GlucoBalance: Non-Invasive Blood Glucose Monitoring via Multi-Wavelength Optical Spectroscopy and Machine Learning

Abstract

Diabetes has become one of the critical health problems globally, with rapidly rising prevalence in developing regions. Nevertheless, many people are still dependent on the finger-pricking method to check blood glucose level. This conventional method is painful and inconvenient for daily use, especially for children and elderly people. To address these limitations, this paper presents GlucoBalance, a non-invasive glucose monitoring system. It uses multi-wavelength optical spectroscopy and machine learning methods to predict blood glucose level. In this device, a multi-wavelength LED array (White, Red, Green, and Infrared), and an AS7341 spectral sensor are used to detect optical signals from the skin. These signals are sent through Bluetooth Low Energy (BLE) to a Flutter-based mobile app by ESP32-C3, which then forwards the data to the model via backend for glucose prediction. With this motive, we tested two machine learning models: a Multiple Linear Regression model and a Multilayer Perceptron (MLP) Regressor. After evaluation, we observed that the MLP results were much better than the simple linear model, achieving an R² of 0.7704 and a Mean Absolute Relative Difference (MARD) of just 5.92%, compared to an R² of 0.2763 and MARD of 12.27% for linear regression. This implies that non-linear models would be more appropriate for the capture of this complicated relationship between the optical signal and glucose. Despite the success achieved by this experiment, there are certain limitations to this experiment such as the limited data and the hardware computing limit. Hence, this approach reflects that affordable, painless, and accessible glucose monitoring solutions are worth developing further.