A Multi-Patient Data-Driven Approach to Blood Glucose Prediction

Continuous glucose monitoring systems (CGMSs) allow measuring the blood glycaemic value of a diabetic patient at a high sampling rate, producing a considerable amount of data. These data can be effectively used by machine learning techniques to infer future values of the glycaemic concentration, allowing the early prevention of dangerous hyperglycaemic or hypoglycaemic states and better optimization of the diabetic treatment. Most of the approaches in the literature learn a prediction model from the past samples of the same patient, which needs extensive calibrations and limits the usability of the system. In this paper, we investigate the prediction models trained on glucose signals of a large and heterogeneous cohort of patients and then applied to infer future glucose-level values on a completely new patient. To achieve this purpose, we designed and compared two different types of solutions that were proved successful in many time-series prediction problems based respectively, on non-linear autoregressive (NAR) neural network and on long short-term memory (LSTM) networks. These solutions were experimentally compared with three literature approaches, respectively, based on feed-forward neural networks (FNNs), autoregressive (AR) models, and recurrent neural networks (RNN). While the NAR obtained good prediction accuracy only for short-term predictions (i.e., with prediction horizon within 30 min), the LSTM obtained extremely good performance both for short- and long-term glucose-level inference (60 min and more), overcoming all the other methods in terms of correlation between the measured and the predicted glucose signal and in terms of clinical outcome