Presented here is an examination of the issues surrounding the analysis of lung sounds and their display. The project is aimed at providing a visual representation of the information that a physician gleans from auscultation of the lungs. Such a tool would be of benefit to those who are hearing impaired and also in teaching auscultation. A second goal is to provide a tool that will allow the examination and quantification of lung sounds thus permitting linkage between the acoustic events and their physical causes. The project is divided into two tasks. The first is the isolation of the wheezes and crackles; the second is their display. The isolation problem is difficult due to the variance in the frequency characteristics of the sounds; wheezes may appear anywhere in a two thousand hertz band and crackles also display a varying spectrum. The difficulty in separation is further compounded by the spectral overlap of the two. These problems preclude any 'simple' filter solution. In order to separate the sounds, filtering methods based on exploiting the statistical differences namely the stationarity of the wheeze and non-stationarity of the crackle are utilized. Of the several methods attempted, the most promising was the Adaptive Line Enhancement process when driven by the Least Mean Squares adaptive algorithm. An important criteria for being able to display the sounds was to access their temporal information. Accomplishing this with the standard short time Fourier transform precludes adequate resolution to identify the frequency characteristics of crackles. Display of the crackle information was facilitated by the use of high resolution time-frequency methods based on Cohen's Class of time-frequency representations. These methods are able to simultaneously provide high time and frequency resolution. A method for automatic adjustment of the parameters involved in the process was developed in order to yield the best display possible. / Thesis / Master of Engineering (ME)

Thesis (MScEng (Mechanical and Mechatronic Engineering))--University of Stellenbosch, 2009. / Pulmonary tuberculosis is a common and potentially deadly infectious disease, commonly affecting the respiratory area. Over one-third of the world’s population is infected with the tuberculosis bacterium. Since pulmonary tuberculosis damages the respiratory area, the sound properties of infected lungs differ from those of non-infected lungs. However, auscultation is often ruled out as a reliable diagnostic technique due to the random position and severity of damage to the lungs as well as requiring the personal and trained judgment of an experienced medical practitioner. This project investigates a possible improvement in the pulmonary diagnostic and treatment field by applying electronic and computer-aided sound analysis techniques to analyze respiratory actions beyond human audible judgment. Respiratory sounds of both healthy subjects and subjects who were infected with pulmonary tuberculosis were recorded from seven locations per lung on both the posterior and anterior chest walls, using self-designed hardware. Adaptive filtering signal and analysis techniques yielded a wide range of signal features. This included analysis for time, frequency and both wheeze and crackle adventitious respiratory sounds. Following the analysis, statistical methods identified the most attractive signal measurements capable of separating the recordings of healthy and unhealthy respiratory sounds. Selected signal features were used with neural network optimization to obtain a successful implementation for the semi-automated identification of healthy and unhealthy respiratory sounds originating from pulmonary tuberculosis, with a performance of over 80% for sensitivity, specificity and accuracy. The success of categorizing the recordings justifies the capabilities of the digital analysis of respiratory sounds and supports an argument for further research and refinement into the assessment of pulmonary tuberculosis by electronic auscultation. Further research is recommended, with improvements justified and highlighted in this report.

