Deep Learning Algorithms for Blood Glucose Forecasting in Type 1 Diabetes

Type 1 Diabetes (T1D) is a chronic metabolic disease characterized by elevated blood glucose (BG) concentrationdue to the immune system's destruction of insulin-producing $\beta$-cells in the pancreas. The standard treatment for lowering BG levels involves the injection of exogenous insulin, with dosages manually computed by the patients based on the current BG levels, recent carbohydrate (CHO) intakes, and previous insulin administrations. However, this approach overlooks other critical factors influencing glucose-insulin dynamics, such as physical activity, stress, and illness, which are more challenging to quantify. As a result, these simplifications contribute to frequent episodes of hypoglycemia ($BG<70$ mg/dL) or hyperglycemia ($BG>180$ mg/dL). While hyperglycemia may not pose an immediate threat, it can lead to severe long-term complications. On the other hand, hypoglycemia can cause immediate and potentially life-threatening issues. Without prompt treatment using fast-acting CHOs, commonly known as hypotreatments, hypoglycemia can quickly escalate to an acute medical emergency, such as seizures, sudden loss of consciousness, or coma, which may be fatal. Effective management of T1D greatly benefits from constant monitoring of BG levels, which can be achieved in real-time using minimally-invasive continuous glucose monitoring (CGM) devices. Forecasting algorithms can be employed to predict BG levels minutes in advance, enabling timely and proactive therapeutic interventions to prevent or minimize adverse events. Deep learning (DL), with its capability to autonomously learn complex non-linear dependencies, represents the state-of-the-art in forecasting BG levels for individuals with T1D. However, its intrinsic black-box nature poses significant challenges for its adoption in clinical practice. The aim of this thesis is manifold: \begin{enumerate} \item To conduct a systematic literature review (SLR) on DL algorithms for BG forecasting in T1D. \item To perform a comprehensive comparative analysis of various DL models and preprocessing strategies across three distinct datasets, aiming to address gaps identified in the SLR. This analysis will evaluate the effectiveness of both commonly used and promising neural network (NN) architectures highlighted in the SLR, focusing on prediction accuracy and physiological fidelity. Particular emphasis will be placed on replicating and assessing the CNN-LSTM model proposed by Jaloli et al. \cite{Long-Term_Prediction}, which is claimed to offer significantly superior performance compared to other state-of-the-art models. \item To develop a new explainable DL model, referred to as Monotonic-NN, that aligns with physiological principles. This model is designed so that predictions increase with the amount of ingested CHOs and decrease with the amount of administered insulin, aiming to overcome the limitations of current state-of-the-art models. \end{enumerate}