Early Detection System Of Diabetes Mellitus Disease Using Artificial Neural Network Backpropagation With Adaptive Learning Rate And Particle Swarm Optimization

Abstract

Diabetes Mellitus (DM) is a health problem that is growing rapidly in Indonesia. According to the International Diabetes Federation (IDF) in 2013, DM patients in Indonesia were around 8.5 million people. Delay in recognizing the initial symptoms of DM can cause complications with other diseases and produce a more difficult treatment process that can even cause death. Early detection of DM is a way to detect the possibility of someone having DM. The problem under study is the clinical symptoms of DM which are observed only in outliser problems, where data can be categorized as a hot coding condition, where the data is in the form of ‘Yes' or ‘No’. In this study we conduced a comparison of several artificial neural network techniques for early detection of DM namely the Standart Backpropagation Neural Network (SBNN), SBNN with Adaptive Learning Rate (SBNN+ALR), SBNN with Particle Swarm Optimization (SBNN+PSO), or SBNN with Particle Swarm Optimization and Adaptive Learning Rate (SBNN+PSO+ALR). The variables used in this study are symptoms and factors supporting DM as many as 9 variables. Research data is taken from medical records at the Health Center (Puskesmas) Brebes. Distribution of training data and test data is determined by K-fold Cross Validation method. The results was showed that the best architecture is obtained SBNN+PSO+ALR. The SBNN+PSO+ALR architecture produced an average accuracy of 88,75%, sensitivity value of 82,5%, specificity value of 95% and Mean Squared Error (MSE) value of 0,02939 in only 30 epoch.