Atakanisik/ICBHI-AST-SAM

Geometry-Aware Optimization for Respiratory Sound Classification (AST + SAM)

This repository contains the model weights for the paper "Geometry-Aware Optimization for Respiratory Sound Classification: Enhancing Sensitivity with SAM-Optimized Audio Spectrogram Transformers".

Description

Respiratory sound classification is often hindered by limited datasets and class imbalance. This framework enhances the Audio Spectrogram Transformer (AST) by using Sharpness-Aware Minimization (SAM). Instead of merely minimizing training loss, this approach optimizes the geometry of the loss surface, guiding the model toward flatter minima that generalize better to unseen patients. The method specifically aims to improve sensitivity, a crucial metric for reliable clinical screening.

Key Results (ICBHI 2017 Official Split)

Metric	Score
Sensitivity (Se)	68.31%
Specificity (Sp)	67.89%
ICBHI Score	68.10%

Links

• GitHub Repository: Atakanisik/ICBHI-AST-SAM
• Paper: arXiv:2512.22564

Citation

If you use this model or code in your research, please cite:

@article{isik2025geometry,
  title={Geometry-Aware Optimization for Respiratory Sound Classification: Enhancing Sensitivity with SAM-Optimized Audio Spectrogram Transformers},
  author={I\c{s}{\i}k, Atakan and Vulga I\c{s}{\i}k, Selin and I\c{s}{\i}k, Ahmet Feridun and Taylan, Mah\c{s}uk},
  journal={arXiv preprint arXiv:2512.22564},
  year={2025}
}