Feature Transformation Through Stacked Autoencoder for Diabetes Classification

Abstract

Diabetes Mellitus (Diabetes) refers to a chronic disability of the human body to process glucose in the bloodstream. Diabetes is widespread globally and is associated with high healthcare costs and can cause problems such as neuropathy and heart attacks along the line. Accurate diabetes prediction is essential for proper diagnosis and treatment thereafter. Therefore, the proposed work aims to develop a diabetes classification system. This work experiments feature learning through stacked auto encoder and feature fusion of low-level features as well as deep features extracted from the Pima Indians Dataset. The classification process is carried out using a shallow and deep neural network. The proposed feature transformation techniques are evaluated using relevant performance metrics. Using feature fusion, the metric scores are 92.71%, 92.37%, 93.44% and 96.80% for accuracy, sensitivity, specificity, and precision respectively. Henceforth, this study could be used to aid medical practitioners for efficient diabetes diagnosis and to provide treatment at an early stage.