Automated heart murmur detection has become increasingly relevant to aid the medical diagnosis. The George B. Moody PhysioNet Challenge 2022 proposed a similar problem to encourage research in this field, by providing a manually labeled dataset of 942 patients.
Our approach focuses on a classification method using segmented S1, S2, systole, and diastole sections from the given recordings. The segmentation stage was performed by detecting the heart sounds, then with further refinement the systole and diastole regions were also determined. Heart sound detection process is based on Shannon entropy envelope and a proper moving average filter on the given detail level of the discrete wavelet transform. Each initial detection was obtained from the positive slope zero crossings of the instantaneous phase of the given smoothed entropy envelope.
Different features were extracted from each record. Multiple frequency features were obtained, such as the highest frequency standard deviation in the spectrum of each section in the 20–40 Hz region. In the time-domain, the most prominent features were the average length of the sections and their standard deviation, as well as the root mean square and the kurtosis of a section. The dimensionality of the extracted feature-space was reduced by principal component analysis and for classification a support vector machine was trained.
The recordings also contained breathing noises in several cases, these signals could be erroneously classified as containing murmurs. To mitigate this behavior, spectral analysis was used on the envelope of the given wavelet decomposed signal, since breathing has different periodicity from heart murmurs.