Forecasting and decision support for type 1 diabetes insulin therapy using machine learning

Resumen

Insulin therapy for Type 1 Diabetes (T1D) has several ramifications with different degrees of automation. The advances in sensors and monitoring devices have led to an increasing availability of data. Additionally, machine learning algorithms usage has sprung, allowing the development of models for Blood Glucose (BG) forecasting with relative ease. Nevertheless, BG forecasting is still a challenging task for prediction horizons beyond 30 min and, even more so, with missing or erroneous data, which is a common burden in the field. This thesis is devoted to generate machine learning models that forecast either BG levels using regression algorithms or postprandial hypoglycemia using classification algorithms. The application of these models range from Multiple Daily Injections (MDI) therapy up to Sensor Augmented Pump (SAP) therapy. On one hand, this work focuses on the prediction of BG values by proposing a hybrid model that uses Grammatical Evolution (GE), an insulin on board model, and a glucose rate of absorption model to predict BG values with a prediction horizon of 120 min. The algorithm relies on the construction of a set of rules that determine the search space for an optimization algorithm based on a Genetic Algorithm (GA). A glucose-specific fitness function leads the evolution of the solution while penalizing deviations based on their clinical harmfulness and a tailored evolutionary grammar. On the other hand, this study delves into the methods to forecast hypoglycemic events aiming to contribute to decision-making tools in T1D therapy. For this reason, a method for training classification models that predict postprandial hypoglycemia is also proposed and validated for MDI and SAP applications, using real patients’ data in free-living conditions. The method relies on well-known machine learning algorithms and, in some cases, a combination of them to anticipate hypoglycemia using an entirely data-driven approach with carbohydrate content estimation, insulin bolus and BG level as common inputs. The aforementioned approaches are evaluated using clinically meaningful metrics that provide insights regarding the practical use of the proposed methods. The obtained results are promising and contribute to the advances in the development of technologies for the management of type 1 diabetes.