Blood glucose forecasting using LSTM variants under the context of open source artificial pancreas system

Abstract

High accuracy of blood glucose prediction over the long term is essential for preventative diabetes management. The emerging closed-loop insulin delivery system such as the artificial pancreas system (APS) provides opportunities for improved glycaemic control for patients with type 1 diabetes. Existing blood glucose studies are proven effective only within 30 minutes but the accuracy deteriorates drastically when the prediction horizon increases to 45 minutes and 60 minutes. Deep learning, especially for long short term memory (LSTM) and its variants have recently been applied in various areas to achieve state-of-the-art results in tasks with complex time series data. In this study, we present deep LSTM based models that are capable of forecasting long term blood glucose levels with improved prediction and clinical accuracy. We evaluate our approach using 20 cases(878,000 glucose values) from Open Source Artificial Pancreas System (OpenAPS). On 30-minutes and 45-minutes prediction, our Stacked-LSTM achieved the best performance with Root-Mean-Square-Error (RMSE) marks 11.96 & 15.81 and Clark-Grid-ZoneA marks 0.887 & 0.784. In terms of 60-minutes prediction, our ConvLSTM has the best performance with RMSE = 19.6 and Clark-Grid-ZoneA=0.714. Our models outperform existing methods in both prediction and clinical accuracy. This research can hopefully support patients with type 1 diabetes to better manage their behavior in a more preventative way and can be used in future real APS context.