Neural Network for Prediction of Glucose Concentration in Type 1 Diabetic Patients

Prediction of biological time series is important in chronic pathologies to allow prevention of health threatening events. In type 1 diabetes several real-time short-term prediction methods, based on time-series modeling of past continuous glucose monitoring (CGM) sensor data, have been proposed with the aim of generating preventive alerts to mitigate hypo/hyperglycemia. Even if scarcely explored, neural network (NN) approaches are promising given their ability of learning nonlinear functions easily integrating, among their inputs, signals of different nature. In this contribution we describe a jump NN predictor (horizon 30 min) that uses past CGM data and information on ingested carbohydrates. The algorithm is optimized on data of 10 type 1 diabetics and assessed on 10 different subjects. Prediction is accurate and give an anticipation sufficient to potentially avoid risky events.