An explainable and accurate transformer-based deep learning model for wheeze classification utilizing real-world...

Area of Science:

• Artificial Intelligence in Medicine
• Pediatric Pulmonology
• Biomedical Signal Processing

Background:

• Auscultation with stethoscopes is vital for diagnosing respiratory diseases but relies on subjective interpretation.
• Existing deep learning models, primarily CNN-based, struggle with complex respiratory sound patterns.
• There's a need for advanced AI to objectively interpret lung sounds for improved pediatric respiratory diagnostics.

Purpose of the Study:

• To develop and validate an AI deep learning model for accurate interpretation of pediatric respiratory sounds.
• To apply the Audio Spectrogram Transformer (AST) model to clinical data for wheezing detection.
• To compare the AST model's performance against CNN-based models and assess its clinical reliability.

Main Methods:

• A prospective study involving children from two South Korean university hospitals (2019-2020).
• Recording of pediatric breath sounds by a pulmonologist and construction of a double-blind verified dataset.
• Development of a deep learning model using pre-trained AST model weights on a dataset of 194 wheezes and 531 other respiratory sounds.

Main Results:

• The AST-based model achieved 91.1% accuracy, 86.6% AUC, 88.2% precision, 76.9% recall, and 82.2% F1-score.
• The model demonstrated high accuracy in detecting wheezing in pediatric respiratory sounds.
• Score-Class Activation Mapping (Score-CAM) visualization confirmed the model's reliable decision-making process.

Conclusions:

• The Transformer-based AST model shows significant promise for accurate wheezing detection in children.
• This AI model offers a reliable and objective tool to support clinical diagnosis of pediatric respiratory conditions.
• The developed AI system is expected to enhance the accuracy of diagnosing pediatric respiratory diseases in clinical settings.