Machine learning–based non-invasive glucose monitoring via dual-signal acquisition

Abstract

Blood glucose monitoring remains a critical component of diabetes management, yet the discomfort and compliance challenges associated with conventional finger-prick methods have long prompted researchers to explore less invasive alternatives. This project comes in response to that sustained demand with a dual-sensor signal acquisition and machine learning-based estimation system that will be a potentially more accessible and patient-friendly way to monitor glycaemia, without requiring invasive blood sampling. The system is based on two different and complementary sensing modalities: Near-Infrared (NIR) spectroscopy and Galvanic Skin Response (GSR). The light transmitted through NIR reflects the biochemical composition of glucose molecules in the subcutaneous tissue, whereas the changes in GSR are related to the electrodermal changes related to the physiological fluctuations associated with glucose dynamics. The dual-acquisition system is not meant to use either signal alone but in an attempt to enhance the robustness of the estimation by utilizing the complementary information contained in each. A custom dataset collected through extensive fieldwork and on-site data acquisition is at the core of the system's functioning, designed to reflect the natural variability in physiological data. It is generally believed that training a machine learning model using domain-specific data of this type is more reliable than adapting generalized benchmarks, especially for glucose estimation due to the sensitivity of these estimations to individual biological variations. The model learns how to relate the output of the various sensors to the reference glucose values and is able to predict glucose values that are based on empirical patterns rather than purely theoretical proxies. The system is implemented to be constantly monitored, and it can be viewed via a web-based interface, without the need for repeated invasive procedures to monitor glucose trends. Although additional clinical testing will be needed before it could be introduced clinically, these results indicate that this sensor fusion method has significant potential as a non-invasive alternative, especially for people who need to test glucose levels more often.