Abstract
Monitoring of fasting blood sugar (FBS) is a critical component in the diagnosis and management of diabetes, one of the most widespread chronic diseases globally. Microwave sensing-particularly through microstrip-based sensors-has recently gained attention as a promising technique for blood glucose monitoring, offering advantages such as low cost, high sensitivity, real-time response capability, and suitability for compact and wearable systems. In this study, a miniaturized microstrip microwave sensor is presented for non-contact FBS detection. Blood samples were directly collected from 78 individuals and analyzed using a clinical-grade auto-chemistry analyzer to determine reference FBS levels. Each sample was measured five times on the microwave sensor, resulting in a total of 390 transmission responses (S21) across a frequency range of 30 kHz to 18 GHz. These responses were recorded under controlled laboratory conditions, ensuring consistency and minimizing environmental interference. To interpret the complex, non-linear features of the sensor response, a convolutional neural network (CNN) was developed and trained using the entire dataset. The network demonstrated highly promising performance in estimating FBS values, achieving a mean relative error (MRE) of 1.31%. The results confirm the feasibility of combining broadband microwave sensing with deep learning techniques to enable reliable non-contact blood glucose measurement. This approach holds strong potential for integration into future wearable health monitoring systems, providing more user-friendly diabetic management tools without the frequent use of conventional blood sampling techniques.