Personalized Non-Invasive Blood Glucose Monitor Using Machine Learning Models

Authors

  • Pradeep Kumar Anand, Dong Ryeol Shin, Mudasar Latif Memon, Ranjita Kumari

Abstract

In the past few decades have witnessed immense growth in non-invasive sensing technologies. Unfortunately, all non-invasive blood glucose monitors developed until now lacks to measure glucose value accurately. This paper presents the concept of “personalized non-invasive blood glucose monitor” to measure glucose levels accurately for all diabetic patients using different machine learning models. These models are random forest, support vector machine (SVM), multi-layer perceptron (MLP), decision tree, and adaptive boosting (AdaBoost). Our concept consists of both invasive and non-invasive sensors on a single device. Initially, during clinical trials, several patients’ blood glucose is measured both invasively and non-invasively. Then the paired data is divided into five different groups. Five machine learning models are trained for each group having paired data. Each machine learning model predicts non-invasive values accurately based on patients’ characteristics. Once, errors in predicted non-invasive values are within the acceptable error range, patient measures blood glucose by non-invasive methods only. Our concept is applied on a baseline simulation data, the MARD is reduced from 36.1% to 12.4% for adaptive boosting machine learning model. The minimum to maximum error is reduced from -221 ~ 55% to -55 ~ 48%.

Downloads

Published

2020-05-18

Issue

Section

Articles