An efficient and robust Phonocardiography (PCG)-based Valvular Heart Diseases (VHD) detection framework using Vision...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Valvular heart diseases (VHDs) are a major cause of global cardiovascular mortality.
• Early diagnosis of VHDs using phonocardiogram (PCG) signals is crucial for effective treatment and mortality reduction.

Purpose of the Study:

• To develop high-performance, deep learning (DL)-based frameworks for accurate VHD detection from PCG signals.
• To compare different DL architectures and feature extraction methods for VHD classification.

Main Methods:

• Proposed three DL frameworks utilizing 1D and 2D PCG signals with features like MFCC, LPCC, and D-CNNs.
• Employed nature-inspired algorithms (PSO, GA) for feature selection and a vision transformer (ViT) with self-attention for enhanced classification.
• Performed comparative analysis of various descriptors, classifiers, and feature selection algorithms.

Main Results:

• The vision transformer (ViT) model achieved superior performance, with a mean average accuracy (Acc) of 99.90% and an F1-score of 99.95%.
• This performance surpasses current state-of-the-art VHD classification models.

Conclusions:

• Developed a robust, efficient, end-to-end DL framework for PCG signal classification.
• The proposed framework facilitates the design of an automated, high-performance VHD diagnosis system.