Meal detection and carbohydrate estimation based on a feedback scheme with application to the artificial pancreas

GODOY, J.L.; SERENO, J.E.; RIVADENEIRA, P.S.

BIOMEDICAL SIGNAL PROCESSING AND CONTROL

ELSEVIER SCI LTD

Año: 2021 vol. 68

1746-8094

Current glucose control systems automatically regulate basal insulin infusion, but users still need to manually announce meals (major disturbances) to dose prandial insulin boluses. This issue needs to be solved to reach a fully automated artificial pancreas. Automatic meal detection and carbohydrate amount estimation from readings of blood glucose (BG) and insulin infusion can improve the artificial pancreas control system from two possible paths: (i) the off-line reconstruction of the carbohydrate intake signal which allows a reliable identification of a control-relevant model, and (ii) the on-line prediction of meal onset and amount of carbohydrates ingested, which allows safety supervision of manually entered meal announcements. The aim of this work is the item (i), for which an automatic algorithm is developed to detect the consumption of a meal and estimate its carbohydrate amount in people with type 1 diabetes. The unknown input estimation is based on a feedback scheme where the measured BG is compared with a BG prediction. Glycemic behavior is predicted using a personalized model by means of the patient´s functional insulin therapy parameters defined by the treating physician. The proposed algorithm is evaluated with data extracted from the 30-patient cohort of the UVA/Padova simulator approved by the FDA and with retrospective data from 11 real patients of a diabetes center. Diabetes care data from free-living adult patients were collected during regular screening and the meals were identified by experts. For the in silico dataset, the detection accuracy is near 100%, the absolute error of the estimation of ingested carbohydrates is 10% on average, and the average bias of meal onset estimation is 5 min. For the clinical dataset, the meal detection performance is 98% and the estimation accuracy measures are 13% and 2 min, respectively. In this work, the impact of reconstructing the carbohydrate intake signal on the identification proved to be beneficial. In addition, the feedback scheme and the easily personalized prediction model make the strategy efficient.