Scholarlink

<p>Accurate classification of heart sounds is critical for the early diagnosis of cardiovascular diseases and the reduction of mortality rates. Traditional diagnostic methods typically rely on expert analysis, which may introduce subjectivity and limitations. To enhance the accuracy and efficiency of heart sound diagnosis, a novel method integrating bispectral analysis with deep learning techniques is proposed. This method extracts two-dimensional high-order spectral feature maps from phonocardiography signals, which are then input into a ResCANet convolutional neural network for classification. The ResCANet model incorporates attention mechanisms, including Squeeze-and-Excitation and Efficient Channel Attention, to improve feature extraction and classification accuracy. Extensive experiments were conducted on two publicly available datasets to demonstrate the efficacy of the proposed method. For the binary classification dataset, the model achieves an accuracy of 93.27%, an F1 score of 92.95%, and a sensitivity of 94.67%. For the five-class classification dataset, the model achieves an accuracy of 99.6%, outperforming traditional methods. The source code is at this link: <a href="https://github.com/yao-May-all-go-well/code">https://github.com/yao-May-all-go-well/code</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhanced heart sound classification using bispectral features and attention-guided ResCANet

  • 期刊论文

  • 发表日期：

    2025年7月13日

查看全文
  • Mengyao Cui,
  • Zhanfang Zhao,
  • Tengchao Yin

摘要

Accurate classification of heart sounds is critical for the early diagnosis of cardiovascular diseases and the reduction of mortality rates. Traditional diagnostic methods typically rely on expert analysis, which may introduce subjectivity and limitations. To enhance the accuracy and efficiency of heart sound diagnosis, a novel method integrating bispectral analysis with deep learning techniques is proposed. This method extracts two-dimensional high-order spectral feature maps from phonocardiography signals, which are then input into a ResCANet convolutional neural network for classification. The ResCANet model incorporates attention mechanisms, including Squeeze-and-Excitation and Efficient Channel Attention, to improve feature extraction and classification accuracy. Extensive experiments were conducted on two publicly available datasets to demonstrate the efficacy of the proposed method. For the binary classification dataset, the model achieves an accuracy of 93.27%, an F1 score of 92.95%, and a sensitivity of 94.67%. For the five-class classification dataset, the model achieves an accuracy of 99.6%, outperforming traditional methods. The source code is at this link: https://github.com/yao-May-all-go-well/code.