Convolutional Recurrent Neural Networks for Glucose Prediction

Glucose control serves as the primary method of diabetes management. Current digital therapeutic approaches for subjects with Type 1 diabetes mellitus (T1DM) such as the artificial pancreas and bolus calculators leverage machine learning techniques for predicting subcutaneous glucose for improved control. Deep learning has recently been applied in healthcare and medical research to achieve state-of-the-art results in a range of tasks including disease diagnosis, and patient state prediction among others. In this work, we present a deep learning model that is capable of predicting glucose levels over a 30-minute horizon with leading accuracy for simulated patient cases (RMSE = 10.02$\pm$1.28 [mg/dl] and MARD = 5.95$\pm$0.64\%) and real patient cases (RMSE = 21.23$\pm$1.15 [mg/dl] and MARD = 10.53$\pm$1.28\%). In addition, the model also provides competitive performance in forecasting adverse glycaemic events with minimal time lag both in a simulated patient dataset (MCC$_{hyperglycaemia}$ = 0.82$\pm$0.06 and MCC$_{hypoglycaemia}$ = 0.76$\pm$0.13) and in a real patient dataset (MCC$_{hyperglycaemia}$ = 0.79$\pm$0.04 and MCC$_{hypoglycaemia}$ = 0.28$\pm$0.11). This approach is evaluated on a dataset of 10 simulated cases generated from the UVa/Padova simulator and a clinical dataset of 5 real cases each containing glucose readings, insulin bolus, and meal (carbohydrate) data. Performance of the recurrent convolutional neural network is benchmarked against four state-of-the-art algorithms: support vector regression (SVR), latent variable (LVX) model, autoregressive model (ARX), and neural network for predicting glucose algorithm (NNPG).