Blood Glucose Prediction Algorithm via TimesNet-MPINN

In the task of blood glucose time series prediction, traditional data-driven models rely highly on high-quality Continuous Glucose Monitoring(CGM) data and are prone to prediction offsets or overfitting in data sparse/noisy scenarios. When the physiological dynamics equation is introduced with additional constraints, the insufficiency of data can be compensated for in the case of prior physiological knowledge, and the generalization ability of the model can be improved. This study proposes the Timesnet-MPINN blood glucose prediction algorithm based on the Bergman minimum model constraint. The specific work is to capture the rich periodic change information in the blood glucose time series data based on the TimesNet encoder. The Mixture of Expert network(MoE) is constructed to be responsible for the fitting of single-channel time series data and the correlation learning among multi-channel features respectively. Meanwhile, the Bergman minimum model constraint is used as the regularization term of the loss function to constrain the expert network to comply with physiological rules when learning the nonlinear relationship among channel features, thereby guiding the model to correctly predict the blood glucose predicted value under this constraint. The comparative experiments show that TimesNetMPINN performs best in blood glucose prediction for patients with type 1 diabetes compared to other models.