On the analysis of data augmentation methods for spectral imaged based heart sound classification using convolutional...

Area of Science:

• Machine learning applications in healthcare
• Artificial intelligence in cardiology
• Digital health and medical diagnostics

Background:

• Machine learning, especially convolutional neural networks (CNNs), shows promise for improving cardiac auscultation accuracy.
• Limited patient data hinders the development of robust models for diverse heart sound variability.
• This study investigates data augmentation techniques to enhance CNN performance in heart sound classification.

Purpose of the Study:

• To evaluate the impact of various data augmentation methods on a CNN model for automated heart sound classification.
• To identify the most effective augmentation strategies for cardiac spectrogram analysis.
• To improve the accuracy and adaptability of machine learning models in cardiology.

Main Methods:

• A standard CNN model was developed for classifying heart sound recordings as normal or abnormal.
• Various data augmentation techniques were applied, including horizontal flipping, PCA color augmentation, HSV perturbations, time/frequency masking, and noise injection.
• The performance was assessed using the PR AUC metric.

Main Results:

• The baseline CNN model achieved a PR AUC of 0.763 ± 0.047.
• Horizontal flipping yielded the best performance improvement, reaching a PR AUC of 0.819 ± 0.044.
• Combining horizontal flipping with PCA and SV perturbations further enhanced model performance, while other methods like noise injection negatively impacted results.

Conclusions:

• Data augmentation can significantly improve classification accuracy by increasing dataset diversity and preventing overfitting.
• Domain-specific augmentation is critical; methods successful in other sound classification tasks may degrade cardiac sound analysis performance.
• Clinically appropriate data augmentation techniques are essential for developing reliable AI tools for cardiac auscultation.