Dinesh Kumar Dinesh Dinesh Kumar CARDIOVASCULAR DISEASE ASSESSMENT DISEASE CARDIOVASCULAR AUTOMATIC HEART FOR SOUND AUTOMATIC ANALYSIS AUTOMATIC HEART SOUND ANALYSIS FOR CARDIOVASCULAR DISEASE ASSESSMENT Doctoral thesis submitted to the Doctoral Program in Information Science and Technology, supervised by Prof. Dr. Paulo Fernando Pereira de Carvalho and Prof. Dr. Manuel de Jesus Antunes, and presented to the Department of Informatics Engineering of the Faculty of Sciences and Technology of the University of Coimbra. September 2014 OIMBRA C E D NIVERSIDADE NIVERSIDADE U September 2014 Thesis submitted to the University of Coimbra in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Information Science and Technology This work was carried out under the supervision of Professor Paulo Fernando Pereira de Carvalho Professor Associado do Departamento de Engenharia Informática da Faculdade de Ciências e Tecnologia da Universidade de Coimbra and Professor Doutor Manuel J Antunes Professor Catedrático da Faculdade de Medicina da Universidade de Coimbra ABSTRACT Cardiovascular diseases (CVDs) are the most deadly diseases worldwide leaving behind diabetes and cancer. Being connected to ageing population above 65 years is prone to CVDs; hence a new trend of healthcare is emerging focusing on preventive health care in order to reduce the number of hospital visits and to enable home care. Auscultation has been open of the oldest and cheapest techniques to examine the heart. Furthermore, the recent advancement in digital technology stethoscopes is renewing the interest in auscultation as a diagnosis tool, namely for applications for the homecare context. A computer-based auscultation opens new possibilities in health management by enabling assessment of the mechanical status of the heart by using inexpensive and non-invasive methods. Computer based heart sound analysis techniques facilitate physicians to diagnose many cardiac disorders, such as valvular dysfunction and congestive heart failure, as well as to measure several cardiovascu- lar parameters such as pulmonary arterial pressure, systolic time intervals, contrac- tility, stroke volume, etc. In this research work, we address the problem of extracting a diagnosis using analysis of the heart sounds. Heart sound analysis consists of three main tasks: i) identification of non-cardiac sounds which are unavoidably mixed with the heart sound during auscultation; ii) segmentation of the heart sound in order to localize the main sound component; and finally, iii) classification of the abnormal heart sounds from the normal heart sounds and in case of abnormal sounds classification is performed further to identify the type of the abnormal sound. In most previous works on heart sound analysis these three problems were tackled jointly. Two main classes of approaches are found in past years: first is to perform analysis using an auxiliary signal, such as ECG, sound signal for ambient noise, carotid pulse, etc which is acquired during auscultation; and the second is without the support of any auxiliary signal. The first group of methods can be ruled out for the sake of keeping analysis cost effective and convenient to the users, how- ever, the second group of approaches is dependent on the use of regularity of heart sound components which may not be found in various cardiac disorders. Hence, our intention aims to develop algorithms for the each problem related to heart sound analysis with the following main requirements: i) not to use any auxiliary; ii) appli- cable to any age, weight, sex, body proportion subject; iii) robustness. Noise detection technique uses the template matching based approach. We pro- pose a new method applicable in real time to detect ambient and internal body nois- iii iv Abstract es manifested in heart sound during acquisition. The algorithm is designed on the basis of the periodic nature of heart sounds and physiologically inspired criteria. A small segment of uncontaminated heart sound exhibiting periodicity in time as well as in the time-frequency domain is first detected and applied as a reference signal in discriminating noise from the sound. For the segmentation problem, the heart is considered as a nonlinear dynamical system. In a first processing stage, Lyapunov exponents are computed from the phase space in order to identify the presence of murmur in the heart sound. Based on this information, the method enters one of two segmentation methods: for heart sound with murmur, an algorithm based on adaptive wavelet-nonlinear feature analysis is applied; for heart sounds without murmur, a less complex approach is followed based on the signal’s envelope analysis. Recognition of S1 and S2 sounds is achieved by inspecting high-frequency