Detection of systolic ejection click using time growing neural network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Accurate classification of heart sounds is crucial for pediatric cardiac diagnostics.
• Short-duration heart sounds present unique challenges for automated analysis.
• Existing neural network models show limitations in classifying subtle cardiac abnormalities.

Purpose of the Study:

• To introduce a novel neural network, the time growing neural network (TGNN), for classifying short-duration heart sounds.
• To evaluate the performance of TGNN against established models like TDNN and MLP.
• To assess TGNN's robustness against noise in cardiac sound classification.

Main Methods:

• Input features derived from spectral power in adjacent frequency bands within time windows of growing length.
• Utilized a dataset of 614 normal and abnormal cardiac cycles from children.
• Compared TGNN performance against Time Delay Neural Network (TDNN) and Multi-Layer Perceptron (MLP).

Main Results:

• TGNN achieved a classification rate and sensitivity of 97.0% and 98.1%, respectively.
• TGNN significantly outperformed TDNN (85.1%/76.4%) and MLP (92.7%/85.7%) on the test dataset.
• TGNN demonstrated superior performance and enhanced immunity to noise compared to TDNN and MLP.

Conclusions:

• The time growing neural network (TGNN) is a highly effective model for classifying short-duration pediatric heart sounds.
• TGNN offers improved accuracy and noise resistance over traditional neural networks for cardiac auscultation analysis.
• This novel approach holds promise for advancing automated diagnosis of pediatric heart conditions like systolic ejection click (SEC).