PPEA: Post-Position Encoding Attention for Imbalanced Lung Sound Classification.

Area of Science:

• Medical Technology
• Artificial Intelligence
• Pulmonology

Background:

• Early detection and continuous monitoring of respiratory diseases are crucial clinical challenges.
• Auscultation is a primary diagnostic tool, but interpretation requires expertise and varies between observers.

Purpose of the Study:

• To present PPEA, a novel deep-learning framework for automated respiratory sound classification.
• To address the challenge of imbalanced clinical data in respiratory sound analysis.

Main Methods:

• Developed PPEA, a deep-learning framework utilizing a post-position encoding attention mechanism.
• Implemented a layer-wise feature fusion strategy to handle imbalanced data.
• Evaluated performance on the ICBHI dataset for classifying six respiratory conditions.

Main Results:

• Achieved a specificity of 0.9933 and a sensitivity of 0.7863 across six respiratory conditions.
• Demonstrated superior performance compared to existing methods.
• Showcased effectiveness in few-shot learning scenarios for respiratory sound classification.

Conclusions:

• PPEA offers a robust solution for automated respiratory sound classification.
• The framework shows significant potential for supporting clinical decision-making in respiratory care.
• Advanced deep learning techniques can overcome challenges posed by imbalanced clinical data.