A Novel RNN-Based Blood Glucose Prediction Approach Using Population and Individual Characteristics

As a common endocrine disease, diabetes has been plagued the lives of patients. An accurate blood glucose (BG) prediction approach can not only be used in daily BG management to reduce the occurrence of hypoglycemia or hyperglycemia, but also regulate the amount and time of insulin injection combined with insulin pump. Data driven methods have become an effective way for predicting BG. While time series analysis methods only deal with one patient at a time and most machine learning approaches simply use multiple patients' data to capture the population characteristics of BG but ignore the individual characteristics. To overcome these shortcomings, we propose a novel neural network approach based on GRU in which both population and individual characteristics of BG fluctuation are well integrated by pre-training and fine-tune processes. The proposed approach not only overcomes the problem of insufficient data for individual patient, but also makes full use of the population and individual differences of BG fluctuation. Compared with other machine learning or neural network approaches, the numerical results suggest that the proposed approach gains significant improvements on prediction accuracy.