Geometry-Aware Optimization Improves Respiratory Sound Classification
Published:Dec 27, 2025 11:39
•1 min read
•ArXiv
Analysis
This paper addresses the challenges of respiratory sound classification, specifically the limitations of existing datasets and the tendency of Transformer models to overfit. The authors propose a novel framework using Sharpness-Aware Minimization (SAM) to optimize the loss surface geometry, leading to better generalization and improved sensitivity, which is crucial for clinical applications. The use of weighted sampling to address class imbalance is also a key contribution.
Key Takeaways
- •Proposes a novel framework for respiratory sound classification using SAM-optimized Audio Spectrogram Transformers.
- •Addresses overfitting and class imbalance issues common in medical datasets.
- •Achieves state-of-the-art performance and significantly improves sensitivity.
- •Demonstrates the model learns robust, discriminative features.
Reference
“The method achieves a state-of-the-art score of 68.10% on the ICBHI 2017 dataset, outperforming existing CNN and hybrid baselines. More importantly, it reaches a sensitivity of 68.31%, a crucial improvement for reliable clinical screening.”