| A number of gases in the breath are known to be indicators of the presence of diseases and clinical conditions. These gases have been identified as biomarkers of the according diseases, and measurement of the development of the diseases. This thesis investigates the potential of breath signals analysis as reliable channels for disease identification, by coupling a breath analysis device with appropriate data analysis methods. The device employs 12 chemical sensors that are especially sensitive to the biomarkers in human breath to obtain distinguishable samples. A set of data analysis issues, including sensor selection, physiological feature extraction, and classifier design, are investigated in order to achieve the best disease identification accuracy. In the device, since each sensor has a specific contribution in identifying a type of disease, it is not necessary, even disadvantageous to use all of the sensors for a specific application. This thesis proposes a sensor selection technique for the particular task of disease identification, by computing the weight of each sensor via LDA. To find the most relevant features, this thesis extracts a variety of features from multiple analysis domains. Then, a mathematical model based on Gaussian functions is developed, to extract coefficient features from the original signal, which is proven especially useful when clustering samples that belong to the same category. Additionally, this thesis proposes two classification methods especially for different applications. For disease diagnosis, it expresses an input sample as the linear combination of all training samples. The coefficients of the linear combination provide useful cues for classification. For physical condition measurement, support vector ordinal regression method is used to find ordinal hyper planes, which separate the training data into different ordered classes. Finally, various applications of breath analysis are presented. Breath samples from healthy persons and patients known to be afflicted with various diseases are collected by the proposed device. The applications including disease diagnosis, diabetes condition monitoring, and evaluation of medical treatment of renal failure are introduced and the performances are evaluated by using the proposed pattern analysis methods. The results show that the system has satisfactory performance in these applications.;Key words: breath analysis, disease identification, diabetes condition monitoring, medical treatment evaluation, pattern recognition, chemical sensor. |
