A fusion framework based on multi-domain features and deep learning features of phonocardiogram for coronary artery...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Phonocardiogram (PCG) signals reflect cardiac mechanical activity.
• Coronary Artery Disease (CAD) can cause subtle heart murmurs, often undetectable by human auscultation.
• Accurate and noninvasive CAD detection remains a clinical challenge.

Purpose of the Study:

• To propose a novel feature fusion framework for enhanced Coronary Artery Disease (CAD) diagnosis using Phonocardiogram (PCG) signals.
• To develop a robust method for identifying weak heart murmurs associated with CAD.
• To evaluate the performance of the proposed framework as a noninvasive screening tool.

Main Methods:

• A dataset of PCG signals from 175 subjects was utilized.
• 110 multi-domain features were extracted, reduced, and selected.
• Mel-frequency cepstral coefficients (MFCC) images were processed using a convolutional neural network (CNN) for deep feature learning.
• A feature fusion approach combined selected traditional features with CNN-derived deep learning features.
• A multilayer perceptron (MLP) was employed for final classification.

Main Results:

• The proposed feature fusion framework demonstrated superior classification performance compared to using multi-domain features or deep learning features independently.
• Achieved an accuracy of 90.43%, sensitivity of 93.67%, and specificity of 83.36%.
• Performance comparison indicated the method's potential as a noninvasive screening tool for CAD.

Conclusions:

• The novel feature fusion framework significantly improves the accuracy of CAD detection from PCG signals.
• The method successfully identifies subtle cardiac murmurs indicative of CAD, surpassing traditional auscultation limitations.
• The proposed approach shows promise as an effective, noninvasive tool for general medical screening of Coronary Artery Disease.