Estimation of the insulin sensitivity profile for the stochastic variant of the ICING model

In this paper the insulin sensitivity profile and the diffusion term as stepwise functions were determined for the grey box variant of the ICING (Intensive Control Insulin-Nutrition-Glucose) model used for virtual trial methodology. The suggested technique can separate system noise, which was earlier lumped into the insulin sensitivity profile itself. In this way one can get smaller residual for the model fitting as well as more precise value for the insulin sensitivity. This fact can improve the prediction process estimating insulin sensitivity value for the future time span and therefore it can ensure better clinical treatment. Maximum likelihood method has been employed to compute the values of the insulin sensitivity profile and the diffusion term for every half hour intervals. The global optimization problem was solved using genetic algorithm, simulated annealing method and Nelder-Mead procedure employing parallel computation. The application of the grey box model was able to reduce the errors sometimes with even more than 50%. According to the likelihood ratio test, the stochastic variant of the ICING model yielded significant improvement comparing to the white box one.