Deep Unsupervised Representation Learning for Abnormal Heart Sound Classification.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Heart disease is a global health concern.
• Early detection of heart abnormalities is crucial for patient outcomes.
• Phonocardiogram (heart sound) analysis is a key diagnostic tool.

Purpose of the Study:

• To compare conventional and deep learning methods for classifying heart sound abnormalities.
• To evaluate deep feature representations from sequence-to-sequence autoencoders.
• To assess performance on the Heart Sounds Shenzhen corpus.

Main Methods:

• Utilized the auDeep toolkit for sequence-to-sequence autoencoders.
• Extracted deep unsupervised features from phonocardiogram recordings.
• Performed a three-way classification: normal, mild, and moderate/severe abnormalities.
• Fused features from different deep learning models.

Main Results:

• A combination of deep unsupervised features achieved the highest performance.
• The fused feature approach yielded an unweighted average recall of 47.9% on the test set.
• Deep learning models demonstrated suitability for automated heart sound classification.

Conclusions:

• Deep unsupervised features are effective for classifying heart sound abnormalities.
• Automated analysis of phonocardiograms using deep learning can aid in patient care.
• Further research can improve accuracy for robust clinical application.