In this paper we consider the problem of predicting future values of subcutaneous glucose (glucose concentration in the interstitial fluid) in Type-1 diabetes patients, exploiting information on injected insulin, carbohydrates intake and past subcutaneous glucose samples measured by a Continuous Glucose Monitor sensor. Prediction can be used to warn the patient of upcoming possibly harmful events, such as hypoglycemia or hyperglycemia. In addition, an effective prediction is a key ingredient in Model Predictive Control, a technique successfully applied in the development of the so called artificial pancreas. Derivation of individualized predictors is crucial to cope with the wide inter-subject variability. For this reason we explored the application of linear black-box identification methods to derive patient-tailored predictors. Together with the mainstream technique in system identification, the Prediction Error Method, we investigate a novel and promising nonparametric method based on Gaussian regression. The considered methods are applied to data collected on 20 subjects, during two 21-hours hospital admissions for each subject. One of the major challenges of this problem is that the available dataset is short and the glucose-insulin system has very high complexity. The quality of the prediction was compared on the basis of three metrics commonly used in system identification, namely Coefficient of Determination, Fit and Root Mean Squared Error. On a system identification perspective the present work is a comparison of a novel system identification approach with the state-of-the-art one on a particularly difficult and relevant case of study. The nonparametric technique improves prediction performance and reduces computational burden associated with predictor identification.
File in questo prodotto:
Non ci sono file associati a questo prodotto.