Meal detection based on non-individualized moving horizon estimation and classification

Meals are one of the greatest challenges to glucose regulation in diabetes mellitus type 1. Several times each day, food causes heavily elevated blood glucose concentrations that may result in long-term complications. Meal-time insulin boluses are administered to mitigate these hyperglycemic periods. Sporadic omissions of prandial boluses impair the outcome of the insulin therapy by leading to significant variations in blood glucose levels. As continuous glucose monitoring (CGM) becomes more common, an automated detection based on CGM data could support patients by reminding about missed boluses. In fully automated systems, meal detection could temporarily modify controller parameters until the meal is mitigated. In the present study, moving horizon estimation (MHE) and linear discriminant analysis (LDA), abbreviated “MHE+LDA”, are proposed for meal detection. An augmented version of Bergman's minimal model is used for the estimator model. Neither the model parameters nor the MHE tuning are individualized. The method is tested in simulations on the UVa/Padova simulator and its performance is compared to two other methods, namely threshold checking of the current estimated glucose appearance and the GRID algorithm. All meals are detected by MHE+LDA within 35 min while the two comparative methods do not detect the smallest simulated meal. The combination of MHE and LDA outperforms the two other methods also with respect to time of detection. The MHE+LDA method's ability to identify even smaller meals without the need for individual tuning suggests that the method should be further investigated