The Artificial Pancreas System (APS) is a device that continuously monitors the blood glucose levels of diabetic patients in real-time and automatically adjusts insulin delivery based on CGM data. However, glucose data obtained from Continuous Glucose Monitoring (CGM) sensors typically exhibit a time delay of approximately 10 min relative to actual blood glucose levels measured by Blood Glucose Monitoring (BGM). This time delay can hinder the timely administration of insulin. To address this issue, blood glucose levels can be predicted in advance to compensate for the delay, with the predicted values complementing real-time data to improve the accuracy and efficiency of the APS. In this paper, we propose GluPatchNet, an enhanced blood glucose prediction model based on PatchTST, utilizing various combinations of input features such as blood glucose (BG), carbohydrate intake (CHO), and insulin administration. GluPatchNet combines Channel-Mixed operations with Residual-Trend Decomposition techniques to effectively model both long-term trends and short-term variations. We evaluated the blood glucose prediction performance using data generated from the UVA-Padova virtual simulator through Simglucose. Simulation results show that GluPatchNet outperforms the original PatchTST and other deep learning models, achieving the highest prediction accuracy in both single-input and multi-input scenarios.
Kim et al. (Fri,) studied this question.