DropConnected neural networks trained on time-frequency and inter-beat features for classifying heart sounds

Area of Science:

• Cardiology
• Biomedical Engineering
• Signal Processing

Background:

• Valvular heart diseases require timely diagnosis, often challenging in primary care or resource-limited settings.
• Echocardiography is the gold standard but lacks accessibility globally.
• Automatic heart sound analysis offers a potential solution for accessible cardiac diagnostics.

Purpose of the Study:

• To develop and evaluate an algorithm for classifying heart sounds as normal or abnormal.
• To assess the algorithm's potential for diagnosing valvular heart diseases in diverse clinical settings.

Main Methods:

• Heart sounds segmented using a hidden semi-Markov model.
• Time-frequency analysis employed continuous wavelet transform, mel-frequency cepstral coefficients, and complexity measures.
• Features extracted for murmur and arrhythmia detection, processed via principal component analysis and a neural network.

Main Results:

• The algorithm achieved 85.2% accuracy on test data in the PhysioNet/Computing in Cardiology Challenge.
• A more realistic 10-fold cross-validation on training data (excluding dataset-e) yielded 74.8% accuracy.
• Performance dropped to 58.1% on test data when dataset-e (with differing stethoscopes) was excluded from training.

Conclusions:

• The developed algorithm demonstrates potential for automatic heart sound analysis.
• Real-world performance may be lower than challenge scores due to data variations.
• Further validation is needed to ensure robust performance across different recording conditions.