Realizing Blood Glucose Prediction by Convolutional Recurrent Neural Networks with Residual Blocks

An advanced convolutional recurrent neural network architecture for forecasting blood glucose is proposed in this paper. To improve the competitiveness of the suggested model, several merits of WaveNet, a deep learning model that performs well on processing audio waveforms, are adapted and implemented. To be more specific, a multi-layer dilated causal convolutional neural network (CNN) with residual blocks followed by a modified recurrent neural network (RNN) with GRU cells is the architecture of our model. 10 virtual adult patients from the UVA/Padova T1D simulator provide the 10 simulated datasets for in-silico experiment, and each dataset consists of 6 channels of time series data, including glucose levels, insulin dosages, carbohydrate intake, the rate of glucose appearance, plasma insulin and the time index. After preprocessing, the data is fed into the network with the aim to forecast the blood glucose level in the next 30 minutes. The obtained prediction results are evaluated by the root mean squared error (RMSE). The average of the best RMSE among the 10 subjects is 8.3050. This RMSE result is better than that of many current algorithms using the same datasets, which shows the superior performance of our proposed model.