Heart function grading evaluation based on heart sounds and convolutional neural networks

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Accurate cardiac function assessment is vital for diagnosing and managing heart diseases.
• Existing methods for assessing cardiac function have limitations in adaptability and application.
• Heart sounds (HS) offer a non-invasive window into cardiac function changes.

Purpose of the Study:

• To develop and validate a novel method for cardiac function classification using heart sound signals.
• To leverage a specialized pruning convolutional neural network (CNN) for automated feature extraction and classification.
• To demonstrate the superiority of the proposed HS analysis method compared to established deep learning models.

Main Methods:

• Heart sound signals were preprocessed using adaptive wavelet denoising and a hidden semi-Markov model for segmentation.
• Continuous wavelet transform (CWT) converted denoised HS signals into spectra for CNN input.
• A custom-designed pruning CNN was developed and compared against AlexNet, Resnet50, Xception, GhostNet, and EfficientNet.

Main Results:

• The proposed pruning CNN method achieved a classification accuracy of 94.34% for cardiac function.
• The method demonstrated superior performance compared to other benchmark deep learning architectures.
• Signal preprocessing steps effectively enhanced the quality and usability of heart sound data.

Conclusions:

• Heart sound analysis, combined with advanced AI techniques like CNNs, provides an effective and non-invasive means for cardiac function classification.
• The developed pruning CNN model shows significant potential for clinical application in cardiology.
• This research highlights promising avenues for utilizing HS analysis in disease diagnosis and treatment strategies.