Automatic pediatric congenital heart disease classification based on heart sound signal

Area of Science:

• Biomedical Engineering
• Pediatric Cardiology
• Artificial Intelligence in Medicine

Background:

• Congenital heart diseases (CHD) are the most common birth defects, necessitating early diagnosis for effective therapy.
• Pediatric CHD diagnosis via intelligent auscultation is challenging due to poor heart sound quality caused by factors like crying and breathing.
• Existing studies on pediatric CHD intelligent auscultation are limited.

Purpose of the Study:

• To develop a novel intelligent auscultation method for accurate pediatric CHD diagnosis using electronic stethoscopes.
• To establish a robust pediatric CHD heart sound database for research and development.
• To improve the classification accuracy of pediatric CHD through advanced signal processing and machine learning techniques.

Main Methods:

• Development of a pediatric CHD heart sound database comprising 941 phonocardiogram (PCG) signals.
• Implementation of a segment-based heart sound segmentation algorithm to isolate cardiac cycles and minimize noise.
• Classification of CHD using a majority voting classifier integrating Random Forest and Adaboost algorithms with 84 time and frequency domain features.

Main Results:

• The proposed method achieved high classification performance for pediatric CHD.
• Key performance metrics included accuracy (0.953), sensitivity (0.946), specificity (0.961), and f1-score (0.953).
• The segment-based approach effectively reduced the impact of local noise on global heart sound analysis.

Conclusions:

• The developed intelligent auscultation method demonstrates competitive performance for pediatric CHD classification.
• This approach offers a promising tool for the early and accurate diagnosis of congenital heart diseases in children.
• The study highlights the potential of AI-driven auscultation in pediatric cardiology.