Automated heart sound classification system from unsegmented phonocardiogram (PCG) using deep neural network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Low patient-to-doctor ratios in low- and middle-income countries necessitate automated diagnostic tools.
• Phonocardiogram (PCG) analysis for cardiovascular disease detection is crucial but often requires expert interpretation.
• Automating PCG analysis can improve accessibility and efficiency of cardiac screening.

Purpose of the Study:

• To propose and evaluate deep neural network (DNN) architectures for automated classification of unsegmented PCG signals.
• To eliminate the need for manual feature engineering and segmentation in PCG analysis.
• To develop a real-time screening system for heart sound classification.

Main Methods:

• Down-sampling PCG signals to 500 Hz and segmenting into 6-second epochs.
• Applying a Savitzky-Golay filter for noise reduction and data smoothing.
• Utilizing one-dimensional Convolutional Neural Networks (1D-CNN) and Feed-forward Neural Networks (F-NN) for classification.
• Training and validating DNN models on 1081 PCG records.

Main Results:

• The Feed-forward Neural Network (F-NN) model achieved an overall accuracy of 85.65%.
• The F-NN model demonstrated a sensitivity of 86.73% and specificity of 84.75%.
• Balanced accuracy was 85.74%, with an Area Under the Curve (AUC) of 0.857 from ROC analysis.
• DNN models showed comparable performance to existing methods without feature engineering.

Conclusions:

• Deep neural networks, particularly F-NN, are effective for automated, real-time heart sound classification from PCG signals.
• The proposed DNN approach simplifies PCG analysis by removing the need for manual feature extraction and segmentation.
• This technology holds potential for improving cardiovascular screening accessibility, especially in resource-limited settings.