Contactless Glucose Sensing Using Miniature mm-Wave Radar and Tiny Machine Learning

Abstract

In this article, we present a contactless glucose sensing system called GlucoRadar, which leverages a miniature low-power mm-wave radar for data collection, data augmentation to boost the training data, and tiny machine learning (TinyML) for the classification of glucose concentration. The frequency-modulated continuous-wave (FMCW) radar operates in the 60 GHz band and comprises one transmitter (TX) and three receiver (RX) channels (1TX–3RX). The radar detects features of glucose aqueous solutions at a distance of 4.8 cm. The data collected by the radar is processed, and multiple features based on the peak magnitude of the spectrum and signal energy are extracted. Data augmentation is applied by adding random noise to generate additional synthetic training data. A tiny convolutional neural network (CNN) is developed to classify 16 classes of glucose concentrations in the range of 50–200 mg/dL with a fine resolution of 10 mg/dL. The tiny CNN achieves a classification accuracy of 91.4%, comprises 153,074 parameters, and occupies 598 kB of memory, making it suitable for implementation on a commercial microcontroller unit (MCU). The developed system, evaluated using in vitro tests, is promising for future wearable electronic devices.