Subspace Identification of a Glucose-Insulin model Using Meal Tracer Protocol Measurements

In this study, the problem of identifying a low complexity state space model describing glucose and insulin dynamics from low sample meal tracer experiments is investigated. Triple tracer meal protocol measurements (sampled as low as 15 samples per meal) together with continuous glucose monitoring measurements, measured concurrently at a rate of 5 minutes per sample, are used. A new formulation to estimate the missing input and output measurements at such low sample rates is developed. Nuclear norm minimization is used to exploit low rankness of the stacked input and output matrix, while the {ell1} norm is used to exploit an available sparse basis for the glucose flux profiles. Simulation results, using the UVa Padova simulator, show that the technique outperforms previous methods and also demonstrate the possibility of identifying state space models from triple tracer measurements with good prediction performance under non-ideal conditions.