Noninvasive Blood Glucose Monitoring via Vein Visualization-Guided Confocal Raman Spectroscopy and Heterogeneous Ensemble Learning

Raman spectroscopy holds promise for noninvasive glucose monitoring. But its practical implementation is constrained by challenges like imprecise sampling locations, poorly resolved glucose-specific features, and small data sets. To overcome these limitations, venous visualization technology (VVT) is, for the first time, integrated with a confocal Raman spectroscopy (CRS) system to improve spectral sampling precision and quality. Moreover, a stacking ensemble learning framework combining multiple heterogeneous base models is proposed to improve model generalization while mitigating overfitting for accurate glucose prediction. Meanwhile, an internal standard method, combined with the Raman spectrum intelligent augmentation engine and a feature selection-driven stacking model, is employed to expand the effective data set size and optimize model weight allocation. For method validation, regression analysis is conducted on synthetic blood samples and in vivo venous blood glucose, with classification of samples based on varying glucose levels. Results indicate that, for in vitro synthetic blood experiment, the glucose peak area ratio is significantly linearly correlated with concentration, with the regression R2 value reaching 0.9897. For the in vivo transcutaneous test, VVT enhances the accuracy of venous localization by approximately 20%, with the regression R2 between the predicted and reference values reaching 0.928. In the classification experiment, reliable results are obtained for synthetic blood using 2 mmol/L intervals during training; when transitioned to the in vivo transdermal test, transfer learning boosts the accuracy by 2 percentage points to 92% for distinguishing between pre- and postprandial glucose states. These results establish a foundation for the application of AI-based noninvasive glucose monitoring technology.