Knowledge-guided recurrent neural networks for glucose–insulin dynamics modeling

Mathematical models of glucose–insulin dynamics are essential for managing type 1 diabetes. Their applications extend to closed-loop control, forecasting glucose trajectories, anticipating and detecting hypo- and hyperglycemia, and supporting real-time decision-making. In this work, we introduce the Compartmental Recurrent Neural Network (COMP-RNN) model, which advances the Biologically-Informed Recurrent Neural Network (BI-RNN) framework for glucose–insulin dynamics modeling. The COMP-RNN extends the data-driven strengths of the BI-RNN by embedding physiological structure directly into the model architecture. Specifically, it leverages structured RNNs aligned with canonical physiological compartments and incorporates prior physiological knowledge into training through an augmented cost function. The COMP-RNN is trained and validated on in silico cohorts. Compared to both BI-RNN and a benchmark linear model, the proposed approach achieves higher predictive accuracy and improved parameter efficiency, while better reflecting the underlying physiological principles.