Leveraging Digital Twins for Personalized Optimization of Insulin Therapy in Type 1 Diabetes

Type 1 diabetes (T1D) is a chronic condition affecting approximately 9.5 million people worldwide, including nearly 2 million children and adolescents. It is caused by the autoimmune destruction of the pancreatic beta-cells, resulting in the absence of insulin secretion and chronically elevated glycemia, i.e., blood glucose (BG) concentration. To survive, people with T1D must adopt lifelong insulin therapy, requiring multiple daily actions to maintain BG within a safe range (70-180 mg/dL): prolonged high BG levels, (hyperglycemia) can indeed lead to long-term damage, dysfunction, and failure of different organs, particularly involving the cardiovascular system; while low BG levels, (hypoglycemia) can lead to short-term risk as seizures, coma, and, if left untreated, even death. Despite technological advances such as continuous glucose monitoring devices and automated insulin delivery systems, optimizing insulin dosing is still a major challenge. Insulin requirements vary widely over time, between individuals, across days, and even within the same day; driven by factors such as growth, activity, stress, illness, and hormonal changes. Clinical guidelines provide structure but cannot fully capture this variability, leaving conventional strategies too rigid to ensure truly personalized and adaptive therapy. These limitations highlight the need for innovative approaches that leverage the increasing wealth of real-world data generated by wearable devices. Digital twins (DTs) represent a promising solution: individualized computational models capable of reproducing patient-specific glucose-insulin dynamics. Unlike traditional simulators, which are limited by fixed assumptions and cannot fully replicate free-living variability, DTs are directly calibrated on patient data, thereby reducing the “reality gap.” This makes them a powerful tool to design, test, and optimize therapeutic strategies under realistic conditions. This Ph.D. thesis proposes and evaluates novel methods to improve insulin dosing in people with T1D, with DTs as the central methodological framework. Across different studies, DTs are employed to assess, personalize, and optimize insulin therapy strategies under real-world variability. The first part focuses on mealtime insulin bolus calculation, where an existing machine learning model is recalibrated using real-world data, and novel gradient boosting models are developed. The second part addresses corrective insulin boluses, introducing a series of novel algorithms of increasing complexity, ranging from heuristic rules to deep learning approaches for BG prediction, with a particular emphasis on proactive strategies aimed at anticipating and mitigating impending hyperglycemic episodes. Subsequently, an iterative optimization algorithm is presented for long-term therapy adaptation, specifically applied to pediatric patients, where physiological changes play a critical role in shaping insulin needs. The development of the insulin dosing algorithms presented before rely on ReplayBG, a physiological digital twinning tool, to construct personalized DTs grounded in real data. In the final part, the thesis also explores data-driven, black-box DTs based on generative models, designed to produce synthetic glycemic traces that preserve fundamental physiological behaviors and approximate unaccounted variability. Together, these contributions demonstrate how DTs can ground insulin optimization strategies in real data, embedding real-world variability into their design. Beyond providing a methodological framework, this thesis delivers a set of novel algorithms for mealtime and corrective bolus dosing, long-term therapy adaptation, and data-driven generative modeling constrained to preserve fundamental physiological behaviors. Collectively, these advances move a step closer to clinically relevant tools for insulin therapy, bridging the gap between technological innovation and the everyday challenges faced by people with T1D.