From Data to Decisions: Making Complex Data Science and Machine Learning Approaches Accessible for Understanding and Managing Hypoglycaemia During Exercise in Type 1 Diabetes

Abstract

Background: Exercise is essential for managing Type 1 Diabetes (T1D), yet it poses the risk of inducing hypoglycaemia, a significant barrier to physical activity for individuals with T1D. Despite advancements in continuous glucose monitoring (CGM) and insulin delivery systems, exercise-induced hypoglycaemia remains a challenge, underscoring the need for tools that can predict and mitigate this risk. Objectives: The thesis aims to (1) establish a clear definition of hypoglycaemia applicable across various CGM devices, (2) develop a user-friendly web application for analysing CGM data to understand glycaemic control around exercise, and (3) create a user-friendly, machine learning-based tool to predict the risk of exercise-induced hypoglycaemia, empowering individuals with T1D to exercise safely. Methods: The study employed a multi-phase approach, starting with an investigation into how CGM recording intervals impact hypoglycaemia detection, using data from 496 adults with T1D. This was followed by the development of Diametrics, a Dash-based web application, designed to facilitate in-depth analysis of CGM data, particularly around exercise. The final phase involved the application of the XGBoost machine learning algorithm to create predictive models for hypoglycaemia risk during exercise, utilising a large dataset of 16,477 exercise sessions from 834 participants. These models were rigorously tested for accuracy and usability, and the simplified model's predictions were visualised through the GlucoseGo heatmap, created in collaboration with the T1D community. Results: The research established that extending CGM recording intervals affects the detection of glycaemic episodes, with interpolation methods offering partial compensation for data gaps. The Diametrics application was successfully developed and validated, demonstrating reliability and ease of use in analysing CGM data around exercise. The machine learning model achieved high predictive accuracy (ROC AUC of 0.89 for the full-featured model and 0.85 for the simplified model) and was translated into the GlucoseGo heatmap, a user-friendly tool for predicting hypoglycaemia risk. Conclusions: This thesis contributes to T1D management by providing a standardised approach to defining and detecting hypoglycaemia across CGM devices, a novel web application for detailed CGM data analysis, and a machine learning-based tool for predicting exercise-induced hypoglycaemia. These advancements empower individuals with T1D to engage in physical activities with greater confidence and safety, addressing a critical barrier to exercise and improving quality of life.