University of Macau Abstract DIAGNOSE CVDS THROUGH DETECTING HEART SOUNDS by WANG Haiyang Thesis Supervisor: Prof. VAI Mang I Thesis Co-supervisor: Prof, DONG Mingchui Department of Electrical and Computer Engineering Cardiovascular diseases(CVDs)have become a great threat to human’ lives aggressively around the world. Currently, the fatality of cardiovascular diseases(CVDs) represents one of the global primary healthcare challenges and necessitates broader population checking for earlier intervention. The American Heart Association studies predict that 40.5% of US population will have some form of CVD by 2030 with associated indirect cost reaching $ 276 billion [1]. CVD affects individuals in their peak mid-life years disrupting the future of the families dependent on them and undermining the development of nations by depriving valuable human resources in their most productive years. An early detection and prevention care is much significant and is a big challenge on the global scale. Clinical check of CVDs is costly and prolonged, such as echocardiogram (ECHO), electrocardiography (ECG), etc. [2]. Besides the cost, they are inconvenient and patients themselves cannot often be checked away from clinics. Based on heart sound (HS), auscultation could also find signs of pathologic conditions conveniently, like many cardiac abnormalities including valvular heart disease, congestive heart failure and congenital heart lesions etc. However, the traditional auscultation involves subjective judgment by the clinicians, and needs long period experience and clinics training. With the development of hardwares and sensors, it becomes a possible chance to realize auto-auscultation. But we have to be aware of limitation-resource of the hardwares and the importance of accuracy in health field. These reasons lead the research on HS signal analysis in order to provide faster, better and more cost-effective healthcare support for the victims of CVDs. HS signals are typically dynamic, complex and non-stationary. For analysis of HS signal, many approaches, such as wavelet decomposition and reconstruction method[3-4], short time Fourier transform (STFT) method [5, 6], and S-transform method [7.8], have been proposed in literatures. Most will solely analyze the time frequency domain for feature extractions. This leads to a problem that feature extractions are less aligned with medical knowledge. From acoustic perspective, a doctor diagnoses the CVDs generally through a stethoscope. Common descriptive terms about what it sounds like in auscultation include rumble, blowing, machinery, scratchy, harsh, gallop, ejection, click, drum, cymbal, etc.From acoustic perspective to analyze such a bio-signal, its main advantage is that engineering is directly aligned with clinical diagnosis. For instance, a diastolic sound resembling the third sound but earlier in timing characterizes the pericardial “knock" sound of pericardial constriction; when such a sound is combined with careful evaluation of the precordial motion and jugular venous pulse, the diagnosis of congestive heart failure is strongly supported. [14] The mel-frequency cepstral coefficient (MFCC) method has been utilized for HS feature extraction [9-13], which is based on the theory that human audition spaces linearly at low frequency band and logarithmically at high frequency. Unfortunately, it is a segmentation-based feature extraction technique and its effect suffers from segmentation error greatly. Due to multi-transforms, its complexity is high. This is a big burden and challenge for portable devices as data of HS would be handled frequently or even continually. Tackling with these challenges and problems, a novel multidimensional model is creatively proposed for HS analysis. The model based on traditional timbre analysis could extract the timbre of HS. When the pathology changes in heart or blood vessels, different timbre information can be sampled, from which can trace back the reason causing such a timbre. Its performance for characterizing features with diagnostic significance from HS is evaluated through experiments. As the model for the HS feature extraction adapts three timbre attributes to interpret HS signal acoustically, it aligns engineering with medical knowledge much more. Without segmentation, it voids computation complexity and mistakes introduced. Hereby, the proposed model shows great potential for pervasive health monitoring with intelligent auscultation function. Finally through experiements and classification, the outcomes exhibit the propsed meathod is low complexity and high accurancy. This will provide people with the low cost, high efficient and reliable health monitor and management. Keywords: Cardiovascular Diseases, Timbre, e-mobile Healthcare, Multidimensional Model, and Segmentation-Free