Integrating Biological-Informed Recurrent Neural Networks for Glucose-Insulin Dynamics Modeling

TL;DR

This study addresses the challenge of modeling glucose-insulin dynamics for Type 1 Diabetes (T1D) management, where inter- and intra-patient variability undermines classic ODE models. It proposes a Biological-Informed RNN (BI-RNN) built on a GRU architecture with physics-informed losses that encode the 5-state physiological model $y(t)=Ay(t)+Bu(t)+E$ and $\gamma(t)=Cy(t)$, plus $IOB(t)$ and $Ra(t)$ definitions. The method yields higher glucose-prediction accuracy and better reconstruction of unmeasured states than a linear baseline, validated on the UVA/Padova simulator under circadian insulin-sensitivity variations. This approach enables personalized, adaptive glucose regulation in AP systems and can be integrated with MPC to improve glycemic control in real-world settings. The results underscore the value of combining mechanistic physiology with data-driven learning for biomedical control.

Abstract

Type 1 Diabetes (T1D) management is a complex task due to many variability factors. Artificial Pancreas (AP) systems have alleviated patient burden by automating insulin delivery through advanced control algorithms. However, the effectiveness of these systems depends on accurate modeling of glucose-insulin dynamics, which traditional mathematical models often fail to capture due to their inability to adapt to patient-specific variations. This study introduces a Biological-Informed Recurrent Neural Network (BIRNN) framework to address these limitations. The BIRNN leverages a Gated Recurrent Units (GRU) architecture augmented with physics-informed loss functions that embed physiological constraints, ensuring a balance between predictive accuracy and consistency with biological principles. The framework is validated using the commercial UVA/Padova simulator, outperforming traditional linear models in glucose prediction accuracy and reconstruction of unmeasured states, even under circadian variations in insulin sensitivity. The results demonstrate the potential of BIRNN for personalized glucose regulation and future adaptive control strategies in AP systems.

Figures (2)

• Figure 1: Performance metrics comparing the BI-RNN and the linear model proposed in abuinArtificialPancreasStable2020 in simulating the test scenario. The evaluation is performed on the 10 adult patients of the commercial version of the UVA/Padova simulator. The selected network performs particularly well for adult patient number 9, with a GoF of 65.04%.
• Figure 2: Testing scenario comparison of BGL, IOB, and Ra for the adult patient number 6 (a. - worst BI-RNN result) and adult patient number 9 (b. - best BI-RNN result) of the UVA/Padova in-silico commercial cohort. The red dotted line represents the ground truth, based on the nonlinear time-variant model from UVA/Padova. The black line corresponds to predictions obtained using the proposed BI-RNN, while the blue line reflects predictions from the linear model described in abuinArtificialPancreasStable2020. Notably, the BI-RNN achieves a closer fit to real glucose data while maintaining accurate estimations for both IOB and Ra.