Optimization and Adaptation of Insulin Therapy Profiles in Type 1 Diabetes: A Personalized Model-Based Framework 37 views

Effective insulin therapy for people with type 1 diabetes (T1D) relies on properly tuned dosing parameters tailored to each individual’s physiology and lifestyle to achieve glycemic control targets. However, current practice consists of burdensome trial-and-error adjustments that fail to efficiently determine optimal dosing profiles considering real-world variability. This dissertation presents a model-based framework for personalizing insulin therapy using subject-specific digital twins and simulation-driven optimization strategies. The framework builds on a minimal glucose-insulin model. An identifiability analysis was conducted to determine which physiological parameters can be reliably estimated from real-world data. Structural and practical identifiability were assessed across simulated and clinical datasets, guiding the selection of parameter subsets for personalization. Results showed that while several model parameters are theoretically identifiable, only a few—such as insulin sensitivity and operating glucose level—could be consistently estimated in practice, with accuracy depending on data quality and study conditions. To support long-term personalization, the single-day model identification process was extended to a multi-day platform. This enabled the construction of digital twins that capture intra-subject variability and can simulate glucose responses over time. In addition, a validation analysis was conducted using data from a clinical trial in which participants experienced changes in automated insulin delivery modalities. The digital twins were able to reproduce glucose outcome metrics across control modes, demonstrating the potential of the replay simulation framework to predict therapy outcomes and evaluate alternative treatment options. The final component of this work introduces a therapy optimization strategy that uses replay simulations to find personalized dosing profiles. The framework was tested in silico across different scenarios: an ideal prediction case, a robustness test with increased variability, and a long-term simulation emulating a clinical protocol. In each case, optimized profiles led to improvements in glycemic control. A multi-profile extension was also implemented to account for phase-dependent insulin requirements, such as those associated with the menstrual cycle. These contributions aim to provide a foundation for decision support systems that help with insulin therapy personalization for people with T1D. By combining physiological modeling and simulation-based optimization, the framework offers a structured approach to generate therapy recommendations.

PHD (Doctor of Philosophy)

Type 1 diabetes; Digital twin; Replay simulation; Personalized medicine

English

All rights reserved (no additional license for public reuse)

2025-07-09