Stochastic seasonal models for glucose prediction in type 1 diabetes

Resumen

Diabetes is a significant global health problem, one of the most serious noncommunicable diseases after cardiovascular diseases, cancer and chronic respiratory diseases. Diabetes prevalence has been steadily increasing over the past decades, especially in low- and middle-income countries. It is estimated that 425 million people worldwide had diabetes in 2017, and by 2045 this number may rise to 629 million. About 10% of people with diabetes suffer from type 1 diabetes, characterized by autoimmune destruction of the beta-cells in the pancreas, responsible for the secretion of the hormone insulin. Without insulin, plasma glucose rises to deleterious levels, provoking long-term vascular complications. Until a cure is found, the management of diabetes relies on technological developments for insulin replacement therapies. With the advent of continuous glucose monitors, technology has been evolving towards automated systems. Coined as "artificial pancreas", closed-loop glucose control devices are nowadays a game-changer in diabetes management. Research in the last decades has been intense, yielding a first commercial system in late 2017 and many more are in the pipeline of the main medical devices industry. However, as a first-generation device, many issues still remain open and new technological advancements will lead to system improvements for better glycemic control outputs and reduced patient's burden, improving significantly the quality of life of people with type 1 diabetes. At the core of any artificial pancreas system is glucose prediction, the topic addressed in this thesis. The ability to predict glucose along a given prediction horizon, and estimation of future glucose trends, is the most important feature of any artificial pancreas system, in order to be able to take preventive actions to entirely avoid risk to the patient. Glucose prediction can appear as part of the control algorithm itself, such as in systems based on model predictive control (MPC) techniques, or as part of a monitoring system to avoid hypoglycemic episodes. However, predicting glucose is a very challenging problem due to the large inter- and intra-subject variability that patients suffer, whose sources are only partially understood. These limits models forecasting performance, imposing relatively short prediction horizons, despite the modeling technique used (physiological, data-driven or hybrid approaches). The starting hypothesis of this thesis is that the complexity of glucose dynamics requires the ability to characterize clusters of behaviors in the patient's historical data naturally yielding to the concept of local modeling. Besides, the similarity of responses in a cluster can be further exploited to introduce the classical concept of seasonality into glucose prediction. As a result, seasonal local models are at the core of this thesis. Several clinical databases including mixed meals and exercise are used to demonstrate the feasibility and superiority of the performance of this approach.