Kalliopi Dalakleidi is defending her doctoral thesis titled "Development of Clinical Decision Support Systems for the Management of Diabetes Mellitus" on July 26, 2017
The event will be held in the premises of the Central Library of NTUA (multimedia center), at 10:00.
In the present PhD thesis, the study and development of decision support systems for the prevention, diagnosis, and treatment of diabetes has been conducted.
In the first part of the thesis, a comparative assessment of different machine learning and statistical methodologies towards the development of risk prediction models for the incidence and the evolution of Type 2 Diabetes Mellitus has been orchestrated. The use of ensembles of classifiers, and specifically ensembles of feed forward neural networks, for the prediction of Diabetes Mellitus for Pima Indian women and Cardiovascular Diseases for patients with Type 2 Diabetes Mellitus has been examined. Several classifiers have been developed, others follow the Bagging paradigm, others are ensembles of Feed-forward Neural Networks with different numbers of hidden neurons or layers, others follow the Binary Logistic Regression paradigm, others follow the Bayesian approach, and others are variations of Decision Trees. It has been shown that the ensembles of classifiers that have been trained following the Bagging approach and the ensembles of feed forward neural networks with different numbers of hidden neurons or layers have achieved the highest levels of prediction accuracy and their predictions are closer to the real risk scores, as indicated by the results of the Hosmer-Lemeshow test. The obtained results justify that ensembles of Artificial Neural Networks can significantly contribute in predicting the incidence of T2DM or its complications by having the capacity to handle the unbalanced nature, which usually occurs in medical datasets, and furthermore to capture an individual’s health evolution. In the second part of this thesis, feature selection is conducted in order to find the most critical clinical features which are strongly related with the incidence of fatal and non fatal Cardiovascular Disease in patients with Type 2 Diabetes Mellitus. The proposed system is based on the use of a Genetic Algorithm with a fitness function that depends on the classification sensitivity and accuracy of a Dual Weighted K-Nearest Neighbours classifier. The best subsets of features proposed by the implemented algorithm include the most common risk factors, such as age at diagnosis, duration of diagnosed diabetes, glycosylated haemoglobin (HbA1c), cholesterol concentration, and smoking habit, but also factors related to the presence of other diabetes complications and the use of antihypertensive and diabetes treatment drugs (i.e. proteinuria, calcium antagonists, b-blockers, diguanides and insulin).
In the third part of this thesis, a food recognition system is proposed, which consists of two modules performing feature extraction and classification of food images, for the automatic assessment of carbohydrates (CHO) in the meals of diabetic patients. In an automatic food recognition system, the user first takes a photograph of the upcoming meal with the camera of his mobile phone. Then, the image is processed so that the different types of food are divided from each other and segmented in different areas of the image. A series of features are extracted from each segmented area and are fed to a classifier, which decides what kind of food is represented by each segmented area. Then, the volume of each segmented area is calculated and the total CHO of the depicted meal are estimated. The combination of Speeded Up Robust Features (SURF), Color and Local Binary Pattern (LBP) features is examined in this thesis, since SURF ensures that spatial intensity patterns are captured, and Color and LBP features ensure stability and distinctiveness. Moreover, a novel modified version of the All-And-One (M-A&O) SVM classifier for multiclass classification problems is proposed and its performance is assessed against classification methods based on SVM or the K-Nearest Neighbour approaches including the One-Against-All (OAA) SVM, the One-Against-One (OAO) SVM, the All-And-One (A&O) SVM, the Weighted K-Nearest Neighbour (WKNN) classifier, the Dual Weighted K-Nearest Neighbour (DWKNN) classifier, and the K-Nearest Neighbour Equality (KNNE) classifier. The results show the importance of color features in discriminating different food classes and the superiority of the M-A&O SVM classifier in terms of classification accuracy.