Non-Invasive Glucose Sensing on Fingertip Using a Mueller Matrix Polarimetry With Machine Learning

Abstract

This study achieved significant predictive results using Mueller Matrix Polarimetry combined with the XGBoost algorithm for non-invasive glucose sensing of biological tissues on human fingertips. The experiment used a 660nm laser in polarimetry and incident angle optimization to enhance measurement capabilities, comprehensively obtaining properties including Linear Birefringence (LB), Circular Birefringence (CB), linear dichroism (LD), Circular Dichroism (CD), and Degree of Polarization (DoP). Phantom models simulated the interference properties of biological tissue polarization measurements. The XGBoost regression model, with feature engineering based on correlation matrices, showed consistent trends in both phantom and human measurements. As a result, the prediction results for glucose concentration in phantom mixtures were R² = 0.96 and Mean Absolute Relative Difference (MARD) = 8.67%. Furthermore, the prediction of glucose concentration on human fingertips achieved R² of 0.89, and MARD of 2.92% using the features: ${{R}_1}$, ${{m}_{32}}$, $S{{1}^{{{{45}}^\circ }}}$, $S{{2}^{{{{45}}^\circ }}}$, $S{{1}^R}$, $DoL{{P}^R}$, and ${{S}_{tol}}$. It is found that for predicting human blood glucose using machine learning, CB, CD, DoP, and Total light intensity are crucial optical properties.