RepAugment: Input-Agnostic Representation-Level Augmentation for Respiratory Sound Classification.

Area of Science:

• Artificial Intelligence in Healthcare
• Biomedical Signal Processing
• Respiratory Medicine

Background:

• Artificial intelligence (AI) is increasingly used as a healthcare assistant, with pretrained models showing success in various tasks.
• Pretrained speech models, intuitively similar to human-originated lung sounds, have not been explored for respiratory sound classification.

Purpose of the Study:

• To investigate the effectiveness of pretrained speech models for classifying respiratory sounds.
• To address the characterization gap between speech and lung sounds using data augmentation.
• To introduce a novel augmentation technique suitable for waveform-pretrained models.

Main Methods:

• Exploration of pretrained speech models for respiratory sound classification.
• Identification of a characterization gap requiring data augmentation.
• Proposal of RepAugment, a representation-level augmentation technique, as an alternative to spectrogram-based SpecAugment.
• Comparison of RepAugment with SpecAugment on respiratory sound datasets.

Main Results:

• A significant characterization gap exists between speech and lung sound data.
• Data augmentation is crucial for bridging this gap.
• RepAugment, an input-agnostic technique, is effective for waveform-pretrained models.
• RepAugment outperforms SpecAugment, particularly improving accuracy for minority disease classes by up to 7.14%.

Conclusions:

• Pretrained speech models show promise for respiratory sound classification but require effective data augmentation.
• RepAugment offers a superior alternative to SpecAugment for waveform-pretrained models in this domain.
• The proposed method enhances diagnostic accuracy, especially for less common respiratory conditions.