Scholarlink

<p>With the widespread use of wearable electrocardiographic (ECG) devices, there’s a growing need for efficient processing of large-scale real-time data to detect cardiovascular diseases. Deep learning, known for its accuracy in ECG signal analysis, has emerged as a crucial tool in computer-aided diagnosis. Leveraging two-dimensional (2-D) representations like time–frequency diagrams, Poincaré plots, and Gramian Angular Fields can enhance deep learning’s capability in capturing edge and texture features. However, the computational complexity of high-resolution images poses challenges for clinical application. Methods: This study proposes a Z-shaped reconstruction method to transform 1-D time series into 2-D modalities. Additionally, this study introduces a multiscale Squeeze-and-Excitation based convolutional neural network (SE-ConvNet) that integrates multiscale convolutional kernels and attention mechanisms. This facilitates rapid localization of key channel information while simultaneously reducing parameter count and computational costs. Our method achieves a significantly lower FLOPs (185&#xa0;M) compared to inputting 2-D images (2529&#xa0;M). The accuracies of the proposed method on the public database and clinical dataset were 99.30 and 99.04%, respectively, with F1 scores of 99.09 and 99.03%. Moreover, we verify its generalization ability, demonstrating its potential for practical clinical use.</p> Graphical abstract <p></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

Efficient transformation of ECG signals from 1-D to 2-D for atrial fibrillation detection using deep learning

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Jiahui Gao,
  • Yongjian Li,
  • Meng Chen,
  • Xiuxin Zhang,
  • Yiheng Sun,
  • Xinge Jiang,
  • Shoushui Wei

摘要

With the widespread use of wearable electrocardiographic (ECG) devices, there’s a growing need for efficient processing of large-scale real-time data to detect cardiovascular diseases. Deep learning, known for its accuracy in ECG signal analysis, has emerged as a crucial tool in computer-aided diagnosis. Leveraging two-dimensional (2-D) representations like time–frequency diagrams, Poincaré plots, and Gramian Angular Fields can enhance deep learning’s capability in capturing edge and texture features. However, the computational complexity of high-resolution images poses challenges for clinical application. Methods: This study proposes a Z-shaped reconstruction method to transform 1-D time series into 2-D modalities. Additionally, this study introduces a multiscale Squeeze-and-Excitation based convolutional neural network (SE-ConvNet) that integrates multiscale convolutional kernels and attention mechanisms. This facilitates rapid localization of key channel information while simultaneously reducing parameter count and computational costs. Our method achieves a significantly lower FLOPs (185 M) compared to inputting 2-D images (2529 M). The accuracies of the proposed method on the public database and clinical dataset were 99.30 and 99.04%, respectively, with F1 scores of 99.09 and 99.03%. Moreover, we verify its generalization ability, demonstrating its potential for practical clinical use.

Graphical abstract