Wavelet Feature Extraction with Deep Learning Models for Heart Sound Recognition
Abstract
Currently, heart disease has risen significantly in recent years due to demographic changes. Both men and women are showing more cases of abnormalities detectable through phonocardiogram (PCG) signals. These signals are widely used to diagnose for identifying a wide range of cardiovascular conditions, including arrhythmias, valvular heart diseases, congenital heart defects, cardiomyopathies, and heart failure. In this research, we apply wavelet scattering transform (WST) and continuous wavelet transform (CWT) as wavelet analysis techniques. For the Deep learning component, we incorporate three primary architectures: a one-dimensional CNN (WaveNet), a two-dimensional CNN (MLNet) and a self-attention Transformer to classify heart sounds from phonocardiogram (PCG). By identifying heart sounds using phonocardiogram (PCG) data, the ensemble technique of our deep learning model achieves better performance than recent studies and is effectively applied to diagnosing and predicting cardiac disorders.