Approcci di Codifica da Serie Temporali a Immagini per la Predizione dei Livelli di Glucosio nel Sangue nel Diabete di Tipo 1 utilizzando Modelli di Deep Learning

Type 1 diabetes (T1D) is a chronic autoimmune disease characterized by the destruction of pancreatic β-cells, leading to impaired insulin production and to elevated blood glucose (BG) levels. Without an accurate glycemic control, individuals with T1D can face frequent episodes of hyperglycemia (BG > 180 mg/dL) and hypoglycemia (BG < 70 mg/dL), which can result in severe complications such as cardiovascular disease, kidney failure, neuropathy, and blindness. An accurate management of T1D requires a complex and burdensome daily routine: fre- quent BG monitoring, carbohydrate counting, precise insulin administration, physical exercise and strict dietary management. Continuous Glucose Monitoring (CGM) are minimally invasive sensors that provide real-time glucose readings, typically every five minutes, as well as visual and acoustic alarms to warn patients of hypo- or hyperglycemic episodes. However, knowing ahead-in-time if glucose is approaching harmful levels would be much more helpful as it would enable proactive interventions. BG forecasting is a well-established research area in the field of T1D with a vast body of literature exploring different methodologies to improve prediction accuracy. A wide range of models, from traditional statistical approaches to advanced machine learning and deep learning techniques. Despite these advancements, the development of BG forecasting algorithms is still an open problem, and its clinical applicability remains an ongoing challenge due to factors such as variability in individual responses, inaccuracies in sensor measurements, and the difficulty in capturing the full complexity of glucose dynamics. One emerging trend in time series forecasting involves transforming sequential data into images, enabling the advanced Convolutional Neural Networks (CNN) which proves promis- ing performance in the field of computer vision techniques for improved prediction accuracy. This novel approach has already demonstrated superior performance in various fields such as financial forecasting, weather pattern classification, and medical diagnostics. This thesis explores the application of time series-to-image encoding for BG prediction in T1D. Particularly, glucose time series data are converted into images through three encoding techniques: Recurrence Plot (RP), Gramian Angular Field (GAF), and Markov Transition Field (MTF). These representations are then fed into a Convolutional Neural Network (CNN) trained to predict BG levels across various prediction horizons. To evaluate this approach, we con- duct a comparative analysis of different input lengths and prediction horizons, benchmarking a Long Short-Term Memory (LSTM) network, which directly processes glucose data. The dataset used in this research consists of CGM data from 100 virtual individuals monitored for 30 days, generated using the UVA/Padova Type 1 Diabetes Mellitus Simulator (T1DMS).