Heart Murmur Classification Using a Capsule Neural Network.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Intelligent detection systems like AI-powered diagnostics have advanced heart condition diagnosis.
• Automated segmentation and classification of heart sounds are crucial for accurate AI-driven analysis.
• Current methods often rely on electrocardiography (ECG)-labeled phonocardiograms (PCGs) or Mel-scale Frequency Cepstral Coefficient (MFCC) feature extraction, which can be limiting for Convolutional Neural Networks (CNNs).

Purpose of the Study:

• To introduce a novel Capsule Neural Network (CapsNet) for improved heart sound classification.
• To address the limitations of traditional CNNs in extracting relevant features from MFCC spectrums of heart sounds.
• To enhance the prediction accuracy of heart murmur classification using advanced AI techniques.

Main Methods:

• Proposed a Capsule Neural Network (CapsNet) utilizing iterative dynamic routing for feature extraction.
• Employed CapsNet to leverage translational equivariance in MFCC spectrum features for better classification.
• Trained and validated CapsNet against CNNs using the 2016 PhysioNet heart sound database and a custom clinical dataset.

Main Results:

• CapsNet demonstrated superior performance in classifying heart sounds compared to traditional CNNs.
• Achieved validation accuracies of 90.29% and 91.67% on the test dataset.
• Successfully fine-tuned hyperparameters and tested results on a real-world clinical auscultation dataset.

Conclusions:

• CapsNet offers a feasible and effective approach for accurate heart murmur classification.
• The proposed method overcomes feature extraction challenges faced by CNNs in heart sound analysis.
• This AI advancement holds potential for improving cardiovascular diagnostics through enhanced heart sound analysis.