Abstract
Recognizing normality and abnormality from heart sound recordings (phonocardiograms or PCG) has promote a scientific research in cardiology. However, only a small number of PCG recordings is publicly available. Also, current recognition approaches have not reached a satisfiable accuracy. In this paper, we apply transfer learning to automate the recognition for heart rates normalities and abnormalities. Mel Frequency Cepstrum Coefficients (MFCC) signal representation is adopted to transform the output feature into a PCG image which is used as an input to pre-trained Convolutional Neural Network (CNN) models. Then, several deep pre-trained CNN models are fine-tuned. The proposed approach is effective. Inception-ResNet-v2 pre-trained model achieves better performance than other models and other recent methods with an 89.5% of average classification accuracy using Pascal Heart Sound Challenge dataset.