Nas últimas décadas uma considerável parcela das pesquisas nas áreas de Física e Matemática tem sido dedicada ao estudo de fenômenos não lineares. Uma possível explicação para isso foi o rápido desenvolvimento de sistemas computacionais, tanto em nível de hardware quanta em nível de software, algoritmos e técnicas de programação que propiciaram ao homem maiores facilidades no tratamento de sistemas não lineares, o que levou a um maior grau de entendimento de sua complexidade. Geralmente, aos sistemas não lineares esta associada uma geometria irregular, onde comum o aparecimento de regimes caóticos, com um conjunto atrator de órbitas cuja dimensão não é um inteiro positivo, mas sim um número real positivo. Por esta razão, tais atratores, são denominados estranhos e ditos possuírem uma geometria fractal. É possível, através de métodos cuidadosamente desenvolvidos, estimar-se as dimensões associadas à dinâmica de séries temporais. Uma das séries de maior dificuldade de análise através do computador, e de particular interesse na medicina, são as séries de sons pulmonares humanos. Desde quando o estetoscópio foi inventado até os dias de hoje não há uma ferramenta plenamente confiável para a análise destas séries. Recentemente, temos trabalhado com estas séries e verificamos que há uma geometria fractal. Esta tese propõe a utilização da análise não-linear para identificação de padrões sonoros. Além da geometria fractal, a análise por wavelets tem sido utilizada no estudo de sinais complexos, sobretudo naqueles que apresentam estruturas fractais. O conjunto de filtros construído através da translação, expansão ou compressão de uma função wavelet mãe tem uma estrutura auto-similar, mostrando-se particularmente apropriado para a verificação da auto similaridade dos sons. A técnica da estimativa dos expoentes de Lyapunov dependente do tempo, a qual e desenvolvida na tese, tem se mostrado bastante adequada para identificação de padrões sonoros de origem pulmonar. / It has been observed that in the last decades, considerable amount of the research in the areas of Physics and Mathematics have been dedicated to the study of nonlinear phenomena. A possible explanation for this fact is the fast development of computational systems occurring in the level of the hardware as in computer languages, algorithms and programming techniques. These developments propitiated to the researchers a broader contact with nonlinear systems, which led to a better understanding of their complexity. In general, for nonlinear systems an irregular geometry is associated, where the appearance of chaotic regimes has an associated attractor set of orbits whose dimension is not a positive integer number, but a real one. Such attractors are called strange and said to possess fractal geometry. It is possible, through carefully developed methods, to estimate the dimension associated to the dynamics of time series. One of the series with high difficulty to be analyzed through a computer and of particular interest in medicine, is the time series generated out of human pulmonary sounds. Since the creation of the stethoscope, there is not yet a fully trustworthy tool for the lung sound analysis. Recently, we have studied these series and verified that they have a fractal geometry nature. The purpose of this thesis is to investigate non-linear analysis as a tool for pattern recognition in lung sounds. In addition to fractal geometry, the wavelet analysis has been used in the study of complex signs, in particular for those presenting a fractal structure. The set of filters constructed through the translation, expansion or compression of a function wavelet mother has an auto-similar structure, being particularly useful for the verification of self similarity of pulmonary sounds. The largest time dependent Lyapunov exponent estimation technique that has been proposed in this thesis has shown a high degree of confidence for the identification of lung sound patterns.

Nas últimas décadas uma considerável parcela das pesquisas nas áreas de Física e Matemática tem sido dedicada ao estudo de fenômenos não lineares. Uma possível explicação para isso foi o rápido desenvolvimento de sistemas computacionais, tanto em nível de hardware quanta em nível de software, algoritmos e técnicas de programação que propiciaram ao homem maiores facilidades no tratamento de sistemas não lineares, o que levou a um maior grau de entendimento de sua complexidade. Geralmente, aos sistemas não lineares esta associada uma geometria irregular, onde comum o aparecimento de regimes caóticos, com um conjunto atrator de órbitas cuja dimensão não é um inteiro positivo, mas sim um número real positivo. Por esta razão, tais atratores, são denominados estranhos e ditos possuírem uma geometria fractal. É possível, através de métodos cuidadosamente desenvolvidos, estimar-se as dimensões associadas à dinâmica de séries temporais. Uma das séries de maior dificuldade de análise através do computador, e de particular interesse na medicina, são as séries de sons pulmonares humanos. Desde quando o estetoscópio foi inventado até os dias de hoje não há uma ferramenta plenamente confiável para a análise destas séries. Recentemente, temos trabalhado com estas séries e verificamos que há uma geometria fractal. Esta tese propõe a utilização da análise não-linear para identificação de padrões sonoros. Além da geometria fractal, a análise por wavelets tem sido utilizada no estudo de sinais complexos, sobretudo naqueles que apresentam estruturas fractais. O conjunto de filtros construído através da translação, expansão ou compressão de uma função wavelet mãe tem uma estrutura auto-similar, mostrando-se particularmente apropriado para a verificação da auto similaridade dos sons. A técnica da estimativa dos expoentes de Lyapunov dependente do tempo, a qual e desenvolvida na tese, tem se mostrado bastante adequada para identificação de padrões sonoros de origem pulmonar. / It has been observed that in the last decades, considerable amount of the research in the areas of Physics and Mathematics have been dedicated to the study of nonlinear phenomena. A possible explanation for this fact is the fast development of computational systems occurring in the level of the hardware as in computer languages, algorithms and programming techniques. These developments propitiated to the researchers a broader contact with nonlinear systems, which led to a better understanding of their complexity. In general, for nonlinear systems an irregular geometry is associated, where the appearance of chaotic regimes has an associated attractor set of orbits whose dimension is not a positive integer number, but a real one. Such attractors are called strange and said to possess fractal geometry. It is possible, through carefully developed methods, to estimate the dimension associated to the dynamics of time series. One of the series with high difficulty to be analyzed through a computer and of particular interest in medicine, is the time series generated out of human pulmonary sounds. Since the creation of the stethoscope, there is not yet a fully trustworthy tool for the lung sound analysis. Recently, we have studied these series and verified that they have a fractal geometry nature. The purpose of this thesis is to investigate non-linear analysis as a tool for pattern recognition in lung sounds. In addition to fractal geometry, the wavelet analysis has been used in the study of complex signs, in particular for those presenting a fractal structure. The set of filters constructed through the translation, expansion or compression of a function wavelet mother has an auto-similar structure, being particularly useful for the verification of self similarity of pulmonary sounds. The largest time dependent Lyapunov exponent estimation technique that has been proposed in this thesis has shown a high degree of confidence for the identification of lung sound patterns.

