Audio for Audio is Better? An Investigation on Transfer Learning Models for Heart Sound Classification.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Cardiovascular disease is a leading cause of mortality worldwide.
• Non-invasive heart sound classification is crucial for health monitoring.
• Current methods for extracting acoustic features are costly and time-consuming.

Purpose of the Study:

• To develop an efficient transfer learning model for heart sound classification.
• To address limitations of image-based pre-trained models in audio analysis.
• To leverage large-scale audio data for improved heart sound representation.

Main Methods:

• Proposed a novel transfer learning model pre-trained on extensive audio datasets.
• Applied the model to the PhysioNet CinC Challenge Dataset for evaluation.
• Compared performance against existing models pre-trained on image data.

Main Results:

• The proposed audio pre-trained model achieved the highest unweighted average recall of 89.7%.
• Demonstrated superior performance compared to image-based pre-trained models.
• Validated the effectiveness of audio-specific pre-training for heart sound classification.

Conclusions:

• Transfer learning with audio-specific pre-training offers a more effective approach for heart sound classification.
• The developed model provides a promising tool for non-invasive cardiovascular health monitoring.
• This method reduces the need for manual feature engineering in heart sound analysis.