Glucose Prediction Based on the Recurrent Neural Network Model

Abstract

An advanced convolutional neural network architecture for forecasting blood glucose is proposed in this paper. Four different measures are introduced in this essay, including Glucose, Meal, Insulin, and Time of the day, which are denoted as G, M, I, and T for short. Past 2-hour historical data of individuals are exploited to predict the future glucose level in 30 minutes with high accuracy. To verify the effectiveness of the blood glucose prediction model, three major methods have been displayed and compared. To be more specific, Recurrent Neural Network (RNN) was the better model for forecasting blood glucose, compared with Neural Network Predictive Glucose (NNPG) and Support Vector Regression (SVM). The metrics of evaluation are Root-Mean-Square deviation (RMSE) and Mean Absolute Relative Difference (MARD). The average of the best RMSE is 7.75, which is largely better than those of the other two models. This result shows the superior performance of RNN in accurate glucose prediction.