Estimation of blood glucose levels techniques

In order to avoid complications in diabetes type 1, a firm management of Blood Glucose Level (BGL) is required, keeping it within a safe range. In this dissertation, several methods such as Extended Kalman Filter (EKF), Nonlinear AutoRegressive with exogenous inputs (NARX) network and Extreme Learning Machine (ELM) for the prediction of future glucose concentrations are presented and their prediction performances are also proposed. The proposed methods are evaluated using one database namely the automated insulin dosage advisor (AIDA) database containing 20 virtual patients, each method was evaluated following Root-Mean-Square Error (RMSE), and Sum of Squares of the Glucose Prediction Error (SSGPE), for 15-min ahead prediction. However, the results highlight the good performance of the NARX network compared to other methods. In other words, the NARX network gives satisfactory predictions and should be used in BGL prediction.