Lung-SRAD: Spectral-Aware Regularized Audio DASS with Dual-Axis Patch-Mix Contrastive Learning for Respiratory Sound Classification

Researchers introduce Lung-SRAD, a novel respiratory sound classification system using State Space Models instead of traditional transformer architectures, achieving 64.48% accuracy on the ICBHI benchmark—a 5% improvement over the Audio Spectrogram Transformer baseline. The approach combines spectral-aware regularization with dual-axis patch-mix contrastive learning to better detect localized abnormal respiratory patterns.

This research addresses a critical limitation in current respiratory sound classification systems: transformer models with CLS-token attention mechanisms exhibit low-pass filtering behavior that diminishes sensitivity to localized abnormalities in audio spectrograms. By switching to State Space Models as the backbone architecture, the researchers identified superior preservation of mid-to-high frequency components essential for detecting subtle respiratory anomalies like crackles and wheezes that indicate lung disease.

The advancement builds on growing recognition within the AI community that different architectural paradigms excel at different tasks. While transformers dominated recent years, SSMs have emerged as competitive alternatives for sequential data processing, particularly for medical applications requiring fine-grained temporal and spectral detail. The researchers' spectral-aware layer regularization using Gaussian convolution strategically suppresses low-frequency noise while maintaining diagnostic signals.

The 5% performance improvement on a benchmark dataset has direct implications for clinical deployment. Healthcare systems implementing respiratory disease screening—from tuberculosis detection to COVID-19 assessment—could achieve higher diagnostic accuracy with reduced false negatives. This matters for telemedicine platforms, point-of-care diagnostic devices, and population-level health screening in resource-limited settings where access to pulmonologists is scarce.

The open-source code release accelerates adoption across research institutions and medical AI companies. Future work should validate performance on diverse patient populations, assess real-world deployment constraints, and explore whether these SSM-based architectures benefit other biomedical sound classification tasks like cardiac auscultation or speech pathology assessment.

• →State Space Models outperform transformer-based audio models for respiratory sound classification by better preserving high-frequency diagnostic signals
• →Spectral-aware regularization and dual-axis patch-mix contrastive learning improved ICBHI benchmark performance by 5% to 64.48%
• →Architecture choice significantly impacts sensitivity to localized abnormalities in medical audio applications
• →Open-source release enables rapid adoption in clinical and research settings for respiratory disease screening
• →SSM-based approaches may generalize to other biomedical audio classification tasks beyond respiratory diagnostics