Data-driven models for type 1 diabetes using generative deep learning

Mujahid, Omer
Share
Modeling biological systems has always been challenging given the complexity of the processes involved in them. Experts have been employing physiological models to approximate the dynamics of biological systems; however, these models are constrained by the limitations of mathematical techniques that can only encompass part of the physical phenomena behind a biological system. Mathematical physiological models of the human glucose-insulin system are considered the gold standard of simulators in type 1 diabetes (T1D) healthcare. Though accurate to a certain degree, these models are not capable of simulating scenarios that could fully capture the real-life dynamics of a T1D patient. The underlying cause for this phenomenon could be attributed to the numerous hidden factors that are ignored during physiological modeling because of increasing model complexity or hurdles in their representation. This work was carried out with a focus on accurate model approximation in T1D. The rationale is built on the hypothesis that generic function approximators such as deep neural networks (DNNs) have the ability to learn all that from data that cannot be modeled mathematically. Since deep generative models (DGMs) are implemented using DNNs, they are capable of learning the underlying probability distribution of a data set. This thesis presents several methodologies based on data-driven models using DGMs for improved model approximation in T1D. Firstly, a systematic review of data-driven models for predicting hypoglycemia is conducted. After that, a methodology for data augmentation in a hypoglycemia classifier using a generative adversarial network (GAN) is developed as part of this thesis. The next work in this series focuses on the conditional synthesis of realistic BG profiles of T1D patients. Finally, building on the work performed thus far, a T1D simulation environment is developed using a sequence-to-sequence GAN (S2S GAN) that is capable of synthesizing realistic patients with T1D. The results obtained from these methods show the efficacy of DGMs for model formation in T1D. It has been demonstrated through these results that a highly precise approximation of the glucose-insulin system of patients with T1D can be obtained from data with the help of DGMs. Moreover, these models have been shown to generate novel data that is statistically similar to real data for all the standardized glycemic metrics. Furthermore, the causal synthesis of realistic T1D data has been shown in the work presented in this thesis. ​
​ADVERTIMENT. Tots els drets reservats. L'accés als continguts d'aquesta tesi doctoral i la seva utilització ha de respectar els drets de la persona autora. Pot ser utilitzada per a consulta o estudi personal, així com en activitats o materials d'investigació i docència en els termes establerts a l'art. 32 del Text Refós de la Llei de Propietat Intel·lectual (RDL 1/1996). Per altres utilitzacions es requereix l'autorització prèvia i expressa de la persona autora. En qualsevol cas, en la utilització dels seus continguts caldrà indicar de forma clara el nom i cognoms de la persona autora i el títol de la tesi doctoral. No s'autoritza la seva reproducció o altres formes d'explotació efectuades amb finalitats de lucre ni la seva comunicació pública des d'un lloc aliè al servei TDX. Tampoc s'autoritza la presentació del seu contingut en una finestra o marc aliè a TDX (framing). Aquesta reserva de drets afecta tant als continguts de la tesi com als seus resums i índexs.