Gaussian Process Regression for Modelling Blood Glucose Dynamics

Abstract

Abstract Type 2 diabetes is a disease characterized by poor control of blood glucose levels. Continuous Glucose Monitoring (CGM) is an increasingly popular technology for studying glucose levels and evaluating treatment effects. Although CGM technology gives potential for granular insights into disease characteristics, more can be done in terms of exploiting this rich and dense data source to the fullest. This study aims to investigate the usefulness of Gaussian process regression as a framework for modelling blood glucose dynamics. The CGM data were collected from a previous clinical trial on a cohort of overweight and obese type 2 diabetes patients. Gaussian process modelling tools were used to capture short-term and recurring trends while adjusting for long-term changes in glucose control. Results indicate that structure such as periodicity can be successfully modelled. Interpreting specific modelling results showed to be challenging due to a high degree of uncertainty in the model hyperparameters. Non-stationary models should be considered to better account for the irregular occurrence of meal-related glucose spikes and differences between day and night glycemic variability. Finally, the periodic properties of blood glucose dynamics should be further explored. :