Applying Deep Learning to Track Food Consumption and Human Activity for Non-intrusive Blood Glucose Monitoring

Abstract

Blood glucose monitoring is a wide area of research as it plays a huge part in controlling diabetes and many of its symptoms. A common human disease 'Diabetes Mellitus' (DM), which is characterized by hyperglycemia, has a number of harmful complications. In addition, the low glucose level in blood caused by hypoglycemia is correlated to fatal brain failure and death. In this paper, we explore a variety of related research to have a grasp on some of the systems and concepts that can assist in forming an autonomous system for glucose monitoring, including deep learning techniques. The proposed system in this paper utilizes non-intrusive Continuous Glucose Monitoring (CGM) devices for tracking glucose levels, combined with food classification and Human Activity Recognition (HAR) using deep learning. We relate the preprandial and peak postprandial glucose levels extracted from CGM with the Glycimc Load (GL) present in food, which makes it possible to form an estimation of blood sugar increase as well as predict hyperglycemia. The system also relates human activity with decrease in blood glucose to warn against possible signs of hypoglycemia before it occurs. We have conducted 3 different experiments; two of which are comparison between deep learning models for food classification and HAR with good results achieved, as well as an experimental result that we obtained by testing hyperglycemia prediction on real data of diabetic patients. The system was able to predict hyperglycemia with an accuracy percentage of 93.2%. © 2021 IEEE.