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

The management of Type 1 Diabetes Mellitus remains a major challenge, as it requires continuous regulation of blood glucose levels. Automated systems, such as the artificial pancreas, address this need by relying on accurate patient-specific models. While linear physiological models are popular in the literature due to their interpretability, their linear structure struggles to capture the complex dynamics of the glucose-insulin system, resulting in low prediction performance. This is particularly challenging given the significant inter-patient variability in glucose-insulin dynamics. More complex models, such as Recurrent Neural Networks (RNNs), can achieve high modeling performance, but their black-box structure lacks interpretability, reducing clinical trust and leading to possible problems for regulatory acceptance. In this work, we propose a hybrid approach using knowledge-guided learning to structure RNNs to model the glucose-insulin dynamics in diabetic patients, merging biological interpretability with high predictive performance. Furthermore, we present a meta-learning approach to address the problem of inter-patient variability. Statistical analysis will show that the proposed methodologies improve prediction performance while maintaining a high level of interpretability.