A neural-ODE based continuous glucose monitoring measurements forecasting approach with knowledge distillation

Abstract Background: With the increasing use and growing popularity of insulin pumps and continuous glucose monitoring(CGM) devices, more reliable sources of data could be obtained to lead to better glucose control, treatment decision and diabetes management. The critical challenge of diabetes management is to develop accurate algorithms that can predict future blood glucose trends and manage the amount of insulin. However, in reality, continuous glucose monitoring(CGM) data is unavailable and difficult to acquire in some scenarios: for example, most type-2 diabetic patients only measure their finger sticks which are far more sparse than CGM. Few work studies glucose prediction when such continuous glucose monitoring (CGM) data is unavailable. Therefore, in this paper, we investigated the possibility of predicting the dense continuous glucose monitoring (CGM) levels purely based on sporadic self-monitoring signals such as finger sticks. Results: We proposed a novel neural ordinary differential equation (neural ODE) based method BGKD-ODE for forecasting continuous glucose monitoring (CGM), and the fundamental idea of this method is to use neural networks to learn the underlying dynamics of blood glucose metabolism in a data-driven manner. Since most of the machine learning methods can not handle such sparse data, to alleviate the training difficulty due to the data sparsity, we integrate the expert knowledge of the existing physiological model into our model. Experiments on real-world data demonstrate the strength of BGKD-ODE in accurately predicting long-term continuous glucose monitoring (CGM) levels only based on sporadic finger sticks. Conclusion: Our proposed method BGKD-ODE could be employed to obtain better insights into foresting blood glucose trends of T1DM patients. Furthermore, in the future, this model can be extended easily to type 2 diabetes with modification on reliable field knowledge.