Gaussian Mixture Models to learn individualized eating behavior in patients with type 1 diabetes

Type 1 diabetes (T1D) is a chronic disease characterized by the absence of endogenous insulin production, whose occurrence is increasingly widespread in the population. To date, controlled simulation environments in which Decision Support Systems (DSS) are validated are simpler than in real life. Consequently, the optimal results observed in such environments are often not reflected in the performance of the DSS when used in real life setting. The thesis work is placed in this context and aims to characterize and develop eating behavioral models in individuals with T1D, in real-life conditions, using individualized Gaussian Mixture Models (GMMs). The goal is therefore to develop an eating model which makes the simulation environment more complex, whose validation was carried out using the UVA/Padova simulator, i.e. the first tool approved by the Food and Drug Administration (FDA) as a substitute of preclinical trails. In this regard, after contextualizing the thesis work in chapter 1, chapter 2 presents the exploratory analysis conducted on a dataset of 100 subjects, the object of study in this thesis. Chapter 3 discusses the learning of their eating behavior using Gaussian Mixture Models (GMMs), whose hyperparameters were set through a Matlab Graphical User interface (GUI). This process returned a set of GMMs for each of which, using the leave-one-out approach, the GMM with the most similar eating pattern (target subject) was identified among the remaining ones, using the Wasserstein Type Distance in the space of GMMs. In chapter 4, after identifying the stable matches, the representative eating prototypes of the studied population are determined. Chapter 5 presents the implementation of the generators with which meals are extracted from the identified eating prototypes, i.e. the inputs of the UVA/Padova simulator. Once the glycemic profiles were simulated, the glycemic control metrics are then computed for both real and simulated subjects. After identifying the most similar simulated subject to each real subject in terms of metrics, a virtual population whose behavior is similar to that of the real one was thus obtained. Chapter 6 discusses the results obtained from the application of statistical tests to both the real and simulated populations of each metric. Some suggestions from which to start are therefore provided for future studies, so as to expand this thesis work.