Photoplethysmography-based glucose estimation with feature selection

Custom low-cost PPG system, 51 subjects recorded and 32 passed quality control; 14 morphological/autonomic features ranked by correlation, partial correlation, mutual information and a composite score. Same pipeline applied to a public four-wavelength PPG dataset (133 subjects, complete metadata), held-out test of 27 subjects. Greedy forward-selected PPG features plus metadata gave R² 0.410 (Random Forest, 730 nm) versus 0.355 metadata alone; gains at other wavelengths ≤ 0.024; the authors state no significance test was performed.

• Form factor: other

Diabetes mellitus remains a major global health challenge, while conventional blood glucose monitoring relies on invasive fingerstick methods that limit continuous and large-scale screening. This study investigates non-invasive glucose estimation using photoplethysmography (PPG) signals processed through an adaptive digital filtering pipeline combined with a multi-criteria statistical feature selection framework. A custom acquisition system was used to collect PPG recordings from 51 subjects, from which 14 morphological and autonomic features spanning time-domain, derivative, frequency-domain, and nonlinear characteristics were extracted. Feature relevance was evaluated using Pearson and Spearman correlations, partial correlation with confounder adjustment, mutual information, and composite ranking. The validated pipeline was subsequently applied to a public four-wavelength PPG benchmark dataset comprising 133 subjects, evaluating three progressively enriched machine learning input configurations. The proposed approach, combining subject metadata with greedy forward-selected PPG features, achieved an R 2 of 0.410 using Random Forest at the 730 nm channel, outperforming the metadata-only baseline by 0.055. All configurations maintained 100% Clarke Error Grid Zone A+B compliance, confirming clinical safety. These findings demonstrate that a compact and interpretable set of PPG-derived autonomic and morphological features can significantly enhance demographic predictors for non-invasive glucose estimation.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: http://creativecommons.org/licenses/by/4.0/

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

Wavelengths

multi (public 4-wavelength dataset incl. 730 nm)