Non-Invasive Blood Glucose Estimation Using Neural Networks and NIR Optical Techniques

Diabetes mellitus represents a prominent global health concern, requiring continuous monitoring of blood glucose levels to prevent severe complications. Current glucose monitoring systems, such as capillary blood testing and continuous glucose monitors, are invasive or minimally invasive, leading to discomfort and risk of infection. Developing a reliable non-invasive method for glucose measurement would therefore represent a major challenge and advancement, significantly improving the quality of life of diabetic patients. The aim of this thesis is thus to study two non-invasive approaches for glucose level detection. The first investigates blood glucose estimation based on photoplethysmografic (PPG) signals combined with neural networks. Publicly available datasets (VitalDB and MUST) were used, which contain PPG recordings, blood glucose values and patient informations. A preprocessing pipeline was applied to first extract clean 10-second and then 90-millisecond signal segments, aligned with reference glucose values to evaluate their performances. Several neural network models were evaluated using VitalDB as training, validation and test datasets and MUST as secondary test dataset. Results obtained from MUST showed an accuracy of 74.25% within the zone A and 25.75% within the zone B of the Clarke Error Grid analysis, being clinically acceptable for the ISO 15197:2013. However, the findings underline the need for an optimized dataset to further improve accuracy and generalization. The second approach focuses on direct glucose detection in aqueous solution using near-infrared (NIR) sensor. The basic phenomenon is based on the absorption of light by glucose molecules. A study carried out by recent literature found distinct glucose absorption peaks in the NIR spectrum. Focusing on the most prominent wavelength peak, a first prototype of the optical system was designed using a 1650 nm LED and a compatible InGaAs photodiode. Due to some hardware limitations, preliminary experiments were carried out using the AFE4403EVM evaluation board from Texas instruments, combined with the custom sensors. Measurements were conducted with glucose solutions ranging from 0 g/dL to 20 g/dL in 2.5 g/dL increments and subsequently at lower concentration, comparable to those present in the human body (100 mg/dL, 150 mg/dL, 200 mg/dL and 500 mg/dL). The results showed a clear linear correlation between the photodiode signal and the glucose concentration, confirming the wavelength dependent absorption phenomenon of glucose molecules. Nevertheless, accurate calibration of the prototype was hindered by light source variability, introducing uncertainty into the measurements. Together, these studies contribute to the understanding and development of future non-invasive glucose monitoring systems. The PPG-based neural network approach demonstrates the potential of deep learning for glucose prediction from physiological signals, while the NIR-based optical experiments confirm glucose absorption in the infrared range.