A NARX modeling-based approach for evaluation of insulin sensitivity

Abstract

Evaluation of insulin sensitivity is of prime importance in the clinical investigation of glucose related diseases. This paper deals with a novel model-based technique for the evaluation of an index for insulin sensitivity. A set of nonlinear autoregressive models are identified from the clinical test data of normal subjects. The two-stage identification procedure involves proper structure selection for approximating the input–output data followed by estimating the parameters of the polynomial model. The models obtained are analyzed to derive an index for insulin sensitivity by determining the effect of insulin on glucose utilization. A median bootstraped correlation (sampling with replacement) of 0.97 with 90% confidence interval of [0.92 0.98], is obtained between the indexes of the proposed model and the widely used minimal model. The proposed model is able to achieve a good fitting performance on the validation dataset. The results also suggest that for representing the dynamics of insulin action on glucose disposal, the proposed model overcomes some of the well known limitations of the minimal model, and thus gives a better representation of insulin sensitivity.