Attention-based multilayer GRU decoder for on-site glucose prediction on smartphone

Abstract

Continuous glucose monitoring (CGM) devices provide a considerable amount of data that can be used to predict future values, enabling sustainable control of blood glucose levels to prevent hypo-/hyperglycemic events and associated complications. However, it is a challenging task in diabetes management as the data from CGM are sequential, time-varying, nonlinear, and non-stationary. Due to their ability to deal with these types of data, artificial intelligence (AI)-based methods have emerged as a useful tool. The traditional approach is to implement AI methods in baseline form, which results in exploiting less sequential information from the data, thus reducing the prediction accuracy. To address this issue, we propose a novel glucose prediction approach within the encoder–decoder framework, aimed at improving prediction accuracy despite the complex and non-stationary nature of CGM data. Sequential information is extracted using a convolutional neural network-based encoder, while predictions are generated by a gated recurrent unit (GRU)-based decoder. In our approach, the decoder is designed with the multilayer GRU attached to an attention layer to ensure the modulation of the most relevant information so that it leads to a more accurate prediction. The proposed attention-based multilayer GRU approach has been extensively evaluated on the OhioT1DM dataset, and experimental results demonstrate the advantage of our proposed approach over the state-of-the-art approaches. Furthermore, the proposed approach is also integrated with our custom-designed Android application called “GlucoWizard” to perform glucose prediction for diabetes.