Improving Multi-Sensor Non-Invasive Glucose Detection through AI: A Domain Generalization Approach

Accurate glucose level monitoring is crucial in diabetes management, aiming to ensure glucose levels are within a safe range and reduce the risk of complications. Inter-patient heterogeneity is one of the most important challenges to achieving accurate non-invasive glucose monitoring. This study employs meta-forests, a novel ensemble based domain generalization approach designed to address this challenge. Our technique is applied to a dataset of 54,280 data points, collected from five subjects over 10 days, using a non-invasive system that integrates near infrared (NIR) spectroscopy, millimeter-wave (mm-wave) sensing, and temperature measurements. Moreover, we significantly enhance model interpretability by incorporating Shapley additive explanations (SHAP) analyses. Importantly, our approach leads to an accuracy for the non-invasive glucose detection system that is comparable to state-of-the art methods, achieving an average root mean square error (RMSE) of 1.10 mmol/L and a mean absolute percentage error (MAPE) of 10.32% in subject-specific experiments.