Development of physiological metabolic models in diabetes based on data mining techniques (Doctoral thesis)
Γεώργα, Ελένη Ι.
In this thesis, we address the problem of the short-term prediction of glucose concentration in the interstitial fluid in people with type 1 diabetes under free-living conditions. This thesis consists of three main parts. In the first part, we approached the specified problem via a time-invariant support vector regression function of multiple input variables, concerning the recent subcutaneous glucose profile, the effect of food and insulin intake, the energy expenditure due to physical activities and the time of the day, which was evaluated individually for each patient. By utilizing different input cases, the effect of each input to the model’s prediction error was quantified and, it was demonstrated that the effective combination of multivariable data can significantly improve the prediction error. The subsequent study on the evaluation of the proposed model with respect to the prediction of single hypoglycaemic events, drove us to introduce new input variables accounting for recurrent nocturnal hypoglycaemia, due to antecedent hypoglycaemia, exercise, and sleep, which resulted in a considerably higher sensitivity and precision values. In the second part of this thesis, we proceeded to feature ranking for assessing, separately for each patient, the importance of the defined feature set with respect to subcutaneous glucose concentration prediction, aiming at the customization of the input of the regression function. To this end, the random forests and RReliefF algorithms were employed, and through a forward sequential feature selection procedure, we investigated the effectiveness of highly-ranked features on the prediction error by kernel-based regression models (support vector regression and Gaussian processes). In the third part of this thesis, we demonstrated the capability of sparse kernel adaptive filtering algorithms (i.e. fixed budget quantized kernel least mean square algorithm, and approximate linear dependency kernel recursive least squares algorithm) to learn online and predict the short-term course of the subcutaneous glucose concentration in type 1 diabetes. In parallel, we verified that multivariate data improve systematically both the regularity and the time lag of the predictions, reducing the errors in critical glucose value regions.
|Institution and School/Department of submitter:||Πανεπιστήμιο Ιωαννίνων. Πολυτεχνική Σχολή. Τμήμα Μηχανικών Επιστήμης Υλικών|
|Keywords:||Type 1 diabetes,Continuous glucose monitoring,Glucose prediction,Machine learning,Kernel based models,Adaptive learning,Physiological system modelling,Διαβήτης τύπου 1,Συνεχής παρακολούθηση γλυκόζης,Πρόβλεψη γλυκόζης,Μηχανική μάθηση,Μέθοδοι βασιζόμενες σε συναρτήσεις πυρήνα,Προσαρμοστική μάθηση,Μοντελοποίηση συστημάτων φυσιολογίας|
|Appears in Collections:||Διδακτορικές Διατριβές|
Please use this identifier to cite or link to this item:This item is a favorite for 0 people.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.