Quality Adaptive Angular Margin Learning for Respiratory Sound Classification

Researchers present QLung, a machine learning framework that uses quality-adaptive angular margin learning to improve respiratory sound classification. The approach achieves 2.46% performance improvement on the ICBHI dataset and demonstrates superior out-of-distribution generalization on the SPRSound dataset compared to existing methods.

QLung represents a specialized advancement in audio classification for medical diagnostics, specifically targeting respiratory sound analysis. The framework addresses a fundamental challenge in machine learning: maintaining model robustness when training data varies significantly in quality and class distribution is imbalanced. By introducing a no-reference audio quality metric derived from spectral entropy and root-mean-square energy, the system dynamically adjusts learning parameters based on actual recording conditions rather than treating all data uniformly.

The research builds on established margin-learning techniques but adapts them for audio domain challenges. Angular margin learning has shown promise in face recognition and other tasks where feature discrimination matters critically. The innovation here centers on making these margins adaptive to audio quality, acknowledging that clinical recordings in real-world settings vary considerably in fidelity. The log-scaled margin approach specifically addresses severe class imbalance, a common problem in medical datasets where certain respiratory conditions are rarer than others.

For healthcare AI applications, this work demonstrates practical value through superior out-of-distribution performance. The fact that QLung generalizes better to the SPRSound dataset after training on ICBHI data suggests the approach captures more generalizable respiratory sound patterns. This generalization capability matters significantly for deployment scenarios where training and deployment environments differ. Medical AI systems that perform well only on matched datasets face substantial real-world limitations.

The availability of open-source code through the RSC-Toolkit repository enables broader adoption and validation by the research community. Future developments might extend this quality-adaptive approach to other audio classification tasks in healthcare, such as cough analysis or cardiac sound interpretation, where recording variability similarly impacts model performance.

• →Quality-adaptive angular margins improve respiratory sound classification by scaling learning parameters based on audio quality metrics
• →The framework achieves 2.46% performance gains on in-distribution data and superior out-of-distribution generalization compared to prior methods
• →Log-scaled angular margins stabilize training under severe class imbalance, a common challenge in medical datasets
• →Superior generalization to unseen datasets demonstrates practical value for real-world clinical deployment scenarios
• →Open-source implementation enables broader adoption and potential extension to other audio-based medical diagnostic tasks