Physics-Informed Neural Networks vs. Physics Models for Non-Invasive Glucose Monitoring: A Comparative Study Under Noise-Stressed Synthetic Conditions

TL;DR

This study tackles non-invasive glucose monitoring under practical, noise-stressed conditions by introducing a synthetic NIR data generator that captures hardware drift, environmental variation, and physiological diversity. It benchmarks six approaches—from physics-engineered Beer–Lambert features to various PINN variants and a shallow DNN—under a fixed low-correlation regime ($ ho olinebreak= olinebreak0.21$). The Enhanced Beer–Lambert model, a physics-informed feature engineering plus ridge regression approach, achieves the best RMSE ($13.6$ mg/dL) with minimal complexity, outperforming deeper neural networks and confirming that carefully engineered physics features can surpass higher-capacity models in noisy settings. The work also provides a validated data-generation framework to support prototype development and emphasizes the importance of deployment-friendly, interpretable models for embedded wearable glucose sensing.

Abstract

Non-invasive glucose monitoring outside controlled settings is dominated by low signal-to-noise ratio (SNR): hardware drift, environmental variation, and physiology suppress the glucose signature in NIR signals. We present a noise-stressed NIR simulator that injects 12-bit ADC quantisation, LED drift, photodiode dark noise, temperature/humidity variation, contact-pressure noise, Fitzpatrick I-VI melanin, and glucose variability to create a low-correlation regime (rho_glucose-NIR = 0.21). Using this platform, we benchmark six methods: Enhanced Beer-Lambert (physics-engineered ridge regression), Original PINN, Optimised PINN, RTE-inspired PINN, Selective RTE PINN, and a shallow DNN. The physics-engineered Beer Lambert model achieves the lowest error (13.6 mg/dL RMSE) with only 56 parameters and 0.01 ms inference, outperforming deeper PINNs and the SDNN baseline under low-SNR conditions. The study reframes the task as noise suppression under weak signal and shows that carefully engineered physics features can outperform higher-capacity models in this regime.

Figures (13)

• Figure 1: Multi-criteria snapshot comparing all six models. Each axis represents a normalized performance metric, with larger area indicating better performance. The Enhanced Beer–Lambert model dominates five of six criteria while matching neural networks on inference speed.
• Figure 2: Accuracy–complexity Pareto front. X axis: physics depth ($0\!=\!$linear, $1\!=\!$RTE-inspired); Y axis: performance score ($1\!=\!$best). Bubble area $\propto$ parameter count. The Enhanced Beer–Lambert point (top–left) dominates both axes, illustrating that increased physics depth does not guarantee better field accuracy.
• Figure 3: Noise-stressed data generation framework incorporating hardware limitations, environmental variations, physiological differences, and glucose dynamics.
• Figure 4: Characteristics of the noise-stressed synthetic dataset showing (a) glucose concentration distribution, (b) NIR intensity correlation with glucose, (c) environmental condition variations, and (d) physiological parameter distributions.
• Figure 5: Detailed comparison of model architectures showing (a) Enhanced Beer–Lambert physics-based feature engineering, (b) Original PINN with Beer–Lambert constraints, (c) Optimized PINN with advanced features, (d) RTE-inspired PINN with physics-regularized light transport, (e) Selective RTE PINN with computational optimization, and (f) SDNN with shallow-deep architecture.