Meta-Learning in Biologically Informed Recurrent Neural Networks for Personalized System Identification of Glucose–Insulin Dynamics

The management of type 1 diabetes mellitus (T1DM) represents a major challenge, as it requires continuous regulation of blood glucose (BG) levels. Automated systems, such as the artificial pancreas (AP), address this need by relying on accurate patient-specific models. Linear models are widely adopted in this context due to their simplicity and suitability for control design. However, their linear structure struggles to capture the complex dynamics of the glucose–insulin system, leading to degraded performance. This issue is further accentuated by the significant interpatient variability in glucose–insulin dynamics. A valid alternative to linear models is represented by recurrent neural networks (RNNs), which present higher expressivity while maintaining a level of complexity suitable for control purposes. However, such networks might lead to overfitting whenever the training dataset is scarce. In this work, we tackle this problem by embedding prior physiological knowledge about the glucose–insulin relation in the training procedure of RNNs, following the framework of knowledge-guided learning (KGL). Furthermore, we employ model-agnostic meta-learning (MAML) to deal with interpatient variability. A systematic results analysis is conducted using the UVA/Padova simulator to show that the proposed methodologies increase identification performance while maintaining an affordable level of complexity.