Data-driven Anomaly Detection for Safer and More Effective Type 1 Diabetes Management

Type 1 Diabetes (T1D) is a chronic autoimmune disease that results in the destruction of pancreatic beta cells, impairing insulin production and leading to altered glucose metabolism. To help manage the disease, several devices have been developed and introduced in the market, including Continuous Glucose Monitoring (CGM) sensors and insulin pumps. While CGMs provide real-time glucose data, allowing timely interventions, insulin pumps automate insulin delivery, reducing the need for manual injections and improving glycemic control. Unfortunately, device malfunctions, such as sensor errors or pump occlusions, pose significant risks to patient safety and treatment efficacy. The ability to detect these malfunctions promptly and accurately is critical for ensuring reliable diabetes management. Beyond immediate patient safety, these faults can also alter the clinical evaluations made by healthcare providers. Since retrospective data analysis plays an important role in clinical decision-making, as it allows clinicians to assess treatment effectiveness and adjust therapies accordingly, malfunctions can lead to inaccurate insights, potentially resulting in suboptimal treatment adjustments or misinterpretation of a patient’s glycemic control. Therefore, improving both real-time detection and retrospective fault identification is essential to enhance the overall quality of care and ensure the accurate long-term management of T1D. This doctoral thesis focuses on developing data-driven anomaly detection methods to identify malfunctions in CGM sensors and insulin pumps. These methods leverage advanced techniques in machine learning, pattern recognition, and direct data-driven approaches to detect faults both real-time and retrospectively. The developed anomaly detection strategies are tested using both in-silico data and real-world data. The first are generated through the UVa/Padova Type 1 Diabetes Simulator, one of the most reliable simulators available in the literature which has been accepted by the US Food and Drug Administration as a viable alternative to animal testing preceding human clinical trials. The available real-world datasets have been collected through collaborations with Dexcom Inc. and the Department of Medicine at the University of Padova: in fact, parts of the real data employed in this thesis have been collected in the Padova Hospital during a clinical trial designed specifically for this purpose. This research aims to advance the safety and effectiveness of diabetes management technologies by enhancing the detection of sensor and pump hardware malfunctions, ultimately contributing to more reliable and efficient care for individuals with diabetes.