Enhancing Heart Sounds Classification Using MFCC And CNN

Abstract

Cardiovascular disease remains one of the leading causes of death worldwide, which increases the need for accurate and efficient early detection methods. In this paper, we utilize heart sound analysis as a non-invasive screening approach to distinguish normal and abnormal cardiac conditions. We apply Mel-Frequency Cepstral Coefficients (MFCC) to extract discriminative spectral features from heart sound recordings and use three Convolutional Neural Network (CNN) architectures, including AlexNet, VGG16, and ResNet18, for classification. To improve model robustness and reduce overfitting, we implement audio data augmentation techniques, including white noise addition, pitch scaling, time stretching, and random gain adjustment. We train all models using a batch size of 32, 25 epochs, and a learning rate of 0.0001. The experimental results show that this learning rate provides stable convergence and optimal performance across architectures. AlexNet achieves the highest accuracy of 100%, followed by VGG16 with 99.5% and ResNet18 with 97%. Overall, this paper demonstrates that the combination of MFCC feature extraction, data augmentation, and CNN modeling provides highly accurate and reliable heart sound classification, with strong potential for practical clinical screening applications.