Multi-Wavelength Photoplethysmography with Combined Entropy, Time, and Frequency Domain Analysis for Noninvasive...

Photoplethysmography (PPG) shows promise for noninvasive blood glucose monitoring, but current methods have yet to fully utilize signal complexity measures. Variations in blood glucose can affect blood viscosity and vessel compliance, potentially leading to subtle changes in PPG signal complexity that conventional analysis may overlook. This study integrates multi-wavelength PPG (525 nm, 660 nm, and 940 nm) with entropy analysis, alongside conventional time- and frequency-domain features. We analyze Shannon, sample, spectral, and wavelet entropy to characterize variations in PPG waveform complexity associated with different blood glucose levels. Using 312 PPG segments from eight participants (reference glucose: 83-189 mg/dL), we demonstrate that entropy features, particularly wavelet entropy from infrared PPG, significantly enhance estimation accuracy when combined with conventional features. Feature importance analysis revealed that wavelet entropy features from IR PPG were most predictive, followed by time-domain features from green PPG. Our Histogram Gradient Boosting regressor achieved an MAE of 10.07 mg/dL and RMSE of 13.3 mg/dL, with 91% of estimations in Clarke Error Grid Zone A. This performance surpasses single-domain approaches by 25-30%, highlighting entropy analysis as a key component for improving PPG-based glucose monitoring.