Deep learning methods for improving diabetes management tools

Abstract

Diabetes is a chronic disease that is characterised by a lack of regulation of blood glucose concentration in the body, and thus elevated blood glucose levels. Consequently, affected individuals can experience extreme variations in their blood glucose levels with exogenous insulin treatment. This has associated debilitating short-term and long-term complications that affect quality of life and can result in death in the worst instance. The development of technologies such as glucose meters and, more recently, continuous glucose monitors have offered the opportunity to develop systems towards improving clinical outcomes for individuals with diabetes through better glucose control. Data-driven methods can enable the development of the next generation of diabetes management tools focused on i) informativeness ii) safety and iii) easing the burden of management. This thesis aims to propose deep learning methods for improving the functionality of the variety of diabetes technology tools available for self-management.

In the pursuit of the aforementioned goals, a number of deep learning methods are developed and geared towards improving the functionality of the existing diabetes technology tools, generally classified as i) self-monitoring of blood glucose ii) decision support systems and iii) artificial pancreas. These frameworks are primarily based on the prediction of glucose concentration levels.

The first deep learning framework we propose is geared towards improving the artificial pancreas and decision support systems that rely on continuous glucose monitors. We first propose a convolutional recurrent neural network (CRNN) in order to forecast the glucose concentration levels over both short-term and long-term horizons. The predictive accuracy of this model outperforms those of traditional data-driven approaches. The feasibility of this proposed approach for ambulatory use is then demonstrated with the implementation of a decision support system on a smartphone application. We further extend CRNNs to the multitask setting to explore the effectiveness of leveraging population data for developing personalised models with limited individual data. We show that this enables earlier deployment of applications without significantly compromising performance and safety.

The next challenge focuses on easing the burden of management by proposing a deep learning framework for automatic meal detection and estimation. The deep learning framework presented employs multitask learning and quantile regression to safely detect and estimate the size of unannounced meals with high precision. We also demonstrate that this facilitates automated insulin delivery for the artificial pancreas system, improving glycaemic control without significantly increasing the risk or incidence of hypoglycaemia.

Finally, the focus shifts to improving self-monitoring of blood glucose (SMBG) with glucose meters. We propose an uncertainty-aware deep learning model based on a joint Gaussian Process and deep learning framework to provide end users with more dynamic and continuous information similar to continuous glucose sensors. Consequently, we show significant improvement in hyperglycaemia detection compared to the standard SMBG.

We hope that through these methods, we can achieve a more equitable improvement in usability and clinical outcomes for individuals with diabetes.