Heart Sound Classification based on Residual Shrinkage Networks.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Cardiovascular diseases (CVDs) are a leading cause of mortality worldwide.
• Early detection of CVDs is crucial for effective treatment and improved patient outcomes.
• Computer Audition (CA) technology offers a non-invasive approach for heart sound analysis.

Purpose of the Study:

• To develop and evaluate a deep learning model for accurate heart sound classification.
• To enhance the early detection of cardiovascular diseases using advanced signal processing and machine learning.
• To investigate the effectiveness of different time-frequency features for heart sound analysis.

Main Methods:

• A deep Convolutional Neural Network (CNN) model was proposed, integrating Residual Network (ResNet) and Long Short-Term Memory (LSTM) architectures.
• The model was trained and evaluated on the PhysioNet/CinC Challenges 2016 datasets.
• Four distinct time-frequency features were extracted: Filterbank (Fbank), Mel-Frequency Spectral Coefficients (MFSCs), and Mel-Frequency Cepstral Coefficients (MFCCs).

Main Results:

• The proposed CNN model demonstrated strong performance in heart sound classification.
• Mel-Frequency Spectral Coefficients (MFSCs) emerged as the most effective feature for the model.
• The model achieved an F1 score of 84.3%, accuracy of 84.4%, sensitivity of 84.3%, and specificity of 85.6% on the test set.
• An accuracy improvement of 4.9% was observed compared to the classical ResNet model.

Conclusions:

• The proposed hybrid ResNet-LSTM CNN model is effective for heart sound classification.
• MFSCs are a promising feature for improving the accuracy of deep learning-based heart sound analysis.
• This approach holds potential for non-invasive, early detection of cardiovascular diseases.