PINN based short-term glucose prediction for type 1 diabetes

Establishing a high-precision blood glucose prediction model is crucial for diabetes treatment and blood glucose management. Both physics-driven and data-driven models offer distinct advantages. Integrating information from both approaches is an effective strategy to enhance prediction accuracy and reduce computational expense. Physics-Informed Neural Networks (PINN) can incorporate physical information into neural networks, representing an emerging method for integrating physics and data. This study is the first to apply the PINN method to blood glucose prediction. Initially, a physiological model of blood glucose metabolism is established based on the principles of glucose metabolism. Subsequently, a Recurrent Neural Network (RNN) structure is designed to establish the mapping relationship between blood glucose levels and the time series of food intake and insulin administration. Building upon this, a penalty function for training the RNN is designed, which includes the physical penalty term and network penalty term. Finally, the proposed method is experimentally validated on the 30 in-silico subjects with type 1 diabetes, derived from the FDA-certified UVa/Padova Simulator. The model achieves a 30-minute blood glucose prediction accuracy of 0.27 mmol/L.