Glycemia Prediction Accuracy of Simple Linear Models with Online Parameter Identification

This paper deals with the glycemia prediction accuracy comparison of 7 simple linear model structures, with prediction based on Continuous Glucose Monitoring (CGM) data. Different model structures presented are ARX, ARMAX and Box Jenkins models and their single and multiple input variations, which can represent known or not known meal intake. For each model structure the recursive parameter identification algorithm is presented. All prediction model structures are tested on 7 selected CGM datasets of several type 1 diabetes subjects, acquired in DiaDAQ project. Prediction performance is evaluated using a model fit metrics and an error grid analysis. The results show good prediction performance with low variability among different model structures, despite the simplicity of the prediction algorithm.