Predicting Upcoming Glucose Levels in Patients with Type 1 Diabetes Using a Generalized Autoregressive Conditional Heteroscedasticity Modelling Approach

Continuous blood glucose monitoring systems (CGMS) capture interstitial glucose levels at frequent intervals over time, and are used by people with diabetes and their health care professionals to assess glycaemic variability. This information helps to adjust treatment to achieve optimum glycaemic control, as well as potentially providing early warning of imminent and dangerous hypoglycaemia. Although a number of studies has reported the possibilities of predicting hypoglycaemia in insulin dependent type 1 diabetes (T1DM) patients, the prediction paradigm is still unreliable, as glucose fluctuations in people with diabetes are highly volatile and depend on many factors. Studies have proposed the use of linear auto-regressive (AR) and state space time series models to analyse the glucose profiles for predicting upcoming glucose levels. However, these modelling approaches have not adequately addressed the inherent dependencies and volatility aspects in the glucose profiles. We have investigated the utility of generalized autoregressive conditional heteroscedasticity (GARCH) models to explore glucose time-series trends and volatility, and possibility of reliable short-term forecasting of glucose levels. GARCH models were explored using CGMS profiles of young children (4 to <10 years) with T1DM. The prediction performances of GARCH approach were compared with other contemporary modelling approaches such as lower and higher order AR, and the state space models. The GARCH approach appears to be successful in both realizing the volatility in glucose profiles and offering potentially more reliable forecasting of upcoming glucose levels.