Multi-Classification of Cardiovascular Diseases Based on Heart Sound Signals Using Audio Spectrogram Features with Pure Attention Transformer

Background and objectives: In computer-aided medical diagnosis or prognosis, the automatic classification of cardiovascular diseases (CVDs) based on heart sound (HS) signals is of great importance since the heart sound signal contains a wealth of information that can reflect the cardiovascular status. Traditional binary classification algorithms (normal and abnormal) currently cannot comprehensively assess CVDs based on analyzing various heart sounds. The differences between heart sound signals are relatively subtle, but the reflected heart conditions differ significantly. Consequently, the multiclassification of CVDs from heart sound signals is of utmost importance from a clinical viewpoint.Methods: For the multi-classification of heart sound signals, we propose a creative non-convolutional neural networks (CNN) pure-attention transformer model. It has achieved remarkable results from four abnormal HS signals and the typical type.Results: According to a fivefold cross-validation strategy, the proposed method achieves a mean classification accuracy of 99% and a mean average precision of 0.99. Further, the classification accuracy for Aortic Stenosis (AS), Mitral Regurgitation (MR), Mitral Stenosis (MS), Mitral Valve Prolapse (MVP), and standard heart sound signals (N) is 99.39%, 99.53%, 99.42%, 99.48%, and 99.67%, respectively.Conclusion: The results indicate that the framework can precisely classify five classes of heart sound signals. Our method provides an efficient tool for CVD multi-classification based on heart sound signals in clinical settings.