Detecting Respiratory Pathologies Using Convolutional Neural Networks and Variational Autoencoders for Unbalancing...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Respiratory Medicine

Background:

• Respiratory sound analysis is crucial for diagnosing lung diseases.
• The International Conference on Biomedical and Health Informatics (ICBHI) Benchmark dataset presents an imbalanced class distribution, with over 88% of samples belonging to the 'Chronic' disease category.
• Accurate classification of respiratory sounds is essential for effective patient management.

Purpose of the Study:

• To detect respiratory pathologies using respiratory sounds.
• To address the class imbalance in the ICBHI dataset using data augmentation techniques.
• To classify respiratory sounds into healthy, chronic, and non-chronic disease categories, and further distinguish between specific pathologies.

Main Methods:

• Preprocessing of respiratory sound data.
• Application of Variational Convolutional Autoencoder (VAE) and oversampling techniques to handle data imbalance.
• Classification using a Convolutional Neural Network (CNN).
• Performing both three-label (healthy, chronic, non-chronic) and six-class (including URTI, COPD, Bronchiectasis, Pneumonia, Bronchiolitis) classifications.

Main Results:

• Achieved an F-Score of up to 0.993 for the three-label classification.
• Achieved an F-Score of up to 0.990 for the more challenging six-class classification.
• Demonstrated the effectiveness of VAE and CNN in classifying respiratory sounds from imbalanced datasets.

Conclusions:

• The proposed method, utilizing VAE for data augmentation and CNN for classification, is highly effective for detecting respiratory pathologies from respiratory sounds.
• The approach successfully addresses data imbalance issues in respiratory sound datasets.
• High classification accuracy in both broad and specific pathology detection highlights the potential for clinical application.