Murmur Separation and Classification from Heart Sound Using Constrained Singular Spectrum Analysis and Wavelet Transform

Detection and classification of heart murmurs are crucial for the early and timely diagnosis and treatment of cardiovascular diseases. However, heart murmurs often overlap with normal heart sounds, making them difficult to detect. In this paper, we propose a two-stage method for separating murmurs. The first stage employs Constrained Singular Spectrum Analysis (CSSA) to utilize the distinct statistical characteristics of heart sounds and murmurs, facilitating their initial separation. The second stage leverages the capability of Wavelet Transform (WT) to effectively localize sound components in both the time and frequency domains, which is essential for separating subtle murmurs that were not completely removed during the first stage. After these two stages, normal heart sounds are precisely extracted, and the remaining audio components, predominantly murmurs, can be separated for classification. Our results demonstrate that pre-separating murmurs enhances the classification performance, achieving an accuracy of over 98%. The proposed method can improve the clarity and accuracy of heart sound analysis, thereby facilitating better diagnostic evaluations and interventions in clinical settings.