Deep learning-based cough classification using application-recorded sounds: a transfer learning approach with VGGish.

Area of Science:

• Bio-acoustics and Medical Informatics
• Artificial Intelligence in Healthcare
• Respiratory Disease Diagnostics

Background:

• Cough sounds contain vital biometric information for assessing respiratory diseases.
• Early diagnosis of respiratory diseases is crucial due to high complication and mortality rates.
• A deep learning approach is proposed to enhance early diagnosis and improve patient outcomes.

Purpose of the Study:

• To develop and evaluate a deep learning model for early diagnosis of respiratory diseases using cough sounds.
• To assess the influence of clinical expertise and diagnostic input on model generalization.
• To provide a tool for both clinicians and individuals with limited healthcare access.

Main Methods:

• A deep learning framework using VGGish for transfer learning, incorporating detection and classification networks.
• Cough events were identified by a detection model and classified as normal or abnormal by a second model.
• Models were trained on smartphone-recorded cough sounds, meticulously labeled by multiple medical experts.

Main Results:

• The cough detection model achieved high accuracy (0.9883).
• The cough classification model demonstrated strong performance across three datasets (0.8417-0.8662).
• Grad-CAM was used for feature visualization, and performance was validated using AUROC and AUPRC.

Conclusions:

• The proposed cough classification model can support individuals with limited healthcare access and less experienced medical professionals.
• This deep learning approach, utilizing smartphone-recorded coughs, facilitates earlier detection and management of respiratory conditions.