Heart Murmur and Abnormal PCG Detection via Wavelet Scattering Transform and 1D-CNN

Abstract

Congenital heart disease (CHD) is the most common type of congenital anomaly, with an estimated prevalence of 8–12 per 1000 live births. CHD results in heart murmurs, which once listened to provide valuable information about mechanical activity of the heart and aid in diagnosis of CHD and other heart valve diseases (HVDs). This work does automatic and accurate heart murmur detection from phonocardiogram (PCG) recordings. Two publicly available PCG datasets from PhysioNet online database are utilized to train and test various custom neural networks (NNs). We first do preprocessing which includes the following key steps: denoising, segmentation, relabeling of noise-only segments, data normalization, and time-frequency analysis of the PCG segments using wavelet scattering transform (WST), mel-frequency cepstral coefficients (MFCCs), and short-time Fourier transform (STFT). We then conduct four experiments, first three (E1–E3) using first dataset, and fourth (E4) using second dataset. It turns out that our custom 1-D convolutional neural network (1D-CNN) with wavelet scattering outperforms all other models. Furthermore, the vanilla 1D-CNN model (with wavelet scattering) outperforms the related work in terms of weighted accuracy, precision, and specificity, for experiment E3. As for experiment E1, our model performs quite close to top-performing work in terms of weighted accuracy, outperforms related works in terms of precision and is on par with the related works in terms of ${F}1$ -score.