Porieva H. Methods of analysis of lung sounds for the state assessment of the human respiratory system

The thesis is devoted to the actual topic of human lung sounds processing and analysis in order to obtain diagnostically valuable parameters. These parameters are used both as stand-alone criteria for assessing the state of the human respiratory system, and for using them as input arguments of classifiers to automate decision making for certain diseases. In this thesis important scientific and technical task was solved: improving the methods of lung sounds’ preprocessing and their analysis for finding new diagnostically valuable parameters of lung sounds on the basis of mathematical apparatus of higher order statistics used for the classification of bronchopulmonary diseases. The advanced method for the initial processing of lung acoustic signals is described. This method is based on the filtering of noise signals caused by random environmental interference and deficiencies in the recording tool. The use of this method improves the quality of the lung sound signals to further obtain the informatively valuable parameters of lung sounds. This method is based on bidirectional filtering and helps to get rid of random bursts in the test signal. The method of separating individual breathing cycles in lung sounds is based on the allocation of respiratory cycles by means of spectral-temporal analysis of the incoming audio signal, synthesis of the main reference signal that simulates respiratory activity. Method of differentiation of normal and pathological respiratory noise based on the analysis of third-order cumulative functions and bispectral functions. In addition, the method of differentiation of crackling sounds and wet fine bubbling wheezes is also proposed. The developed method of diagnosing chronic obstructive pulmonary disease (COPD) and chronic bronchitis based on an iterative approach to find diagnostically valuable parameters of lung sounds, such as values of bispectrum and their corresponding frequencies, values of bicoherence functions and their corresponding frequencies, skewness and kurtosis coefficients is proposed in the thesis. This method also analyzes the parametric bispectrum. Some relationships and patterns between the group of these parameters and the corresponding category of the respiratory system state are determined. Different classifiers were examined with a different set of input arguments. Thus, the analysis determined that the best results were obtained using three classifiers: the support vector method, the decision tree, and neural networks. It was found that the best results were obtained by applying seven parameters of lung sounds obtained from higher order spectra. The general structure of the developed diagnostic software complex to provide an automated initial recommendation and a brief description of all stages of complex work are proposed.