Transfer and Incremental Learning Method for Blood Glucose Prediction of New Subjects with Type 1 Diabetes

Blood glucose prediction is now valuable for diabetes patients to control the blood glucose level and then manage it. However, it may have the problem of insufficient modelling data and modelling burden if a completely new model has to be developed for every new subject. This work proposes a data-driven transfer learning method to quickly get the new prediction model for new subjects based on other subjects' models and a small number of new data. Besides, an incremental learning method is developed to consecutively improve the new model with more new data available. By simultaneous analysis of transfer learning and incremental learning, the prediction model for new subjects can be readily obtained with few calculation efforts and meanwhile the accuracy is guaranteed. Finally, the proposed method is compared with the general model and individual model with the data from the UVA/PADOVA Type I Diabetes Simulator. The proposed method has shown superiority for both calculation efforts and accuracy.