Predicting continuous blood glucose level using deep learning

Abstract

Diabetes is among the most common chronic diseases nowadays; in diabetes management control of blood glucose is essential. Significant attention has been paid to get the accurate prediction of diabetes. Various deep learning techniques are already proposed, such as multiple types of Neural Networks, SVR, LVX, ARX, LSTM models, and many more. The error rate of existing predicting models are very high. Error rate in prediction can cause several false positive notifications, which results in a decreasing the accuracy if the model. This study presented a hybrid model for predicting blood glucose levels based on two different kinds of neural networks CNN and GRU. The proposed model can predict blood glucose levels with leading accuracy of (MSE = 26.88 ± 17.87 [mg/dl] for 15 mins, MSE = 39.82 ± 22.19 [mg/dl] for 30 mins, MSE = , 66.33 ± 25.2 [mg/dl] for 60 mins) and (RMSE = 4.84 ± 1.83 [mg/dl] for 15 mins, RMSE = 6.04 ± 1.84 [mg/dl] for 30 mins, RMSE = 8.12 ± 1.46 [mg/dl] for 60 mins) on simulated T1D patient. The proposed model used CGM data and used extra features as input, like carbohydrates and insulin. The proposed model is then evaluated on 10 simulated patients of different ages generated using the UVA/Padova simulator.