Physiology-Aware Temporal Attention and Frequency Band Selection for Robust Respiratory Sound Analysis

Area of Science:

• Medical technology
• Artificial intelligence in healthcare
• Bioacoustics

Background:

• Respiratory sound classification is crucial for diagnosing lung conditions.
• Existing methods often struggle with data imbalance and computational efficiency.
• There is a need for robust and real-time respiratory sound analysis tools.

Purpose of the Study:

• To develop a physiology-aware and efficient framework for respiratory sound classification.
• To improve the accuracy and robustness of automated auscultation systems.
• To enable real-time respiratory sound analysis in resource-limited environments.

Main Methods:

• A CNN backbone combined with Phase-Informed Saliency Temporal Attention (PISTA) for feature extraction and attention.
• Importance-guided Frequency Band Selection (FBS) to reduce spectral dimensionality.
• Multi-axis Group Distributionally Robust Optimization (GroupDRO) to enhance robustness against data imbalances.

Main Results:

• State-of-the-art performance on SPRSound 2022/2023 and competitive results on ICBHI 2017.
• FBS module reduced spectral dimensionality and computation by up to 50%.
• Real-time inference demonstrated feasibility on a Raspberry Pi 3, confirming edge computing capability.

Conclusions:

• The proposed framework enables accurate and efficient respiratory sound classification.
• The system supports reliable, real-time auscultation in low-resource and point-of-care settings.
• This technology has the potential to significantly improve respiratory diagnostics globally.