Scholarlink

This paper presents an innovative method to classify ECG arrhythmias by using graph neural networks (GNNs) and continuous wavelet transform (CWT) scalograms together. While current deep learning techniques treat ECG data as pure time series without modeling structural interdependencies, traditional methods are unable to capture complex temporal-spatial relationships. We use CWT with Morlet wavelets to convert ECG signals into time-frequency representations, and then we extract 11-dimensional feature vectors from overlapping 8 × 8 patches with k-nearest neighbor connectivity to turn scalograms into graph structures. These representations are processed by a hierarchical Graph Attention Network in order to classify arrhythmias A thorough assessment on the MIT-BIH Arrhythmia Database, which comprises 1,09,446 heartbeat segments from five AAMI standard classes, shows excellent performance, yielding an accuracy of 95.37%. By efficiently capturing spatial relationships in time-frequency domain cardiac representations, the suggested GNN methodology shows off the potential of graph-based architectures for biomedical signal analysis and offers a strong basis for clinical arrhythmia detection systems.

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

Automated ECG Arrhythmia Classification Using CWT Scalogram Graph and Attention Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sushmita Pramanik Dutta,
  • Sriparna Saha,
  • Soamdeep Singha

摘要

This paper presents an innovative method to classify ECG arrhythmias by using graph neural networks (GNNs) and continuous wavelet transform (CWT) scalograms together. While current deep learning techniques treat ECG data as pure time series without modeling structural interdependencies, traditional methods are unable to capture complex temporal-spatial relationships. We use CWT with Morlet wavelets to convert ECG signals into time-frequency representations, and then we extract 11-dimensional feature vectors from overlapping 8 × 8 patches with k-nearest neighbor connectivity to turn scalograms into graph structures. These representations are processed by a hierarchical Graph Attention Network in order to classify arrhythmias A thorough assessment on the MIT-BIH Arrhythmia Database, which comprises 1,09,446 heartbeat segments from five AAMI standard classes, shows excellent performance, yielding an accuracy of 95.37%. By efficiently capturing spatial relationships in time-frequency domain cardiac representations, the suggested GNN methodology shows off the potential of graph-based architectures for biomedical signal analysis and offers a strong basis for clinical arrhythmia detection systems.