A Robust Deep Learning Approach for COPD Automated Detection.

Area of Science:

• Medical acoustics
• Artificial intelligence in healthcare
• Respiratory medicine

Background:

• Chronic Obstructive Pulmonary Disease (COPD) is a major respiratory illness requiring timely diagnosis.
• Current diagnostic methods for respiratory conditions can be limited.
• Respiratory sound analysis offers a non-invasive diagnostic avenue.

Purpose of the Study:

• To develop and validate a novel deep learning framework for COPD detection using respiratory sounds.
• To adapt the ECAPA-TDNN architecture for medical acoustic analysis.
• To enhance the interpretability and clinical applicability of AI-driven respiratory diagnostics.

Main Methods:

• Utilized the ECAPA-TDNN architecture, originally for speaker verification, for respiratory sound classification.
• Applied rigorous signal preprocessing, log-Mel spectrograms, and data augmentation.
• Incorporated Attentive Statistics Pooling for feature summarization and Grad-CAM for explainability.

Main Results:

• Achieved a mean validation accuracy of 96.8% across a five-fold cross-validation.
• Demonstrated high F1-scores and recall rates, indicating robust performance.
• Showcased superior diagnostic precision and robustness compared to conventional methods.

Conclusions:

• The ECAPA-TDNN framework represents a significant advancement in acoustic-based COPD detection.
• This approach offers a highly accurate, interpretable, and potentially clinically applicable tool for respiratory disease screening.
• Establishes a new benchmark for AI-powered analysis of pathological breathing sounds.