Quality Adaptive Angular Margin Learning for Respiratory Sound Classification

Technical Summary: Quality Adaptive Angular Margin Learning for Respiratory Sound Classification

Problem Statement
Respiratory Sound Classification (RSC) faces two primary challenges that hinder model generalization to real-world clinical data:

1. Variable Recording Quality: Public datasets like ICBHI contain recordings with varying audio quality. Low-quality samples often introduce noise that mimics pathology, leading models to learn spurious correlations rather than true disease features.
2. Severe Class Imbalance and Fine-Grained Overlap: RSC datasets suffer from a significant imbalance between normal and abnormal classes. Furthermore, abnormal categories (e.g., crackles and wheezes) often occur simultaneously, creating fine-grained classification challenges where acoustic boundaries are subtle and overlapping.

Existing methods, including data augmentation and metadata integration, have struggled to fully mitigate these issues, resulting in models that lack robust decision boundaries and perform poorly on out-of-distribution (OOD) data.

Methodology: QLung
The authors propose QLung, a framework designed to enforce intra-class compactness and inter-class separability through a Quality-Adaptive Angular Margin Learning scheme. The methodology consists of four core components:

1. Angular Classifier:
   To ensure that decision boundaries rely purely on angular similarity rather than feature magnitude (which is heavily influenced by recording loudness and quality), the framework employs an angular classifier. This module normalizes both feature vectors and class weight vectors to the unit hypersphere. Decisions are made based on the cosine similarity between the normalized feature and the class weight, scaled by a fixed logit scale ($s_a$).

2. Audio Quality Margin ($m_q$):
   The framework introduces a no-reference Audio Quality Score (AQS) derived from spectral entropy and root-mean-square (RMS) energy.

   • High entropy indicates noise-like signals; low RMS indicates weak recordings.
   • The AQS is calculated as a clipped linear combination of normalized entropy and RMS.
   • This score is used to dynamically scale the angular margin: high-quality inputs receive larger margins to encourage strong separation, while low-quality inputs receive smaller margins to prevent overfitting to noise.

3. Log-Scaled Class Imbalance Margin ($m_c$):
   To address severe class imbalance without causing training instability (which can occur with standard inverse-frequency scaling), the authors propose a log-scaled margin.

   • The margin is defined as the logarithm of the inverse class frequency.
   • This logarithmic mapping ensures smooth margin growth, emphasizing minority classes without excessively penalizing majority classes, thereby stabilizing the trade-off between head and tail classes.

4. Dual Factor Angular Margin (DFAM) and Unified Loss:
   The final margin ($m_d$) is a weighted average of the quality margin and the imbalance margin ($m_d = \gamma m_q + (1-\gamma)m_c$).

   • This composite margin is applied as an additive penalty to the target angle in the softmax loss function.
   • The total training objective combines the standard Cross-Entropy (CE) loss with the DFAM regularization term, weighted by a hyperparameter $\lambda$.

Key Contributions

• First Application of Angular Margins to RSC: The paper claims to be the first to apply angular margin-based frameworks to respiratory sound classification, specifically targeting the dual problems of overlapping acoustic events and class imbalance.
• Unified Dual-Factor Formulation: The introduction of a margin that adaptively scales based on both recording quality (via AQS) and class frequency (via log-scaling).
• Angular Classifier Design: The implementation of a classifier that constrains decisions to the unit hypersphere, ensuring margin penalties are applied consistently regardless of signal amplitude.
• Performance Improvements: The framework achieves state-of-the-art results on standard benchmarks without relying on complex data augmentation strategies like Patch-Mix or additional LoRA training.

Experimental Results
The authors evaluated QLung on the ICBHI 2017 dataset (in-distribution) and the SPRSound dataset (out-of-distribution).

• In-Distribution (ICBHI):
  • QLung applied to an AST backbone improved the official Score (arithmetic mean of Specificity and Sensitivity) from 59.55% (baseline AST fine-tuning) to 62.01%, a 2.46% absolute improvement.
  • When applied to the Audio-CLAP backbone, the Score improved from 62.56% to 63.39%, surpassing the previous state-of-the-art (BTS, 63.54% is noted as the previous best in the table, though the text highlights QLung's competitiveness and improvement over baselines).
• Out-of-Distribution (SPRSound):
  • QLung demonstrated superior generalization compared to prior state-of-the-art methods.
  • On the SPRSound dataset, QLung (on Audio-CLAP) achieved a Score of 59.80%, significantly outperforming previous methods which suffered from performance degradation (e.g., BTS scored 53.42%, Audio-CLAP scored 56.29%).
• Ablation Studies:
  • Incremental addition of components (Fixed angular margin, Audio quality margin, Class imbalance margin, Angular classifier) confirmed that each component contributes to the final performance gain.
  • Qualitative analysis (t-SNE and confusion matrices) showed that QLung produces more compact and well-separated clusters, particularly for the "both" (co-occurring crackle and wheeze) class, and reduces false alarms for normal cases.

Significance
The paper asserts that QLung provides a robust solution for real-world clinical scenarios where data is noisy, heterogeneous, and imbalanced. By adaptively scaling margins based on recording quality and class frequency, the framework enhances the model's ability to form clear decision boundaries. The significant improvement in out-of-distribution performance on the SPRSound dataset suggests that QLung effectively mitigates the risks of overfitting to low-quality data and improves generalization to unseen clinical environments.