Paediatric respiratory sound classification by integrating local feature extraction and global context modeling

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Medical Informatics

Background:

• Automated respiratory sound classification is vital for respiratory disease diagnosis.
• Paediatric respiratory sound analysis is less developed than adult analysis.
• Paediatric respiratory sounds present unique high-frequency challenges for classification models.

Purpose of the Study:

• To develop an advanced deep learning architecture for paediatric respiratory sound classification.
• To address the scarcity of research in paediatric respiratory sound analysis.
• To improve diagnostic precision and efficiency in paediatric respiratory disease.

Main Methods:

• Proposed a novel architecture integrating local feature extraction (Mel_Grouper) with a global context model (Transformer-based Mel_Encoder).
• The Mel_Grouper module enhances local pathological representations.
• The Mel_Encoder fuses global context for improved classification.

Main Results:

• Achieved state-of-the-art performance on the SJTU Paediatric Respiratory Sound (SPRSound) dataset.
• Outperformed previous best results by significant margins across four subtasks (3.37% to 6.83%).
• Validated performance on a real-world paediatric respiratory sound dataset.

Conclusions:

• The proposed method demonstrates significant performance in paediatric respiratory sound classification.
• The architecture effectively handles the unique characteristics of paediatric respiratory sounds.
• The developed model shows promise for clinical application in paediatric respiratory diagnostics.