Deep Learning Applied to Blood Glucose Prediction from Flash Glucose Monitoring and Fitbit Data

Blood glucose (BG) monitoring devices play an important role in diabetes management, offering real time BG measurements, which can be analyzed to discover new knowledge. In this paper we present a multi-patient and multivariate deep learning approach, based on Long-Short Term Memory (LSTM) artificial neural networks, for building a generalized model to forecast BG levels on a short-time prediction horizon. The proposed framework is evaluated on a clinical dataset of 17 patients, receiving care at the IRCCS Policlinico San Matteo hospital in Pavia, Italy. BG profiles collected by a flash glucose monitoring system were analyzed together with information collected by an activity tracker, including heart rate, sleep, and physical activity. Results suggest that a model with good prediction performance can be obtained and that a combination of HR and lifestyle monitoring signals can help to predict BG levels.