ATF-MGIAM: Medically-guided interpretable attention mapping for robust pertussis cough sound recognition
Siheng Zhang1,2, Yan Xu1, Ying Xiao2
1Zhejiang Provincial Center for Disease Control and Prevention, Hangzhou, Zhejiang, China.
Plos One
|May 4, 2026
Summary
This study introduces an interpretable deep learning model for recognizing pertussis cough sounds. The novel framework achieves high accuracy and provides clear insights into diagnostic features.
Area of Science:
- Artificial Intelligence
- Medical Acoustics
- Signal Processing
Background:
- Pertussis (whooping cough) diagnosis faces challenges due to complex cough sounds.
- Existing methods for cough sound recognition lack interpretability.
Purpose of the Study:
- To develop an interpretable deep learning framework for accurate pertussis cough sound recognition.
- To enhance the understanding of diagnostic features in cough sounds.
Main Methods:
- An Adaptive Time-Frequency Fusion Transformer (ATF) was used to extract multi-scale features.
- A Medically-Guided Interpretable Attention Mapping (MGIAM) module was employed for explicit interpretability.
- Experiments were conducted on three public pertussis cough sound datasets with rigorous data splitting.
Main Results:
- The proposed framework achieved high AUC scores (0.994, 0.984, 0.996), outperforming baselines by ~2%.
- Ablation studies confirmed the effectiveness of the ATF and MGIAM modules.
- The model demonstrated strong noise robustness with <3% performance degradation.
Conclusions:
- The framework offers a unified approach to high accuracy and interpretability in pertussis sound recognition.
- It provides a reusable paradigm for medical acoustic analysis.
- The model shows potential for clinical applications due to its reliability and generalization.
