Exploring Pre-trained General-purpose Audio Representations for Heart Murmur Detection

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Automating cardiac auscultation is crucial for reducing reliance on skilled clinicians.
• Deep learning for heart sound analysis requires large datasets, which are often limited.
• General-purpose audio models offer pre-trained representations beneficial for specialized tasks.

Purpose of the Study:

• To investigate the efficacy of general-purpose audio representations for transfer learning in heart murmur detection.
• To evaluate the performance of the Masked Modeling Duo (M2D) model in this context.

Main Methods:

• Utilized transfer learning with general-purpose audio representations, specifically the Masked Modeling Duo (M2D) model.
• Conducted experiments on the CirCor DigiScope heart sound dataset.
• Explored ensembling M2D with other models to further enhance performance.

Main Results:

• The Masked Modeling Duo (M2D) model achieved a weighted accuracy of 0.832 and an unweighted average recall of 0.713.
• Ensembling M2D with other models led to further performance improvements.
• Demonstrated the effectiveness of general-purpose audio representations in processing heart sounds.

Conclusions:

• General-purpose audio representations, particularly M2D, are effective for transfer learning in automated heart murmur detection.
• This approach shows promise for improving cardiac auscultation, especially with limited specialized data.
• The findings pave the way for broader applications of AI in cardiovascular diagnostics.