Deep learning in diabetes management

Abstract

Diabetes is a group of chronic metabolic disorders that affect almost half a billion people worldwide. Despite the rapid advancement of wearable devices, such as continuous glucose monitoring (CGM), maintaining blood glucose (BG) levels in a therapeutically appropriate range has been a heavy daily burden for people living with type 1 diabetes. Due to the large inter- and intra-subject variability, finding optimal personalised treatment is still an open problem. In this thesis, a wide range of novel deep learning technology is investigated to enhance diabetes management and tackle the challenges in BG prediction, glycaemic control, BG time series generation, and digital health systems. Combing the latest advances in evidential deep learning and meta-learning, this thesis proposes the Fast-adaptive and Confident Neural Network (FCNN), a novel deep learning framework for personalised BG prediction, which incorporates model confidence and enables fast adaptation to address the challenges in clinical settings. The proposed algorithm was evaluated on three clinical datasets and achieved state-of-the-art performance. Then, the physiological data measured by wearable wristband sensors were integrated into BG prediction using the FCNN framework, which significantly improved the model performance. Subsequently, deep reinforcement learning (DRL) is explored in glycaemic control. First, a novel algorithm based on double deep Q-learning is proposed to control basal insulin and glucagon delivery in the artificial pancreas for single- or dual-hormone therapy. Then, an actor-critic algorithm is applied to develop a novel insulin advisor to recommend meal insulin bolus. The results of in silico trials demonstrated that the proposed DRL control algorithms significantly enhanced the percentage of time spent in the target BG range and reduced hypoglycaemia and hyperglycaemia. Furthermore, a novel offline DRL and off-policy evaluation framework is proposed for basal insulin control, which enables DRL models to be developed in a safe and offline process. The framework was evaluated on both in silico and clinical datasets and improved various clinical metrics. To generate synthetic BG time series for data augmentation, this thesis introduces GluGAN, a novel framework based on generative adversarial networks. By integrating a supervised learning loss into adversarial training, GluGAN captures autoregressive temporal dynamics and generates high-quality synthetic BG data for the three clinical datasets. In the experiments on BG prediction, training data augmented by GluGAN significantly improved the performance of three classic data-driven algorithms. Finally, this thesis proposes a novel Internet of Medical Things (IoMT) framework for digital health systems in diabetes management. The centre of the proposed framework is an IoMT-enabled wearable wristband that comprises a low-cost and low-power system on a chip to communicate CGM and provide decision support by edge computing. In addition, a smartphone app is designed for data visualisation, while desktop and cloud platforms are proposed for data storage and model training. As a use case, an embedded BG prediction algorithm is developed through the FCNN framework and implemented on the wristband. The optimised hardware design results in extremely low energy consumption for edge inference and wireless connectivity on the wristband. The use of the IoMT framework notably improved glycaemic control in a hardware-in-the-loop in silico trial.