Long short-term memory neural network for glucose prediction

Diabetes is a chronic disease that affects a high percentage of the world population and produces different and serious complications to patients. Most diabetes complications may be avoided by controlling the blood glucose levels exhaustively. Moreover, a prediction of future glucose levels has shown to be fundamental in helping patients to plan and modify their treatment in real-time. In this paper, a glucose predictor based on long short-term memory neural networks is designed. Three input parameters are fed to the predictor: past glucose levels obtained from a continuous glucose monitoring sensor, the insulin units administered by an insulin pump and the patient's carbohydrates intake. Different prediction times and input dimensions have been evaluated in order to provide the best prediction to patients. Results encourage the use of glucose predictions to avoid the occurrence of hypoglycemias, anticipate correction actions, and to increase the quality of life of these patients.