Model-Based Techniques for Safety-Critical Events Detection in Type 1 Diabetes Therapy

Artificial pancreas systems, also known as automated insulin delivery systems, are emerging therapeutic options for the management of type 1 diabetes, that automatically regulate blood glucose levels. Promptly detecting malfunctions and anomalies in artificial pancreas is essential for ensuring the safety, effectiveness, and reliability of the device, ultimately improving the quality of life of individuals with type 1 diabetes and reducing their overall healthcare burden associated with diabetes management. As a matter of fact, the efficacy of glucose regulation achieved through such systems can be significantly compromised in the event of hardware failures or incorrect interactions between users and artificial pancreas itself, potentially endangering the patient’s safety. Hence, the timely and reliable detection of system anomalies and malfunctions is of critical practical importance. In this framework, this doctoral thesis focuses on the detection of anomalous events associated with the management of type 1 diabetes. These events encompass pressure-induced artifacts in glucose sensors, discontinuation of insulin delivery due to pump malfunctions, and user failures to communicate upcoming meals or physical activity to the system. All the proposed detection methodologies rely on dynamic models of the system or the type 1 diabetic user, and are designed for both real-time and retrospective detection applications. The effectiveness of these newly proposed detection strategies, along with a robustness analysis, were assessed using in-silico or real-world datasets. The first are generated through the UVa/Padova Type 1 Diabetes simulator, which has been accepted by the US Food and Drug Administration as a viable alternative to animal testing preceding human clinical trials with an artificial pancreas. The available real-world datasets have been collected through collaborations with Dexcom Inc. and Harvard University. The proposed detection strategies are specifically designed to be integrated into a multi-module architecture aimed at identifying hardware-based malfunctions and dealing with the challenges posed by the human-in-the-loop aspects.