Heart Sound Classification Network Based on Convolution and Transformer

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Electronic auscultation is crucial for diagnosing cardiovascular diseases (CVDs).
• Current heart sound classification methods often require complex signal segmentation and feature extraction.
• These limitations hinder efficient and accurate CVD diagnosis.

Purpose of the Study:

• To develop an innovative and simplified approach for heart sound classification.
• To introduce a novel Convolution and Transformer Encoder Neural Network (CTENN) for automated feature extraction.
• To improve the accuracy and efficiency of cardiovascular disease detection.

Main Methods:

• A new method named Convolution and Transformer Encoder Neural Network (CTENN) was developed.
• CTENN integrates a 1D-convolution module and a Transformer encoder for automatic feature extraction.
• The approach bypasses the need for traditional, precise signal segmentation and feature engineering.

Main Results:

• The CTENN method achieved high accuracies of 96.4%, 99.7%, and 95.7% on three different datasets.
• Demonstrated superior performance in both binary and multi-class heart sound classification tasks.
• Outperformed existing similar approaches in experimental evaluations.

Conclusions:

• The CTENN method offers a simplified and effective solution for heart sound classification.
• This advancement has the potential to significantly enhance cardiovascular disease diagnosis.
• The automated feature extraction capability of CTENN promises broader clinical applicability.