Regularization networks for glucose system identification

Abstract

A framework for non-linear identification of glucose kinetics using neural networks is presented. The framework combines: recursive input-output system representation (Non-linear AutoRegressive model with eXogenous inputs (NARX)); approximation method derived from regularization theory and based on radial basis function neural networks; and validation methods for non-linear systems. System identification was performed using: (1) simulated data from a mathematical model of glucose kinetics in a diabetic state with exogenously infused soluble insulin and monomeric insulin analogues and (2) measured subcutaneous tissue glucose time-series from healthy subjects, respectively.