Photoplethysmography-based glucose estimation with feature selection

Abstract

Diabetes mellitus is a major global health challenge, and routine blood glucose monitoring still depends on invasive fingerstick sampling. This study examines non-invasive glucose estimation from photoplethysmography (PPG) using a signal quality pipeline and a multi-criteria feature selection framework. PPG recordings from 51 subjects were collected with a custom low-cost system, and 32 passed quality control. Fourteen morphological and autonomic features were extracted and ranked using Pearson and Spearman correlation, partial correlation with confounder adjustment, mutual information and a composite score. The same processing was then applied to a public four-wavelength PPG dataset of 133 subjects with complete metadata, comparing three machine learning input configurations on a held-out test set of 27 subjects. Adding greedy forward-selected PPG features to subject metadata gave a coefficient of determination of 0.410 with Random Forest at 730 nm, compared with 0.355 for metadata alone. Gains at the other wavelengths were at most 0.024, and no significance test was performed. Agreement with the ISO 15197:2013 criterion ranged from 74% to 85%, below the 95% the standard requires. A compact set of amplitude-variability and spectral features may add modest information to demographic predictors, but this needs confirmation in larger cohorts, with longer recordings and independent validation.