A fusion framework based on multi-domain features and deep learning features of phonocardiogram for coronary artery disease detection

Phonocardiogram (PCG) signals reflect the mechanical activity of the heart. Previous studies have reported that PCG signals contain heart murmurs caused by coronary artery disease (CAD). However, the murmurs caused by CAD are very weak and rarely heard by the human ear. In this paper, a novel feature fusion framework is proposed to provide a comprehensive basis for CAD diagnosis. A dataset containing PCG signals of 175 subjects was collected and used. A total of 110 features were extracted from multiple domains, and then reduced and selected. Images obtained by Mel-frequency cepstral coefficients were used as the input for the convolutional neural network for feature learning. Then, the selected features and the deep learning features were fused and fed into a multilayer perceptron for classification. The proposed feature fusion method achieved better classification performance than multi-domain features or deep learning features alone, with accuracy, sensitivity, and specificity of 90.43%, 93.67%, and 83.36%, respectively. A comparison with existing studies demonstrated that the proposed method was a promising noninvasive screening tool for CAD under general medical conditions.