Literature-Grounded Simulation of Non-Invasive Glucose Monitoring Using NIR Wearable Sensor Archetypes: Accuracy Metrics, Candidate Confounders, and an Adaptive Calibration Framework

Non-invasive continuous glucose monitoring remains an important technological challenge in diabetes management. This study was revised as a literature-grounded simulation and methodological proof-of-concept rather than an experimental clinical validation. The analysis used 320 independent synthetic reference–NIR glucose pairs generated from physiological ranges and performance distributions reported in the cited literature, together with an illustrative 576-point, 48-h synthetic time series. Three sensor archetypes were informed by published optical systems. Descriptive agreement was evaluated using mean absolute relative difference (MARD), Pearson correlation, Bland–Altman analysis, and the supplied Clarke Error Grid (CEG) categories. Candidate physiological and environmental factors were examined using Pearson correlations with 95% confidence intervals, raw p-values, and Holm correction. The synthetic dataset produced an overall MARD of 10.38% (SD = 7.67%), r = 0.958 (R2 = 0.918, p < 0.001), a Bland–Altman bias of 4.32 mg/dL, and 89.4%/10.3%/0.3% of records in CEG zones A/B/D, respectively. All candidate-factor correlations were small and non-significant after multiplicity correction; the largest was hematocrit (r = 0.090, 95% CI −0.019 to 0.198; Holm-adjusted p = 0.852). Differences in MARD among sensor archetypes were also non-significant (ANOVA p = 0.369; Kruskal–Wallis p = 0.543). Adaptive calibration is therefore presented as a literature-informed framework for future prospective validation, not as a performance improvement demonstrated by the present synthetic dataset. No FDA or ISO compliance or clinical applicability is claimed.