Real-Time Non-Invasive Detection and Classification of Diabetes Using Modified Convolution Neural Network

Non-invasive diabetes prediction has been gaining prominence over the last decade. Among many human serums evaluated, human breath emerges as a promising option with acetone levels in breath exhibiting a good correlation to blood glucose levels. Such correlation establishes acetone as an acceptable biomarker for diabetes. The most common data analysis strategies to analyze the biomarkers in breath for disease detection use feature extraction and classification algorithms. However, snags such as computational cost and lack of optimal feature selection on application to real-time signals reduce the efficiency of such analysis. This paper explores the use of a one-dimensional (1-D) modified convolution neural network (CNN) algorithm that combines feature extraction and classification techniques. The approach proposed in this paper is found to significantly reduce the limitations associated with using these techniques individually and thereby improving the classifier's performance further. This paper proposes to apply a modified 1-D CNN on real-time breath signals obtained from an array of gas sensors. The experimentation and the performance of the system is carried out and evaluated.