signatures. This marker is physiologically motivated by the accentuated pressure differences found across heart valves, both in native and prosthetic valves, which leads to distinct high-frequency signatures of the valve closing sounds. Since heart murmurs originate from numerous anomalies in the heart, these show different characteristics, temporal, frequency and complexity. These features are extracted in order to classify heart murmur using a supervised classifier. A new set of 17 features extracted in the time, frequency and in the state space domain is suggested. The features applied for murmur classification are selected using the floating sequential forward method (SFFS). Using this approach, the original set of 17 features is reduced to 10 features. The classification results achieved using the proposed method are compared on a common database with the classification results obtained using the feature sets proposed in two well-known state of the art methods for murmur classification. These algorithms have been tested on the database prepared with the help of the University Hospital in Coimbra. Three classes of the heart sound databases were prepared, i) normal heart sounds from the native valves; ii) abnormal heart sounds from the native valves; iii) heart sounds from artificial valve implants. Ex- perimental results suggest that the algorithms can be applicable for cardiac applica- tion. Keywords: Heart sound, valvular disease, prosthetic valve, segmentation, heart murmur, classification, noise detection, time-frequency analysis, wavelet decomposi- tion, nonlinear dynamics. SUMÁRIO As doenças cardiovasculares são a maior causa de morte em todo o mundo, ultrapas- sando de forma significativa a mortalidade devida aos diabetes e ao cancro. Dado o envelhecimento acentuado da população mundial e atendendo a que a incidência das doenças crónicas, em particular das doenças cardiovasculares, está fortemente corre- lacionada com a idade, observa-se uma nova tendência de cuidados de saúde focada nos cuidados de saúde preventivos, com vista a reduzir o número de episódios agu- dos numa tentativa de reduzir custos e de melhor a qualidade de vida dos pacientes. A implementação destas estratégias requer sistemas fiáveis de telemonitorização, porém simples e baratos por forma a permitir o seu uso prolongado por populações pouco motivadas e muitas vezes com iliteracia tecnológica. A auscultação é a técnica mais antiga e menos onerosa para examinar o coração. Os recentes avanços em estetoscópios com tecnologia digital está a renovar o interesse no uso da auscultação como ferramenta de diagnóstico, nomeadamente para aplica- ções no contexto da tele-monitorização e da gestão da doença. Uma auscultação as- sistida por computador abre novas possibilidades na gestão da saúde ao permitir a avaliação do estado mecânico do coração, usando-se métodos não invasivos e pouco onerosos. As técnicas de análise de som cardíaco assistidas por computador facilitam o diagnóstico médico de muitos problemas cardíacos, tais como a disfunção valvular e a insuficiência cardíaca congestiva, permitindo também medir vários parâmetros cardiovasculares, como por exemplo a pressão arterial pulmonar, intervalos de tem- po sistólicos, contractilidade, volume sistólico, entre outros. Neste trabalho de investigação abordamos o problema de produzir um diagnóstico usando a análise de sons cardíacos. De uma forma geral, a análise de sons cardíacos consiste em três tarefas principais: i) na identificação de contaminações por sons não cardíacos, que se misturam inevitavelmente com o som do coração durante a auscul- tação; ii) na segmentação do som cardíaco por forma a localizar as sua principais componentes; e finalmente iii) na classificação dos sons cardíacos anormais e distin- ção dos sons cardíacos normais. No caso da existência de sons anormais, procede-se a uma classificação, por forma a identificar o tipo de som irregular. Na literatura sobre análise de som cardíaco, estes três problemas têm sido aborda- dos conjuntamente. De uma forma geral, podemos afirmar que emergiram duas clas- ses de abordagens principais: a primeira classe consiste na condução da análise usando um sinal auxiliar, tal como o ECG, um sinal sonoro para o ruído ambiente, o v vi Abstract pulso da carótida, etc., sinal esse que é adquirido durante a auscultação; na segunda classe de métodos a tarefa é realizada sem o suporte a qualquer sinal auxiliar. O primeiro grupo de métodos não é muito prático para aplicações clínicas, dado o seu custo superior e o facto de exigir um número de sensores mais elevado o que, neces- sariamente, obriga a uma complexidade logística e operacional superior. No entanto, o segundo grupo de abordagens depende do uso da regularidade de componentes