Heart murmur detection with deep learning based mel spectrogramanalysis from phonocardiogram data

Heart auscultation is a fundamental clinical method used to evaluate heart sounds and detect valve diseases. However, its accuracy largely depends on the clinician's experience, which may lead to variability in diagnostic outcomes. This study proposes a deep learning–based automatic system for classifying heart sounds to reduce these limitations. Phonocardiogram (PCG) signals are first transformed into Mel-spectrogram representations to extract discriminative time–frequency features. These features are used to classify heart sounds into four categories: normal, murmur-related abnormal, normal outcome murmurs, and abnormal sounds without murmurs. Convolutional Neural Networks (CNNs) are employed, utilizing transfer learning architectures such as ResNet50 and ResNet152. To address class imbalance, targeted data augmentation techniques are applied. Experimental results show that the ResNet152-based model achieves the highest performance with approximately 84.6% accuracy. The findings demonstrate that the proposed approach provides strong discriminative capability and has significant potential for use in clinical decision support systems.