Open-loop insulin dosing personalization in type 1 diabetes using continuous glucose monitoring data and patient characteristics

Patients with type 1 diabetes (T1D) require lifelong insulin therapy in order to maintain their blood glucose (BG) concentration within the euglycemic range preventing long-term complications associated with hyperglycemia and avoid- ing dangerous episodes of hypoglycemia. To achieve proper glycemic con- trol, people with T1D need to perform a constant learning process about how daily conditions (e.g. insulin administrations, meals schedule and composi- tion, physical activity, and illness) affect BG levels. More than 500,000 op- erations can be needed during the lifetime of a T1D patient to manage the therapy. For this reason, management of diabetes is burdensome for patients, and results in deteriorating their quality of life. One of the major issues in the daily management of T1D concerns with the amount of insulin that has to be administered, by a subcutaneous bolus injection, in order to compensate the increase of BG associated with meals. So far, a standard simple mathematical formula (SF), designed by clinical investigators on an empirical basis, is com- monly used by patients to calculate the size of insulin boluses. SF leverages on the current BG level obtained from self monitoring blood glucose (SMBG) samples, the estimated amount of carbohydrates (CHOs) present in the meal, and patient specific therapy parameters. While the SF is well-established in clinical practice, the insulin amount determined through its use could be sub- optimal due to several reasons, including the error patients make in estimating CHO, the intrinsical sparseness of SMBG, and the inability of accounting for many important factors such as patients’ intra-/interday variability. Margins of improvement over the SMBG-based SF emerged in the past decade, when diabetes management has been transformed by the introduction of min- imally invasive continuous glucose monitoring (CGM) sensors, which have been recently approved by regulatory agencies, such as the Food and Drug Ad- ministration (FDA), to be usable to make treatment decisions, such as insulin dosing. Of course, CGM provides an increased amount of available features on BG, such as the rate of change (ROC), that could be exploited to improve insulin standard therapy. As a matter of fact, several attempts have been pro- posed in the literature to account for CGM-derived information and adjust SF accordingly, but unfortunately, they fall short in personalizing such an adjust- ment patient-by-patient. In this thesis we propose new methodologies for determining a dose of in- sulin bolus able to effectively account for the "dynamic" information on BG provided by the ROC and patient characteristics, the final aim being to per- sonalize the standard insulin therapy and eventually improve the glycemic control. In particular, to identify the possible margins of improvement, in the first part of the thesis we assess and analyze the criticalities of three popular literature techniques that exploit the ROC magnitude and direction to adjust the insulin bolus amount computed through SF. To such a scope, we designed ad-hoc in silico clinical trials implemented using a popular powerful simula- tion tool, i.e. the UVa/Padova T1D Simulator. Then, in the second part, we propose two novel machine learning based algorithms that, being fed by in- formation on current patient status and characteristics, provide patients with new tools to adjust SF in a personalized manner. Finally, in the third part of the thesis, we abandon the idea of using the insulin bolus provided by SF as a sort of initial estimate to be simply adjusted, and we design a brand new formula for insulin bolus determination that naturally takes into account for CGM-derived information and current patient status and characteristics. This represents an innovation in the literature because no insulin bolus formulae specifically designed for use with CGM have been proposed yet.