Evaluating a Comparing Deep Learning Architectures for Blood Glucose Prediction

To manage their disease, diabetic patients need to control the blood glucose level (BGL) by monitoring it and predicting its future values. This allows to avoid high or low BGL by taking recommended actions in advance. In this paper, we conduct a comparative study of two emerging deep learning techniques: Long-Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN) for one-step and multi-steps-ahead forecasting of the BGL based on Continuous Glucose Monitoring (CGM) data. The objectives are twofold: 1) Determining the best strategies of multi-steps-ahead forecasting (MSF) to fit the CNN and LSTM models respectively, and 2) Comparing the performances of the CNN and LSTM models for one-step and multi-steps prediction. Toward these objectives, we firstly conducted series of experiments of a CNN model through parameters selection to determine its best configuration. The LSTM model we used in the present study was developed and evaluated in an earlier work. Thereafter, five MSF strategies were developed and evaluated for the CNN and LSTM models using the Root-Mean-Square Error (RMSE) with an horizon of 30 min. To statistically assess the differences between the performances of CNN and LSTM models, we used the Wilcoxon statistical test. The results showed that: 1) no MSF strategy outperformed the others for both CNN and LSTM models, and 2) the proposed CNN model significantly outperformed the LSTM model for both one-step and multi-steps prediction. © 2021, Springer Nature Switzerland AG.