Uncertainty-aware fourier-encoded attentive Bi-GRU for non-invasive blood glucose estimation from multimodal PPG and GSR signals

Noninvasive estimation of glucose from wearable physiological signals remains challenging because sensor–glucose relationships vary across individuals and point estimates do not convey reliability. This article presents an uncertainty-aware regression framework that combines photoplethysmography and galvanic skin response. Synchronized signal segments are transformed using deterministic harmonic Fourier features and processed by a dense projection, a bidirectional gated recurrent unit with additive attention, and a multilayer regressor. Training uses a composite objective comprising mean absolute error, input-gradient smoothness, and a soft constraint on standardized outputs, while Monte Carlo dropout estimates model uncertainty during inference. The framework was evaluated on PhysioCGM recordings from 10 adults with type 1 diabetes, with continuous glucose monitor measurements used as reference values. A random sample-wise 80/20 split yielded a mean absolute error of 3.09 mg/dL, a root mean-square error of 4.72 mg/dL, and a coefficient of determination of 0.960; 96.3% and 99.1% of predictions were located in Clarke Error Grid Zone A and Zones A+B, respectively. In contrast, ten-fold leave-one-subject-out evaluation yielded a mean absolute error of 16.75 mg/dL with a between-fold standard deviation of 0.44 mg/dL and substantially weaker agreement in the hypoglycemic range. These findings indicate strong within-cohort interpolation but limited generalization to unseen individuals and underrepresented glycemic extremes. Larger participant-disjoint studies and calibrated uncertainty estimates are required before clinical or wearable deployment.