Evaluating Clinical Accuracy of Models for Predicting Glycemic Behavior for Diabetes Care

In this work we present a deep study about the real prediction capabilities of two different models: Wiener type and autoregressive with exogenous inputs (ARX) in the field of Diabetes Mellitus Type 1 (DMT1). With this purpose we recorded real data from 16 patients divided in three different experiments at the Center for Diabetes Technology. We analyzed several aspects concerning this topic such as the excitation persistency of the input signals, the inclusion of filters for determining plasma insulin and glucose rate of appearance and the factors to emulate the circadian rhythm effect on the insulin sensitivity to obtain good predictions for 1 and 2 hs. ahead of blood glucose concentration. We have data derived from two studies with 15 patients of almost 30 days and the study of one patient with five months of data. Therefore, we can conclude about the prediction performance of the models depending of the available amount of information taking a part for estimation purpose and other for validation. We can analyze what happens with the validity of the model along the time and the incidence of the selection of data sets under different moments of the patient life. To compare with other works, we present here a summary of the achieved results on this topic by other authors using real data to obtain other kind of models. To present rigorous comparisons we were based on two index very well known in the field of health. They are the Root Mean Square Error (RMSE) and the Clarke Grid Error Analysis (p-CGA). Based on these results we can present the conclusions about the best representative models to be used in the field of diabetes care and on the development of novel technology and tools to assist on improving its treatment.