Multi-path Heart Sound Detection Based on Wavelet Scattering and Attention Mechanism

Changes in heart sounds can often be the earliest signs of organic heart disease, and algorithms for the automatic classification of normal and abnormal heart sounds are essential for early diagnosis of cardiovascular disease. This work proposes a multi-path heart sound detection network based on wavelet scattering and attention mechanism algorithms. The wavelet scattering transform, long-short term memory (WST-LSTM) is used to extract the wavelet scattering features of the heart sound signal; the spatial attention network (SAN) is used to extract the spatial domain features of the time-frequency map of the heart sound signal; and the temporal attention network (TAN) is used to extract the time domain features of the time-frequency map. Sixteen features are extracted from each of the three paths, and after combining the features extracted from multiple path networks, a total of 48 fusion features are obtained, the final classification is performed using a support vector machine (SVM). The method proposed in this work has been well validated on the 2016 PhysioNet/CinC Challenge database, with accuracy (Acc), sensitivity (Se), specificity (Sp), and mean accuracy (Macc) of 96.94%, 97.42%, 96.82%, and 97.12%, respectively. This work also investigates the advantages of multi-path learning, discusses the effectiveness of different features and visualises them. The results show that the proposed method can effectively classify both normal and abnormal heart sound samples with good generalisation ability