Predicting Diabetes using Deep Belief Network

Abstract

Diabetes is a common disease affecting millions of people in the United States. The prevalence of diabetes has been steadily increasing over the past few decades. Left untreated diabetes leads to a risk factor for multiple complications such as heart disease, stroke, kidney and nerve damage. However, diabetes can be effectively managed when early diagnosed. Recent clinical data have applied various advanced technologies to diagnose or predict people with diabetes. Therefore, this paper aims to  predict diabetes by using a Deep Belief Network (DBN) with Rectified Linear Unit (ReLU) activation function. Learning parameters from data through unsupervised path using minimized contrastive divergent algorithm, followed by supervised path using back-propagation algorithm. Diabetes Dataset from the National Institute of Diabetes and Digestive and Kidney Diseases was used, consisting of  female patients aged at least 21 years old of Pima Indian heritage. The result shows that the DBN with ReLU activation function provides an 81% accuracy in diabetes prediction.