Comparative Analysis of Regressor Models on Non-invasive Blood Glucose Dataset

Diabetes affects more than 285 million people globally according to estimates by the International Diabetes Federation. A compact and non-invasive monitoring system can measure blood glucose continuously without posing many problems and easy to use for the diabetic population. In this paper, a near-infrared optical sensor-based non-invasive system has been designed and calibrated against conventional invasive glucose measuring techniques. A synthetic dataset of a sufficiently large population has been developed in a suitable range as defined by the Medical Council of India. Upon development of the laboratory prototype of this device, by analyzing the variation in voltages received after reflection of incident light in the cases the approximate glucose level of the individual is going to be predicted using statistical models. A compact framework for non-invasive blood glucose measurement has been designed and tested successfully for a set of 5000 populations of fasting as well as random blood glucose samples of different patients. This paper emphasizes on different approaches using basic regression methods and classical machine learning algorithms like support vector machines, K-nearest neighbor, random forest with their hybrid regression methods for the predictions of blood glucose. Here, classical and hybrid machine learning algorithms and techniques have been applied as a measuring tool, to the large synthetic datasets to come out with laboratory-based prototypes in an attempt to automate the analysis of large and complex patient data attributes. Also, this work envisages performance optimization using grid search CV and randomized search to find the best suitable values for the hyperparameters and its impact on improvising accuracy of algorithmic models leading to more accurate prediction of blood glucose values.