Non-invasive blood glucose detection system using combined visible and near-infrared spectral data
Abstract This pilot study evaluated non-invasive estimation of continuous glucose monitoring (CGM)-reported glucose from fingertip visible–near-infrared (VIS–NIR) reflectance in ten adults. Optical measurements (approximately 12 per minute) were retained as separate 70-feature vectors and paired with 1-minute FreeStyle Libre 2 CGM readings checked approximately 10 min afterward using clock-based timing; multiple optical records therefore shared one CGM label. The dataset contained 6,816 optical samples. Seven regression models were compared using sample-level five-fold and subject-independent leave-one-subject-out (LOSO) cross-validation. Under LOSO, Random Forest achieved MAE 11.248 mg/dL, RMSE 17.326 mg/dL, MARD 10.379%, and R² 0.767. In Clarke Error Grid analysis against CGM values, 82.40% and 85.30% of predictions were in Zones A and A + B, respectively. LOSO DNN classification accuracies were 92.74% and 92.43% for 30 and 15 mg/dL categories. These findings support feasibility but indicate inter-subject variability. The manually timed 10-minute offset was approximate and not an individually validated physiological lag correction. No independent capillary or laboratory blood glucose reference was available; thus clinical blood glucose measurement accuracy or safety was not established.
Authors
- Byoung Ik Kim
- Tae-Wuk Bae (ORCID: https://orcid.org/0000-0001-5546-9111)
- Kee-Koo Kwon (ORCID: https://orcid.org/0000-0002-2417-1256)
- Young Choon Kim
Institutions
- Electronics and Telecommunications Research Institute (KR)
- Samjin Pharm (South Korea) (KR)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1038/s41598-026-80548-4
- Primary Topic
- Optical Imaging and Spectroscopy Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Electronics and Telecommunications Research Institute