Nunzio Camerlingo, Italy
University of Padova Department of Information EngineeringPresenter of 1 Presentation
BEHAVIORAL MODEL OF POST-MEAL INSULIN CORRECTION BOLUS INJECTIONS IN TYPE 1 DIABETES INDIVIDUALS UNDER FREE-LIVING CONDITIONS
Abstract
Background and Aims
Besides describing the physiology of patients with type 1 diabetes (T1D), realistic simulation tools for in-silico trials should also mimic the behavioral aspects of patients’ lifestyle, which can notably affect glucose control. In a previous work (Camerlingo et al., J. Diabetes Sci. Technol., 2020) we modelled meal amount and timing variability, here we focus on the timing of post-meal insulin correction bolus (CB) injections.
Methods
A multicenter study involving 30 patients with T1D, monitored in free-living conditions for 1-month (Kovatchev et al., Diabetes Technol. Ther., 2017) was used to extract 539 CBs for 1,963 meals. 7-hour post-prandial windows were divided in 30-min portions, labelled as “1” or “0”, based on the occurrence of a CB injection. Three different binary classification techniques were implemented to predict the labels: support vector machine (SVM), decision tree (DT) and logistic regression (LOG), based on 13 features extracted from continuous glucose monitoring (CGM), insulin, and meal data as well as from patient’s characteristics. Average area under the receiver operating characteristic curve (AUROC) over 10-fold cross validation was used to select the best model.
Results
SVM provided an AUROC of 0.76±0.04 (mean±std), performing slightly better than LOG (0.75±0.04) and DT (0.73±0.04). The 8 most representative predictors were: time from last bolus and from last meal, current CGM reading and rate-of-change, daytime, patient’s age, body weight, and correction factor.
Conclusions
Once refined using larger datasets, the new model can be incorporated in T1D simulators. By mimicking patient behavior in self-administering CBs, the model will allow more realistic in-silico trials.