Machine Learning-Enhanced Hybrid Microwave Biosensor for High-Sensitivity Glucose Detection in Fruit Juices

Abstract

Rapid assessment of glucose in physiologically and commercially consumed liquids is important to health and food-quality monitoring. However, conventional enzyme-based methods are invasive and inefficient for continuous or real-time glucose quantification. This paper proposes a miniature U-shaped microwave biosensor integrated with an interdigital capacitor (IDC) and fabricated on a Rogers RO4003C substrate for <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in vitro</i> dielectric characterization of glucose-based liquids. To realize bandpass response and enhance selectivity, a 1.5 pF series capacitor is incorporated into the feedline. The sensor demonstrates distinct and sharp resonance dips in both reflection coefficient (S<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">11</sub>) and transmission coefficient (S<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">21</sub>), validating its strong resonant response suitable for dielectric sensing. Glucose-water solutions with concentrations of 5%, 10%, 15%, and 25% were measured, demonstrating minimal resonant frequency drift and a repeatable 0.8 dB variation in the magnitude of S<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">11</sub>, corresponding to an average sensitivity of approximately 0.05 dB/% glucose. Comparative measurements of fresh and processed mango, guava, orange, and grape fruit juices indicated that different S-parameter responses are related to the concentration of sugar and the effect of processing. For automated data interpretation, a hybrid machine learning (ML) framework combining Random Forest (RF)–based feature selection, XGBoost refinement, and artificial neural network (ANN) classification was employed for juice-type discrimination. Glucose concentration is predicted using a deep neural network (128–64–32–1), achieving MSE of 0.2297, MAE of 0.21%, and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.979, with over 95% of predictions within ±2% error, demonstrating reliable<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> in vitro</i> glucose sensing and food-quality assessment.