Multispectral PPG Sensing Framework for Noninvasive Glucose Prediction via Multi-Feature Fusion and...

Noninvasive blood glucose prediction using near-infrared spectroscopy remains challenging due to weak glucose-specific absorption, tissue variability, and environmental interference. This study proposes a multispectral photoplethysmography (PPG)-based framework integrating multi-feature fusion, local temperature and humidity compensation, and machine learning regression. The system employs four near-infrared wavelengths (940, 1300, 1550, and 1650 nm) while simultaneously measuring finger contact-site temperature (T) and local relative humidity (H) within the finger clip. A 39-dimensional feature set incorporating optical ratios, pulse-wave morphology, frequency-domain descriptors, participant characteristics, and environmental parameters was constructed. Among six optimized regression models, Extreme Gradient Boosting (XGBoost) achieved the best performance in group-isolated 10-fold cross-validation, with R² = 0.9135, root mean square error (RMSE) = 0.2991, and mean absolute relative difference (MARD) = 2.71%. Clarke error grid analysis (CEGA) showed that 97% of predictions fell within zone A and 100% within zones A and B. Feature ablation demonstrated that local temperature and humidity measurements improved prediction accuracy, with humidity providing a greater improvement than temperature. SHapley Additive exPlanations (SHAP) analysis quantified the contributions of optical, environmental, and participant-level features. A separate session-holdout evaluation yielded R² = 0.7261 and MARD = 6.07%. These findings highlight the potential of the proposed framework for noninvasive glucose prediction. However, the small, predominantly healthy cohort and limited glucose range constrain generalizability, necessitating validation in larger, more diverse populations to establish performance in previously unseen participants and assess clinical applicability.