NOVEL ENVELOPED-MORPHOLOGY HEART SOUND FEATURE EXTRACTION FOR INTELLIGENT CARDIOVASCULAR DISEASE PROGNOSIS IN E-HOME HEALTHCARE by YAO HaoDong Thesis Supervisor: Prof. VAI Mang I Thesis Co-supervisor: Prof. DONG MingChui Master of Science in Electrical and Computer Engineering Nowadays, cardiovascular diseases (CVDs) are the leading fatality causes around the world, coupled with a soaring increase. However, the traditional medical resources are limited and unevenly distributed, which is unable to dispose the CVDs crisis. Therefore, early CVDs detection is of great significance for decreasing the death. Heart auscultation has served as an important primary approach to prognosis CVDs for a long time. However, traditional auscultation is of distinct shortages, such as unable to record, replay, and store signals, subjecting to the experience of physicians. With the advent of computer analyzing technologies, echocardiogram (ECHO), electrocardiogram (ECG), photoplethysmogram (PPG) are invented for CVDs diagnosis. Unfortunately, these approaches are operated in hospital and with high expenses, which are not suitable for early detection. What’s more, heart auscultation is also employed by physicians before requesting for ECHO, ECG, etc. Therefore, automatic heart sound (HS) analysis is suggested as the competent approach for early CVDs detection, which is with the advantages: (1) automatic recording, storing, and replaying; (2) effective detection of early cardiac malfunction; (3) convenient using at home; (4) cost-efficient expenditure compared to traditional approaches. In the automatic HS analysis, the bottleneck problem is researching a HS analysis method which should accord with the requirements of multi-type applicability, high accuracy, reliable interpretation, and low complexity. Tackling this bottleneck problem, a novel enveloped-morphology HS feature extraction method is proposed in this thesis. By adapting discrete wavelet transform (DWT), heart murmur and non-murmur HS are separated into different sub-signals. Thereafter, the feature extraction processes for HS with heart murmur and non-murmur HS are designed respectively, since the envelope characteristics for classifying specific heart murmur are different from the characteristics for categorizing each concrete non-murmur HS. In feature extraction of heart murmur, the first step is abstracting the envelope of heart murmur, which is achieved by employing Shannon envelope algorithm as well as disposing the problems of identifying exact peak on each wave and alleviating the influence of residual fundamental HS. On the basis of abstracted envelopes, two diagnostic features of heart murmur are designated and extracted for further analysis. Regarded to the feature extraction of non-murmur HS, the envelope is obtained by using Shannon envelope algorithm. Thereafter, the exact peaks on the envelope are accurately detected by handling the conflict of threshold setting among different type non-murmur HS signals. Subsequently, feature parameters of non-murmur HS are extracted upon the detected peaks. Through experiments and classifications, the outcomes exhibit the classification accuracy is up to 94.4%. Furthermore, a comparison with other advanced HS analysis methods is taken, which strongly indicates the proposed method is able to classify multi-type HS signals in a high accuracy as well as with the least feature parameters.