Le-LWTNet: A Learnable Lifting Wavelet Convolutional Neural Network for Heart Sound Abnormality Detection

Automatic heart sound abnormality detection plays a vital role in the preliminary diagnosis of cardiovascular diseases (CVDs). Many handcraft-designed or learning-based methods have been proposed in recent years. However, due to the influence of the environment, the divergence of different stethoscopes, and the data collection protocol, the pattern of heart sound signals are so complex that fixed pattern feature extraction or learning features directly from the signal cannot enough lead to the final accurate classification. For this issue, a learnable lifting wavelet transform (Le-LWT) block, which embeds the trainable convolutional neural network (CNN) into the lifting wavelet transform (LWT), is proposed in this article. Le-LWT can utilize the nonlinear learning ability of CNN while maintaining the multiresolution time–frequency analysis ability of wavelet transform (WT), as well as more interpretation than the deep networks designed as black boxes. Based on the Le-LWT module, we propound an end-to-end Le-LWTNet that has stronger nonlinear characterization capabilities and few parameters for automatic abnormality detection of the heart sound. Experimental evaluations are performed on a ten-fold cross-validation task using the 2016 PhysioNet/CinC Challenge dataset (PCCD) and the new publicly available pediatric heart sound dataset (PHSD) we collected. Results demonstrate that the proposed method excels the state-of-the-art models both in abnormality detection and parameter consumption. Moreover, the proposed method can give an interpretation of the current classification results, which will help doctors do a second review.