[Classification of heart sound signals in congenital heart disease based on convolutional neural network]

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Cardiac auscultation is fundamental for congenital heart disease (CHD) diagnosis and screening.
• Accurate and robust classification of CHD heart sounds remains a challenge.

Purpose of the Study:

• To develop and evaluate a new classification algorithm for CHD based on convolutional neural networks (CNNs).
• To analyze and classify CHD heart sounds using machine learning for improved diagnostic capabilities.

Main Methods:

• Clinically collected CHD heart sound signals were preprocessed to extract Mel-Frequency Cepstral Coefficients (MFCCs).
• One-dimensional heart sound signals were transformed into two-dimensional feature samples.
• A CNN was trained using 1,000 feature samples and optimized with the Adam optimizer.

Main Results:

• The CNN model achieved a training accuracy of 0.896 and a loss value of 0.25.
• Testing on 200 samples yielded an accuracy of 0.895, sensitivity of 0.910, and specificity of 0.880.
• The proposed algorithm demonstrated improved accuracy and specificity compared to existing methods.

Conclusions:

• The developed CNN-based algorithm effectively enhances the robustness and accuracy of heart sound classification for CHD.
• The findings suggest potential for application in machine-assisted auscultation systems for improved CHD screening.