Nas últimas décadas uma considerável parcela das pesquisas nas áreas de Física e Matemática tem sido dedicada ao estudo de fenômenos não lineares. Uma possível explicação para isso foi o rápido desenvolvimento de sistemas computacionais, tanto em nível de hardware quanta em nível de software, algoritmos e técnicas de programação que propiciaram ao homem maiores facilidades no tratamento de sistemas não lineares, o que levou a um maior grau de entendimento de sua complexidade. Geralmente, aos sistemas não lineares esta associada uma geometria irregular, onde comum o aparecimento de regimes caóticos, com um conjunto atrator de órbitas cuja dimensão não é um inteiro positivo, mas sim um número real positivo. Por esta razão, tais atratores, são denominados estranhos e ditos possuírem uma geometria fractal. É possível, através de métodos cuidadosamente desenvolvidos, estimar-se as dimensões associadas à dinâmica de séries temporais. Uma das séries de maior dificuldade de análise através do computador, e de particular interesse na medicina, são as séries de sons pulmonares humanos. Desde quando o estetoscópio foi inventado até os dias de hoje não há uma ferramenta plenamente confiável para a análise destas séries. Recentemente, temos trabalhado com estas séries e verificamos que há uma geometria fractal. Esta tese propõe a utilização da análise não-linear para identificação de padrões sonoros. Além da geometria fractal, a análise por wavelets tem sido utilizada no estudo de sinais complexos, sobretudo naqueles que apresentam estruturas fractais. O conjunto de filtros construído através da translação, expansão ou compressão de uma função wavelet mãe tem uma estrutura auto-similar, mostrando-se particularmente apropriado para a verificação da auto similaridade dos sons. A técnica da estimativa dos expoentes de Lyapunov dependente do tempo, a qual e desenvolvida na tese, tem se mostrado bastante adequada para identificação de padrões sonoros de origem pulmonar. / It has been observed that in the last decades, considerable amount of the research in the areas of Physics and Mathematics have been dedicated to the study of nonlinear phenomena. A possible explanation for this fact is the fast development of computational systems occurring in the level of the hardware as in computer languages, algorithms and programming techniques. These developments propitiated to the researchers a broader contact with nonlinear systems, which led to a better understanding of their complexity. In general, for nonlinear systems an irregular geometry is associated, where the appearance of chaotic regimes has an associated attractor set of orbits whose dimension is not a positive integer number, but a real one. Such attractors are called strange and said to possess fractal geometry. It is possible, through carefully developed methods, to estimate the dimension associated to the dynamics of time series. One of the series with high difficulty to be analyzed through a computer and of particular interest in medicine, is the time series generated out of human pulmonary sounds. Since the creation of the stethoscope, there is not yet a fully trustworthy tool for the lung sound analysis. Recently, we have studied these series and verified that they have a fractal geometry nature. The purpose of this thesis is to investigate non-linear analysis as a tool for pattern recognition in lung sounds. In addition to fractal geometry, the wavelet analysis has been used in the study of complex signs, in particular for those presenting a fractal structure. The set of filters constructed through the translation, expansion or compression of a function wavelet mother has an auto-similar structure, being particularly useful for the verification of self similarity of pulmonary sounds. The largest time dependent Lyapunov exponent estimation technique that has been proposed in this thesis has shown a high degree of confidence for the identification of lung sound patterns.