Optimization of Backpropagation for Early Detection of Diabetes Mellitu

Abstract

Diabetes mellitus is one of the urgent health problems in the world. Diabetes is a condition primarily defined by the level of hyperglycemia giving rise to risk of micro vascular damage. Those who suffer from this disease generally do not realize and tend to overlook the early symptoms. Late recognition of these early symptoms may drive the disease to a more concerning level. One solution to solve this problem is to create an application that may perform early detection of diabetes mellitus so that it does not grow larger. In this article, a new method in performing early detection of diabetes mellitus is suggested. This method is backpropagation with three optimization namely early initialization with Nguyen-Widrow algorithm, learning rate adaptive determination, and determination of weight change by applying momentum coefficient. The observation is conducted by collecting 150 data consisting of 79 diabetes mellitus patient and 71 non diabetes mellitus patient data. The result of this study is the suggested algorithm succeeds in detecting diabetes mellitus with accuracy rate of 99.33%. Optimized backpropagation algorithm may allow the training process goes 12.4 times faster than standard backpropagation.