A Voting Approach for Heart Sounds Classification Using Discrete Wavelet Transform and CNN Architecture

Cardiovascular diseases (CVDs) are a group of diseases that affect the heart or blood vessels and are the leading cause of mortality around the world. The main focus of this work is to classify heart sounds accurately before the condition of the heart worsens. Over the past few decades, audio-based classifications of medical data have received high consideration among various researchers. However, there are mainly two drawbacks while working with audio data. (1) variable length of the audio data and, (2) an inadequate number of class quintessential audio samples. In this work, reflection operation and the sliding window approach are employed to generate fixed-sized audio data. The proposed method applied Discrete Wavelet Transform (DWT) and Mel-Spectrogram for feature extraction. Furthermore, different augmentation techniques such as time-stretching and pitch-shifting are utilized in the method so that the proposed deep learning-based CNN architecture can be trained on a large amount of data. The proposed method is verified using two datasets provided by the ‘PASCAL Heart Sounds Challenge’, each of which contains a small number of heart sound samples of various lengths. In comparison, the experimental outcomes exhibit that the proposed approach outperforms many state-of-the-art methods with respect to sensitivity, specificity, precision, Youden index, etc.