Novel audio characteristic-dependent feature extraction and data augmentation methods for cough-based respiratory...

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Signal Processing

Background:

• Respiratory diseases pose a global health challenge, necessitating accurate and early diagnosis.
• Cough sounds contain valuable information for identifying different respiratory pathologies.
• Existing AI approaches for cough analysis often overlook specific characteristics of cough signals.

Purpose of the Study:

• To design an AI system for classifying respiratory diseases based on cough sounds.
• To develop novel audio feature extraction and data augmentation methods tailored for cough signals.
• To enhance the accuracy and robustness of AI-driven respiratory disease diagnosis.

Main Methods:

• Proposed maximum overlapping mel-spectrogram to capture rapid transitions in cough sounds.
• Implemented diverse data augmentation strategies, including a novel self-energy-based method, to address limited labeled data.
• Utilized test-time augmentation to fuse results from original and augmented audio for improved classification.

Main Results:

• Achieved an average absolute performance improvement of 3.33% in macro Area Under the Receiver Operating Characteristic (macro AUC).
• Demonstrated an average absolute performance improvement of 3.10% in Unweighted Average Recall (UAR).
• Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirmed feature contributions to classification decisions.

Conclusions:

• The proposed audio characteristic-dependent methods offer a significant advancement in AI-based respiratory disease classification from cough sounds.
• These techniques enhance diagnostic accuracy, potentially leading to earlier and more effective patient management.
• The study highlights the importance of signal-specific feature engineering and augmentation in medical AI applications.