Chaotic time series analysis approach for prediction blood glucose concentration based on echo state networks

Abstract

Blood glucose prediction plays a very critical role in the treatment of diabetes. With the development of continuous glucose monitoring system (CGMS), it becomes possible to know the blood glucose level at real time. In this literature, we establish a predictive model using echo state neural networks (ESN) due to its excellent performance in chaotic time series forecasting. In order to further improve the performance of the network, we optimized the ESN with leakage integral neurons and ridge regression learning algorithm. Under the same condition, the proposed method is compared with the Extreme Learning Machine and Back Propagation algorithm in terms of Root mean square error (RMSE), Time gain (TG) and the Continuous glucose-error grid analysis (CG-EGA). The experimental results demonstrate that ESN is a very suitable prediction model for blood glucose time series.