A Lightweight 1-D Convolution Neural Network Model for Multi-class Classification of Heart Sounds

Analysis of heart sound signals provides ample features to diagnose cardiovascular diseases (CVDs) at an early stage. However, the role of cardiac auscultation is limited only to performing the preliminary screening. It is due to the subjectivity of nature in the diagnosis. Automatic analysis of heart sounds will address this issue as well as it will also reduce the burden of the already-stretched medical facility. This paper proposes a lightweight 1D-CNN model to analyse and classify heart sound signals into five categories. The CNN model is trained on the multi-resolution domain features obtained using the discrete wavelet transform (DWT). The signal is first pre-processed and then decomposed up to five levels using ’coif5’ as the mother wavelet. The obtained detailed level and approximation level coefficients are applied to the 1-D CNN model. The proposed method yields 98.9% accuracy, 99.01% sensitivity, and 99.72% specificity, showing the proposed method’s superiority on various methods proposed in the literature recently.