# Investigation of Printed Slot Antenna for Non-Invasive Glucose Sensing Using FR4 Substrate Material

## TL;DR

This study explores using a printed slot antenna to measure glucose levels non-invasively by analyzing microwave signals affected by a finger's dielectric properties.

## Contribution

The paper introduces a novel RF-based non-invasive glucose sensing framework using machine learning for calibration.

## Key findings

- Machine learning models, especially Random Forest, outperformed linear models in predicting glucose levels from RF data.
- The antenna's resonant frequency and reflection coefficient changes correlate with glucose values under controlled conditions.
- Nonlinear models achieved an R2 score of 0.72 with a mean absolute error of 5.16 mg/dL.

## Abstract

This paper provides a feasibility study of a non-invasive microwave-based glucose-sensing system based on a small printed slot antenna with etched step-impedance resonators (SIRs) on an FR4 substrate in the ground plane at approximately 5.7 GHz. The sensor proposed takes advantage of the effect of the antenna resonant frequency and reflection coefficient (S11) perturbation due to the dielectric loading of a human finger placed in the antenna near field. Instead of declaring direct glucose specificity, this paper is dedicated to understand whether the measures of RF can be translated to the invasive glucose values under the condition of controlled positioning. A vector network analyzer was used to measure the experimental values where resonant frequency and S11 magnitude were obtained at the point of peak sensitivity due to fixed finger placement at the point. These RF properties were associated with invasively measured glucose values using three modeling methods: a simple analytical linear formula, a second-degree Polynomial Ridge regression model, and a Random Forest machine learning model. The comparative analysis has established that nonlinear data-driven models outperform the analytical formulations significantly with the highest predictive accuracy being the Random Forest model (R2 = 0.72, RMSE = 10.57 mg/dL, MAE = 5.16 mg/dL). The findings affirm that the impacts of antenna loading control the raw measurements, but the trend related to glucose can be extracted upon machine learning calibration under controlled conditions. The research provides a methodological framework of RF-based non-invasive glucose sensing and the need to employ various phantom-based validation, sub-subject-based modeling, or clinically based evaluation metrics in future studies.