Non-Invasive Blood Glucose Estimation on Edge Computing Devices Using Linear Regression and Deep Neural Networks

Abstract

This study presents an alternative non-invasive blood glucose estimation based on an ESP-32 microcontroller as an IoT device that serves as the main control unit and a near-infrared (NIR) sensor operating in transmission mode to detect light scattering through human tissue, then sense the voltage and current, as well as convert it into an analog value. The IoT device collaborates with a web application for real-Time collected data and trains linear regression and DNN models to predict blood-glucose levels in a prototype. We compare the results of those machine learning models by inference on-device computing. The DNN model has a 16% lower error rate than linear regression in predicting blood-glucose levels compared with readings from a glucose meter.