Transferring Cross-Corpus Knowledge: An Investigation on Data Augmentation for Heart Sound Classification

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Cardiovascular Diagnostics

Background:

• Human auscultation is a traditional, cost-effective method for diagnosing cardiovascular diseases, but requires extensive training.
• Computer audition (CA) offers automated heart sound classification using machine learning, yet faces accuracy challenges in cross-corpus testing.

Purpose of the Study:

• To address the decline in machine learning accuracy for heart sound classification on unseen datasets.
• To develop a robust Convolutional Neural Network (CNN) model for cross-corpus heart sound classification using data augmentation.

Main Methods:

• Utilized the PhysioNet CinC Challenge 2016 Dataset for cross-corpus evaluation.
• Implemented and combined novel data augmentation techniques to enhance CNN model robustness.
• Compared augmented model performance against a baseline without augmentation.

Main Results:

• The proposed data augmentation strategy significantly improved CNN performance in cross-corpus evaluations.
• Achieved an average improvement of 20.0% in sensitivity and 7.9% in specificity across six databases.
• The improvements were statistically significant (p < .05) on four of the tested databases.

Conclusions:

• Data augmentation is crucial for developing generalizable CNN models for heart sound classification.
• The developed techniques enhance the reliability of computer audition for diagnosing cardiovascular diseases across different datasets.
• This approach overcomes limitations of previous methods, paving the way for more accurate automated cardiac diagnostics.