A Hybrid Deep Learning Model for the Blood Glucose Prediction

Blood glucose (BG) control is important for diabetics to avoid hyperglycemia and hypoglycemia events in their daily life. And the forecasting of BG becomes one of the most important sections of the blood glucose control system. Accurate prediction makes patients can react to the abnormal values in advance so that precisely controlling their BG level can be realized. In this study, we propose a novel method to predict BG levels for 30 min of prediction horizon (PH) with a hybrid deep learning model, which integrates multi-layer perceptron, stacked bidirectional gated recurrent unit (Bi-GRU) based recurrent neural network (RNN), and the attention mechanism (AM). Firstly, after pre-processing, the real patient data (in this work, the history BG only) as the input is transmitted to the network used for training and testing. Subsequently, the stacked Bi-GRU layers are used for extracting temporal characteristics of time sequence data, the AM is used for paying more attention to the temporal information which has higher relevance with outputs by assigning weights and then multi-layer perceptron for rapid gradient descent and calculating the regression output of our neural network (in this work, the predict BG). Finally, some important criteria of predictive evaluation are used to demonstrate the effect of our method.