A physics-informed glucose-insulin neural network model for glucose prediction

Accurate glucose prediction plays an important role in glucose management and closed-loop insulin delivery for subjects with diabetes. Due to its powerful data mining capability, neural networks are used to grasp the glucose trends from continuous glucose monitoring (CGM) data. However, this approach requires a large number of individual data and plentiful computing resources, but has relatively poor interpretability. Given that the glucose metabolism mechanism model contains abundant physiological information, fusing the information with neural networks can reduce the demand for data and computing resources, and improve interpretability. In this study, a physics-informed glucose-insulin neural network (PIGNN) model is proposed, of which the structure and loss function are designed based on the glucose-insulin dynamic model. According to the experiments of 22 real subjects and 30 in-silico subjects with type 1 diabetes, the prediction accuracy of this method achieves 0.726 ± 0.126 mmol/L. Compared with models without physical information, the proposed PIGNN shows a significant improvement in glucose prediction for real subjects with a limited sample size (only 48 data samples), resulting in an accuracy improvement of 12.33%. In addition, it is proved that with limited data more physical information can improve the glucose prediction accuracy significantly.