Deep prediction model: The case of online adaptive prediction of subcutaneous glucose

Abstract

In this paper, we propose the concept of the deep prediction model for subcutaneous glucose concentration. The concept is based on several layers of prediction models. One aim of this approach is to eliminate time lag, which is more severe in longer prediction horizons. Thus, the prediction accuracy of the algorithm might be increased, even for longer prediction horizons. The second goal is to create new, potentially good predictors that could be obtained by combining existing predictors. The effectiveness of the proposed model is illustrated in several examples of two-layer networks. In the first layer, a specific linear/non-linear prediction model is used. In the second (correction) layer, an extreme learning machine is used, due to its rapid learning capabilities. In almost all experiments, the proposed method has reduced the time lag and improved the accuracy of the method.