  The sample data consists of 2,000 Thai people presenting at Bangkok hospital, Thailand from 2010 to 2012.  From these experiments, it was found that an appropriate number of hidden nodes was equal to 50 nodes.  Each proposed factor, i.e., FMH, alcohol consumption factor, smoking factor, and WC gave a better accuracy (correct in prediction) compared with a baseline model.  Their accuracies were 83.35%, 83.50%, 83.60% and 83.65%, respectively.  Subsequently, the new risk factor model performance was increased by tuning the neural network parameter learning rate.  Our previously proposed factors for tuning BNN parameters introduced a high accuracy compared with the baseline model up to 1.2%. In this paper, the new proposed factors model introduces a better performance in Root Mean Square Error (RMSE) than the baseline factors model up to 25.75%, which are trained by the same sample data (2000 cases).  Finally, the new model is implemented to be the diabetes prediction tool based on PHP web application, which works in conjunction with Matlab for predicting calculation.  The threshold value returned is used to make a decision of whether or not patients have diabetes. 

Keywords: diabetes, prediction tools, back-propagation, neural network, neural network tuning, risk factors

Download Full Paper.

DOI: 10.14456/rjas.2012.22