Plasma Glucose Reconstruction Based on Data Driven Real-Time Monitoring of CGM Devices

Real time and accurate monitoring of plasma glucose concentration in diabetes patients is necessary for their condition management. Continuous glucose monitoring (CGM) devices can perform plasma glucose monitoring tasks non-invasive by converting the collected interstitial fluid (ISF) plasma glucose into plasma glucose. Reconstructing plasma glucose from ISF has become an important time series prediction problem. Previous research has mainly focused on constructing a chamber model of ISF and plasma for plasma glucose reconstruction. Accurate parameter estimation is the guarantee of the performance of this model driven approach, and this precise parameter estimation is a difficult part of model driving. We propose an end-to-end approach: to complete the plasma glucose reconstruction task in a data-driven manner, obtaining the corresponding plasma glucose sequence at the corresponding time based on the ISF glucose sequence, and the model driven parameter estimation process will no longer be necessary. This end-to-end reconstruction method is one of the advantages of deep learning. This article uses popular time series analysis models in deep learning: LSTM, TCN, Transformer, Informer for plasma glucose reconstruction. Experiments on publicly available datasets have shown the superiority of this data-driven approach over baseline methods.