Clinical trials in free-living conditions is key in the development of an Artificial Pancreas (AP) for T1D subjects. Since the scenario plays a key role in the synthesis and validation of AP control algorithms, a probabilistic approach is proposed to automatically design meal scenarios. In particular, we exploit our real-life data to design realistic in silico scenarios.
The amount and time-of-day of ingested carbohydrates in a 1-month in 13 patients for a total of 1500 meals. have been considered. The joint distribution of these variables has been estimated via a copula function, in order to model their dependence. The use of a copula allows to generate Monte Carlo scenarios by drawing random samples, which represent a pair of amount and time-of-day.
A Gaussian copula resulted suitable for the description of the dependence in the meal dataset with a p-value of 0.005 according to the χ2 test based on Rosenblatt’s transformation. A bootstrap version of the test shows that the estimate of the Spearman correlation coefficient (ρ) is sufficiently accurate with respect to the correlation (ρ) directly computed from the data (ρ=0.13, ρ=0.12).
The availability of a copula statistical model able to represent the food habits of a T1D population allows to design realistic eating patterns to run in silico simulations under free-living conditions.