Scholarlink

Heart murmurs are abnormal sounds produced by turbulent blood flow in or near the heart, often signaling structural or functional cardiac abnormalities such as valve stenosis or regurgitation. Early detection is critical for the diagnosis and management of cardiovascular diseases, yet access to expert cardiac evaluation is limited in many healthcare environments. This study addresses this challenge by leveraging deep learning techniques for automated murmur detection using the CirCor DigiScope Dataset, which provides phonocardiographic recordings from multiple patients. Our core contribution is a dual-purpose convolutional autoencoder (AE), which performs both data purification and feature engineering. The AE is trained on normal heart sounds to establish a baseline for reconstruction error. This error is then used to filter out murmur-labeled segments that are acoustically indistinguishable from normal sounds and serve as a quantitative anomaly feature for the remaining samples. Finally, these purified spectrograms, augmented with their AE-derived error features, are classified by a Multimodal-based network.

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

Heart Murmur Detection in Phonocardiogram Signals Using an Autoencoder and Multimodal-Based Classification Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Héctor U. Chávez-Loya,
  • Said A. Delgado-Portillo,
  • Natalia G. Sámano-Lira,
  • Javier Camarillo-Cisneros,
  • Abimael Guzmán-Pando

摘要

Heart murmurs are abnormal sounds produced by turbulent blood flow in or near the heart, often signaling structural or functional cardiac abnormalities such as valve stenosis or regurgitation. Early detection is critical for the diagnosis and management of cardiovascular diseases, yet access to expert cardiac evaluation is limited in many healthcare environments. This study addresses this challenge by leveraging deep learning techniques for automated murmur detection using the CirCor DigiScope Dataset, which provides phonocardiographic recordings from multiple patients. Our core contribution is a dual-purpose convolutional autoencoder (AE), which performs both data purification and feature engineering. The AE is trained on normal heart sounds to establish a baseline for reconstruction error. This error is then used to filter out murmur-labeled segments that are acoustically indistinguishable from normal sounds and serve as a quantitative anomaly feature for the remaining samples. Finally, these purified spectrograms, augmented with their AE-derived error features, are classified by a Multimodal-based network.