Heart sound analysis

• Main Author: Xu, Ran
• Other Authors: Ser Wee
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Wireless communication systems
• Online Access: http://hdl.handle.net/10356/64301
• Institution: Nanyang Technological University

Heart auscultation is a common diagnosis for heart diseases, but the skill is highly dependent on personal experience. Technologies have already made the digital signal-processing methods practical for this application. This project was aimed to analyze the heart sound waveform and find an algorithm to extract features efficient for classification. In the project, Matlab was used as the work platform and on which discrete wavelet transform was applied for feature extraction, combining with the method of principal components analysis (PCA) on the detail wavelet coefficients for feature selection, the heart sound samples were classified with the application of support vector machine (SVM). The kurtosis of the detailed wavelet coefficients were firstly used as feature vectors for classification, then the parameters after PCA were selected. Different wavelets at different levels were attempted for premier results. In total 10 NHS samples and 10 AS samples downloaded from websites [8-11] were used in this project, and with the decomposition and reconstruction using wavelet ‘db4’ at fourth level, after PCA, with the 5 most essential parameters chosen as the feature vector, the algorithm achieved an accuracy of 80% of classification for normal heart sound (NHS) and aortic stenosis (AS), where cross validation classification method was used to overcome the limited number of samples. In addition, according to the median differences for one parameter between two classes as well as their overlap areas, the parameters after PCA were reordered as a purpose to determine the best performed feature for classification. This proposed method had achieved an average accuracy of 90% for the classification with only the first 2 parameters chosen as feature vector and with noise injected.