Stochastic Seasonal Models for Glucose Prediction in Type 1 Diabetes under Free-Living Conditions

Stochastic Seasonal Models for Glucose Prediction in Type 1 Diabetes under Free-Living Conditions


Accurate predictions of blood glucose (BG) concentration with longer prediction horizon (PH) might improve type-1 diabetes (T1D) therapy by allowing patients to adjust the therapy based on BG future values. Therefore, if an online prediction of BG concentration more accurate than the existing approaches in the literature is possible then all the systems based on BG predictions can be improved. For example, complete supervision systems based on these values can be devised, or better MPC (Model Predictive Control) algorithms for artificial pancreas (AP) systems (automatic insulin infussion systems) can be developed.

The concept of seasonality has demonstrated successfully its accuracy in the BG prediction models for long PHs [1], but it requires similarity between the series (glucose responsed) used to develop seasonal models. Automatic clustering of time series obtained from T1D patients has also been demonstrated previuosly [2]. However, the proposals done in [1] and [2] are not directly applicable for the use in the normal life.

The proposal is a prediction method developed, first of all, by enforcing the concept of seasonality in normal life data and then integrating the predictions of a set of seasonal local models (each of them corresponding to different glucose profiles observed along historical data). In the modeling step, the number of sets and their corresponding glucose profiles characteristics are obtained by clustering techniques (Fuzzy C-Means). Then, Box-Jenkins methodology is used to identify a seasonal model for each set. Finally, the online BG prediction is obtained by local model integration through realtime membership-to-cluster estimation.

Additionally to the BG prediction, an online monitoring system informs about prediction confidence and abnormal behavior detection through a trust and normality index, respectively.

As a proof of concept, the framework has been tested over 6 months data of UVA/Padova simulator extended with several variability sources. The framework exhibits high prediction accuracy for large PHs: a MAPE of 4.10%, 5.95%, 8.43%, 11.32%, 13.65%, and 13.97% has been achieved for 15- , 30-, 60-, 120-, 180-, and 240-min PHs, respectively.

The results allow us to conclude that the proposal exhibits high prediction accuracy for large PHs (a 240-min-ahead prediction) and, therefore, it could be helpful to allow the diabetic patients anticipate therapeutic decisions. As well, the proof of concept has demonstrated the feasibility for using the trust index for measuring the confidence in the estimation and the normality index for detecting the abnormal behavior/states (time series not used in the model identification) before the action occurs.

Estado de Protección: Nacional: US17/143,572 – 07/01/2021