Enhancing glucose measurement efficiency: a non-invasive approach using machine learning

Abstract This research introduces a cutting-edge approach to glucose monitoring, which is essential in many applications. The study developed a new non-invasive glucose monitoring system utilizing machine learning techniques. This system examines the reflection coefficient data gathered from glucose solutions using a Vector Network Analyzer. To showcase the system’s accuracy in predicting glucose levels, two distinct datasets were employed. The first dataset comprised glucose solutions with concentrations spanning from 0 to 200 g/L, while the second dataset included solutions ranging from 15 000 to 20 000 mg/L for enhanced precision. The system measured both datasets, and three machine learning algorithms – Decision Tree, Random Forest, and Support Vector Regression – were applied to the collected data. Furthermore, a grid search method was employed to optimize the hyperparameters for each model’s optimal performance. The findings revealed that the Random Forest yielded the best results across both datasets. For gram scale, the R 2 value was 0.9995, indicating that 99.95% of the glucose level variance was accounted for, with a low RMSE of 1.1589 mg/dL. Moreover, in milligram scale dataset, the R 2 value was 0.9932, and RMSE was 1.1119 mg/dL, confirming the model’s high accuracy. These experimental outcomes demonstrate that the proposed system can effectively predict glucose levels.