An autonomous channel deep learning framework for blood glucose prediction

Abstract

Accurate prediction of blood glucose (BG) is conducive to avoiding abnormal blood glucose events and improving blood glucose management for Type 1 diabetes (T1D) patients. Recently, many deep learning-based methods for BG prediction have been proposed with encouraging results. However, most deep prediction methods do not consider the time-dependent scale discrepancy of different variables on BG dynamics and use the same time window for all input variables. This neglect will directly lead to information redundancy on short-term related variables or information incompletion on long-term related variables, which is not conducive to prediction accuracy. In this regard, we proposed an autonomous channel deep learning framework for personalized multivariate BG prediction. The autonomous channel network in the proposed framework learns representation from input variables with reasonable sampling periods and sequence lengths based on the domain knowledge of time-dependent scale between variables, thereby effectively avoiding input information redundancy and incompletion. The framework was evaluated on a clinical dataset, OhioT1DM Dataset, with experimental results in terms of root mean square error (RMSE) (18.930 ± 2.155 mg/dL) with the mean absolute relative difference (MARD) (9.218 ± 1.607%) for prediction horizons (PH) = 30 min. These are the best-reported results for BG prediction when compared with other methods including the support vector regression (SVR), the long short-term memory network (LSTM), the dilated recurrent neural network (DRNN), the temporal convolutional networks (TCN), and the deep residual time-series forecasting (DRTF).