Scholarlink

<p>Phonocardiogram (PCG) signals are valuable non-invasive biomarkers for early cardiovascular disease (CVD) detection. Advances in digital stethoscopes and signal processing have enabled early diagnosis of abnormal heart sounds. This study presents a classification and interpretable framework that combines cross-correlation-based features with an attention-based bidirectional long short-term memory (Bi-LSTM) network for automated analysis of heart sounds. The cross-correlation technique identifies temporal relationships among heart sound segments, allowing the model to effectively distinguish between healthy and pathological cardiac patterns. The derived features are then analyzed using a Bi-LSTM network enhanced with an attention module, which emphasizes the most relevant portions of the input sequence. This architecture delivers strong performance in terms of accuracy, sensitivity, and specificity across various categories of heart sounds when tested on a standard PCG dataset. Furthermore, attention visualization maps are produced to display the regions that receive greater importance during classification, offering a clearer understanding of how the model processes information. Temporal attention distributions are also illustrated to interpret the reasoning behind the network’s predictions, identifying specific time intervals that play a key role in decision-making. These visual insights correspond closely with clinical assessments of abnormal heart sounds, thereby improving interpretability and demonstrating the model’s practical value in cardiac diagnostics.</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

Cross-correlation aided feature extraction for automated diagnosis of phonocardiogram signals using attention based Bi-LSTM network

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Moumita Sarkar,
  • Biswarup Ganguly,
  • Debangshu Dey

摘要

Phonocardiogram (PCG) signals are valuable non-invasive biomarkers for early cardiovascular disease (CVD) detection. Advances in digital stethoscopes and signal processing have enabled early diagnosis of abnormal heart sounds. This study presents a classification and interpretable framework that combines cross-correlation-based features with an attention-based bidirectional long short-term memory (Bi-LSTM) network for automated analysis of heart sounds. The cross-correlation technique identifies temporal relationships among heart sound segments, allowing the model to effectively distinguish between healthy and pathological cardiac patterns. The derived features are then analyzed using a Bi-LSTM network enhanced with an attention module, which emphasizes the most relevant portions of the input sequence. This architecture delivers strong performance in terms of accuracy, sensitivity, and specificity across various categories of heart sounds when tested on a standard PCG dataset. Furthermore, attention visualization maps are produced to display the regions that receive greater importance during classification, offering a clearer understanding of how the model processes information. Temporal attention distributions are also illustrated to interpret the reasoning behind the network’s predictions, identifying specific time intervals that play a key role in decision-making. These visual insights correspond closely with clinical assessments of abnormal heart sounds, thereby improving interpretability and demonstrating the model’s practical value in cardiac diagnostics.