Advanced Heart Sound Classification for Valvular Heart Disease Detection Using a Hybrid CNN-LSTM-GRU Model

Cardiovascular diseases (CVDs), particularly valvular heart disease (VHD), pose significant health risks globally, underscoring the need for advanced diagnostic techniques. This study presents a hybrid Convolutional Neural Network (CNN) - Long Short-Term Memory (LSTM) - Gated Recurrent Unit (GRU) model for the classification of heart sounds, leveraging phonocardiography (PCG) to enhance early detection of VHD. By incorporating Mel-frequency cepstral coefficients (MFCCs) for feature extraction and employing a robust preprocessing pipeline, our model achieves high accuracy in distinguishing between five different heart sounds classes. The Yaseen dataset, demonstrating exceptional Train Accuracy (0.9943), Validation Accuracy (0.9667), and Test Accuracy (0.9800). The model's effectiveness is further validated through comprehensive classification metrics, including precision, recall, and F1-score for multiple heart conditions. This research not only contributes to the automation of heart sound analysis but also offers promising insights for improving clinical decision-making in the management of valvular heart disease.