Studies on Genetic Programming Techniques for the Short and Medium Term Predictions of the Interstitial Glucose of Diabetic Patients

Abstract

Diabetes Mellitus (DM) is a chronic disease that increases the morbidity and mortality, and causes a significant deterioration in the quality of life. There are mainly two types of diabetes: • Type 1 Diabetes Mellitus (T1DM): due to an autoimmune process, the pancreas is not able to generate enough insulin to process the sugar produced after carbohydrate intake. • Type 2 Diabetes Mellitus (T2DM): the insulin produced by the pancreas does not work properly, in a phenomenon known as insulin resistance. T1DM can only be treated with synthetic insulin injected into the bloodstream. Depending on the amount of insulin in the body two scenarios can happen. On the one hand, an excess of insulin can cause hypoglycemia, defined as a Blood Glucose (BG) value of less than 70 mg/dl. If this situation continues over time, it can cause short-term complications. On the other hand, if the insulin dose is insufficient, it can lead to hyperglycemia, defined as a BG value greater than 180 mg/dl, which can lead to long-term complications. The goal is to keep BG levels within the target range most of the time (defined as a BG value between [70, 180] mg/dl). BG control in insulin-dependent patients requires predicting future glucose values to determine the amount of insulin to inject. This amount depends on many factors that People with Diabetes (PwD) have to estimate manually, following two different primary therapies: Continuous Subcutaneous Insulin Infusion (CSII) and Multiple Doses of Insulin (MDI). One of the most important factors is Glucose Variability (GV), defined as the fluctuation of glucose levels during a day. GV makes the prediction process more complicated. There are different types of BG control strategies (Manual, Semi-automated and Automated solutions), but in all of them, it is important to develop mathematical models or Artificial Intelligence (AI) systems that describe the interaction between the glucose system and insulin using the measurements and stored data. Different ways to produce models for BG prediction are used depending on the information available at the time of the forecast. One option is the What-if scenario, where the model predicts future glucose values by taking into account not only past values of the three variables (glucose, carbohydrates, and insulin) but also future carbohydrate intakes and insulin injections from the present until the prediction horizon. For instance, the model can predict the BG level, supposing that the patient eats a certain amount of carbohydrates m minutes from the prediction time. Another option is the Agnostic scenario, that produces model without information about future events in the prediction phase. This type of model needs to implicitly predict those events. For example, the model must identify the fasting periods or the physical exercise. Both scenarios are used in this thesis. There are three main methods commonly used to solve the glucose prediction problem. The first, called Physiological models, are linear mathematical models that simulate the physiology of the glucose-insulin regulatory system, and require specific Physiological knowledge. The second method is Data-driven models, which can predict glucose concentration based only on existing input and output data. The last method combines both solutions in a Hybrid way, where the models take the simplest parts of glucose-insulin physiology and include data to determine the parameters of the models. All the models created in this thesis are based on Data-driven method.