Temporal convolutional neural network-based feature extraction and asynchronous channel information fusion method for...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Auscultation is key for pediatric cardiac abnormality screening, especially in resource-limited areas.
• Phonocardiogram (PCG) signal variability and distinguishing murmurs pose clinical challenges.

Purpose of the Study:

• To develop a robust framework for detecting pediatric cardiac abnormalities using PCG signals.
• To address signal variability and improve diagnostic accuracy.

Main Methods:

• A temporal convolutional network (TCN)-based framework for feature extraction and information fusion.
• Integration of asynchronously acquired PCG recordings at the patient level.
• Application of transfer learning to overcome domain adaptation challenges.

Main Results:

• The framework demonstrated robust performance on public PCG datasets.
• Transfer learning significantly improved external validation performance (AUC 0.931±0.027, AUPRC 0.867±0.064).
• Combining datasets enhanced model generalizability for detecting cardiac abnormalities.

Conclusions:

• The proposed framework offers a flexible and accurate solution for pediatric cardiac abnormality detection.
• It accommodates multi-channel, variable-length PCG recordings, suitable for low-resource settings.
• Open-source code is available for wider adoption and research.