Data-driven nonlinear control of glucose-insulin dynamics in patients affected by Type I Diabetes

Diabetes is one of the top ten causes of death in adults and, in the last decades, became a global health problem. People with T1DM suffer from a metabolic disorder characterized by pancreas inability to produce a sufficient amount of insulin, the peptide hormone that plays a vital role in regulating the way cells absorb glucose and use it as an energy source. As a result, these individuals need to constantly control their blood glucose levels through insulin administration. In recent years, significant technological progress in continuous glucose monitoring and insulin delivery systems has enabled researchers to develop automated methods for managing diabetes, often referred to as the Artificial Pancreas. The development of control algorithms for this purpose is a highly active field of research. While traditional control approaches have been the primary focus up until now, machine learning (in particular, neural networks) seems to offer a promising alternative framework that has not yet been thoroughly explored. This thesis focuses on data-driven approaches, with the aim of designing controllers, both linear and nonlinear, starting from the input-output data, collected from the patient, bypassing the modeling step. The work is divided into two main sections. The first one revolves around the development of two linear controllers employing the Least squares and the Set-membership methods. The second, instead, concerns the design of a nonlinear controller taking advantage of recurrent neural networks. Through extensive testing and validation, the results demonstrate that, while linear controllers perform adequately in maintaining safe glucose levels, the nonlinear controller significantly outperforms them. The neural network-based approach provides more accurate and responsive insulin delivery, offering a superior and more robust solution for AP systems. These findings highlight the potential for advanced neural network algorithms to become a new milestone for diabetes management through improved automated control mechanisms.