Neural identification of Type 1 Diabetes Mellitus for care and forecasting of risk events

Abstract

Glucose–insulin models, testing glucose sensors and support systems for health care decisions play an important role in synthesis of glucose control algorithms. In this work we propose an online glucose–insulin identification using the Recurrent High Order Neural Network (RHONN). Then, the model obtained is used to predict n-steps forward of glucose levels, also by RHONN. The used data for identification is from a Type 1 Diabetes Mellitus (T1DM) patient, it was collected from the Continuous Monitoring Glucose System (CMGS) by MiniMed Inc ® and an insulin pump by Paradigm Real-time Insulin Pump ®. RHONN is trained online by Extended Kalman Filter (EKF). The results suggest that it is possible to make a prediction of up to 35 min in the future, which it would help to prevent risky events (hypoglycemia and hyperglycemia). Also shows that, it could be directly connected to a CGMS to help the patient improve the glucose control and even an automatic glucose control algorithm. The proposed Neural Network shows good performance compared to baseline methods in terms of evaluation criteria.