Impact of data augmentation techniques on deep transfer learning for heart sound classification

Cardiovascular diseases remain one of the primary causes of death worldwide, highlighting the importance of early detection. PhonoCardioGraphy (PCG) provides a non-invasive approach to identify cardiac abnormalities. However, its effectiveness is limited by scarce datasets and signal variability. Deep learning (DL) models have shown promising results, but their performance strongly depends on the availability of sufficient training data. In this study, PCG signals were converted into spectrogram representations, and four data augmentation techniques were applied individually and in combination, followed by a systematic evaluation. A baseline convolutional neural network (CNN) and several transfer learning models were trained on both original and augmented datasets. The results demonstrate that combining augmentation techniques yields the best performance. The proposed CNN achieved an accuracy of 94%, representing a 5.6% improvement over its performance without data augmentation. In contrast, transfer learning such as VGG16, ResNet50, and DenseNet121 showed improvements of 3.4%, 3.4%, and 2.2%, respectively. These findings indicate that the proposed augmentation strategy effectively improves the classification performance and generalization capability of DL models for heart sound classification.