Automated detection of pediatric congenital heart disease from phonocardiograms using deep and handcrafted feature...

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence

Background:

• Congenital heart disease (CHD) is a prevalent birth defect, affecting ~1% of newborns globally.
• Current gold-standard diagnosis via echocardiography is expensive and inaccessible in low-resource settings.
• Delayed diagnosis of CHD results from a shortage of skilled experts and significant inter-clinician variability in interpreting pathological heart sounds.

Purpose of the Study:

• To develop an accessible, automated diagnostic method for early detection of CHD using digital stethoscopes.
• To integrate deep and handcrafted features for enhanced diagnostic accuracy in Phonocardiography (PCG) analysis.
• To create a cost-effective screening tool for CHD in resource-limited environments.

Main Methods:

• Collected PCG recordings from 751 pediatric subjects (1 month-16 years) in Bangladesh across four auscultation sites.
• Utilized a deep feature fusion approach, combining deep and handcrafted features for automated CHD detection.
• Trained and validated the model using a patient-wise split (70% training, 20% validation, 10% testing).

Main Results:

• The proposed model achieved 92% accuracy, 91% sensitivity, and 91% specificity.
• The model demonstrated a high Area Under the ROC Curve (AUROC) of 96% and an F1-score of 92%.
• The method shows significant promise for real-time remote CHD detection.

Conclusions:

• The digital stethoscope combined with deep feature fusion offers a viable, cost-effective solution for early CHD detection.
• This approach can significantly improve diagnostic accessibility in low-resource settings, reducing diagnostic delays.
• The model's high performance metrics indicate its potential as a reliable screening tool for pediatric CHD.