Non-Invasive Glucose Monitoring Using mμSORS: A Population-Based Validation Study Across Diverse Patient Variability

Frequent glucose monitoring is essential for diabetes management, yet non-invasive approaches have been limited by poor generalizability across patients. This study aimed to evaluate whether a depth-resolved Raman spectroscopy approach can provide consistent glucose prediction accuracy across physiological variability.

We enrolled 200 patients with type 2 diabetes and acquired depth-resolved Raman spectra using a multiple μ-spatially offset Raman spectroscopy (mμSORS) system. An overall partial least squares (PLS) model was developed using 4768 paired venous plasma glucose and concatenated spectra at 100 μm (offset 2) and 150 μm (offset 3), and validated by subject-wise tenfold cross-validation. Model robustness was assessed through stratified analyses across physical characteristics, skin parameters, and serum biomarkers, with size-matched comparisons to control for sample size effects. Prediction accuracy was evaluated using mean absolute relative difference (MARD), root mean square error (RMSE), and consensus error grid (CEG).

The dataset covered a wide glucose range (3.8-31.6 mmol/L; 68.4-568.8 mg/dL) and substantial physiological heterogeneity. The overall PLS model achieved a MARD of 14.5% (CEG A + B: 99.4%), with no significant differences across stratified subgroups (P > .05). Stratified subgroup models performed significantly worse than the overall model (P < .05), yet after controlling for sample size, their performance was comparable with that of the overall model.

A single, overall model for mμSORS-based glucose monitoring achieves high accuracy across diverse patient subgroups, demonstrating its strong robustness. These findings support the clinical potential of depth-resolved Raman spectroscopy for non-invasive glucose monitoring.

ClinicalTrials.gov NCT05921344; https://clinicaltrials.gov/study/NCT05921344.