Blood Glucose Level Prediction For Type 1 Diabetes Using Deep Learning

Diabetes is a common disease all around the world. It is a difficult and incurable but controllable disease, so it is important to control and prevent its complications. Thus, low error and smart methods are used to predict blood glucose levels and prevent dangerous complications to control it effectively. In this regard, different methods were investigated. In our research, two models based on deep learning technique are used which produce efficient and optimal results. These models are composed of different combinations of long short-term memory and feed forward neural networks which predict the patient's future blood glucose levels with considerable accuracy and speed. To achieve more comprehensive model, 81,200 blood glucose data was evaluated through 203 patients. In addition, 27 effective features in patients' blood glucose levels are considered in this regard. Furthermore, cross-validation method which is suitable for time series was used for more accurate evaluation. The results showed that Autoregressive Integrated Moving Average model could not predict blood glucose levels considering this amount of data and system hardware limitations, while the models based on deep learning had good performance and good speed. Furthermore, the second proposed model for the prediction horizons of 5, 10, and 15 minutes outperformed the first one with 13.8%, 16%, and 18.9%, respectively. Therefore, the second proposed model can be more reliable for predicting blood glucose. So, it can be used in smart warning systems to prevent hypoglycemia, which is a dangerous and widespread problem of type 1 diabetes.