Effect of input excitation on the quality of empirical dynamic models for type 1 diabetes

Accurate prediction of future blood glucose trends has the potential to significantly improve glycemic regulation in type 1 diabetes patients. A model‐based controller for an artificial β‐cell, for example, would determine the most efficacious insulin dose for the current sampling interval given available input–output data and model predictions of the resultant glucose trajectory. The two inputs most influential to the glucose concentration are bolused insulin and meal carbohydrates, which in practice are often taken simultaneously and in a specified ratio. This linear dependence has adverse effects on the quality of linear dynamic models identified from such data. On the other hand, inputs with greater degrees of excitation may force the subject into extreme hypoglycemia or hyperglycemia, and thus may be clinically unacceptable. Inputs with good excitation that do not endanger the subject are shown to result in models that can predict glucose trends reasonably accurately, 1–2 h ahead.