Multi-modal deep learning for non-invasive blood glucose prediction from terahertz reflection spectra: A simulation...

Non-invasive blood glucose monitoring using terahertz (THz) spectroscopy has attracted significant attention due to the sensitivity of THz radiation to water molecule dynamics and hydrogen bonding networks in biological tissues. However, the fundamental challenge lies in the weak glucose-induced spectral changes being obscured by inter-individual variations in skin morphology and hydration. In this study, we present a comprehensive simulation framework that combines a physically rigorous four-layer skin dielectric model with deep learning to investigate THz-based blood glucose prediction. The skin model incorporates a glucose-modified Double Debye relaxation model for interstitial fluid in the dermis layer, Maxwell-Garnett effective medium theory for tissue-water mixing, and the Transfer Matrix Method (TMM) for computing normal-incidence reflection spectra across 0.1-2.0 THz. A dataset of 50,000 samples was generated by systematically varying seven physiological parameters covering realistic inter-individual variability. We evaluated four convolutional neural network (CNN) architectures: three spectrum-only models (SimpleCNN, ResCNN, AttentionCNN) and a novel multi-input CNN that fuses the reflection spectrum with auxiliary skin morphology parameters. The spectrum-only models achieved limited prediction accuracy (best RMSE = 76.09 mg/dL, R² = 0.44), confirming that skin parameter confounding represents the primary bottleneck. In contrast, the multi-input CNN, which simulates the availability of auxiliary measurements from complementary modalities such as optical coherence tomography, achieved dramatically improved performance (RMSE = 33.03 mg/dL, R² = 0.89, MARD = 15.82%) with 99.7% of predictions falling within Zones A+B of the Clarke Error Grid. Sensitivity analysis and frequency importance mapping further revealed that the 0.2-0.6 THz band carries the most glucose-relevant information. These results demonstrate that multi-modal sensing, combining THz spectroscopy with auxiliary skin characterization, is essential for achieving clinically relevant glucose prediction accuracy, and provide a physics-informed framework for guiding future experimental designs.