Temporal Analysis of Glycaemic Variability Metrics

Introduction/Background: Glycaemic Variability (GV) is a widely used measure in managing type 1 diabetes mellitus, describing the fluctuations in blood glucose (BG) levels over time. High GV is linked to chronic complications like micro- and macrovascular diseases. Factors contributing to GV include both external factors (diet, activity, medications) and internal factors (glucose absorption, insulin sensitivity). High GV has also been linked to an increased risk of hypoglycaemia and reduced quality of life. Therefore, minimising GV is an important goal in diabetes management. GV can be measured using various statistics, such as standard deviation, coefficient of variation (CV%), glucose management indicator (GMI), which estimates lab-tested HbA1c, average daily risk range (ADRR) measuring daily risk, high BG index (HBGI) and low BG index (LBGI) for hyperglycaemia hypoglycaemia risk, J-index for glucose control quantification, time in range (TIR), time outside range (TOR), and Glycaemia Risk Index (GRI) summarising glycaemia quality, among other methods. Methods: This study analyses the OhioT1DM dataset using GV metrics from continuous glucose monitoring (CGM), employing a rolling window approach. Each metric assesses a different aspect of GV, quantifying BG control. GMI estimates average BG over 3 months, while ADRR, LBGI, and HBGI classify hypoglycaemia and hyperglycaemia risks into different levels. Additionally, the J-index assesses glucose control using mean and standard deviation, while time in range measures the duration within the target range. GRI provides a comprehensive risk summary. Analysing these metrics collectively offers insights into type 1 diabetes management for individuals using CGM and insulin pump therapy. Each statistic is calculated over a 14-day rolling window, shifting one day at a time. This method captures trends and trajectories for individual statistics effectively. Subsequently, various time series forecasting algorithms are explored including Autoregressive Integrated Moving Average (ARIMA), Seasonal Autoregressive Integrated Moving-Average with Exogenous Regressors (SARIMAX) and Support Vector Regression (SVR) to predict future values, followed by a comparative assessment of these algorithms. Results: Subjects present conflicting results in various diabetes management statistics. While some show GMI within the target range, indicating satisfactory medium to long-term control, the J-index and HBGI suggest inadequate control and high hyperglycaemia risk. This demonstrates the necessity for a comprehensive assessment of metrics for diabetes control evaluation. The conflicting results might stem from statistical biasness towards hypoglycaemia or hyperglycaemia. GRI resolves this by combining both risks of hypoglycaemia and hyperglycaemia. Additionally, analysis by rolling window reveals trends towards an increased risk of hypoglycaemia and hyperglycaemia among specific subjects. Closer examination of the trend lines demonstrates similar trajectories between several metrics. Conclusion: This study holds potential to significantly influence diabetes self-management by offering valuable insights into disease management. Employing various measures of GV allows for a comprehensive analysis of BG control and enhances the understanding of self-management practices. The utilisation of a rolling window not only reveals trends and trajectories but also aids in predicting future values and assessing the risk of complications among individuals with diabetes. The comparison of foundational forecasting algorithms serves as a crucial basis for further investigations and analyses in the respective field.