Assessing the Utility of Audio Foundation Models for Heart and Respiratory Sound Analysis.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Machine Learning

Background:

• Foundation models are crucial in AI, extending to audio analysis for medical applications.
• Off-the-shelf audio foundation models show potential as feature extractors for respiratory and heart sounds.
• Limited benchmarking hinders the understanding of their real-world effectiveness and state-of-the-art (SOTA) compatibility.

Purpose of the Study:

• To evaluate the practical effectiveness of off-the-shelf audio foundation models.
• To compare their performance against SOTA fine-tuning methods on respiratory and heart sound tasks.
• To identify factors influencing model performance, such as data quality.

Main Methods:

• Comparative analysis of audio foundation models against SOTA fine-tuning.
• Evaluation across four distinct respiratory and heart sound diagnostic tasks.
• Assessment of model performance on both clean and noisy datasets.

Main Results:

• Foundation models achieved SOTA performance on tasks with clean data.
• Performance degraded significantly on tasks involving noisy respiratory and heart sounds.
• General-purpose audio models demonstrated superior performance compared to specialized respiratory sound models.

Conclusions:

• Audio foundation models offer practical utility in diagnosing respiratory and heart conditions.
• Data noise is a critical challenge impacting model effectiveness in clinical audio analysis.
• General-purpose audio models exhibit broad applicability, suggesting potential for wider adoption in medical diagnostics.