Scholarlink

<p>Electrocardiogram (ECG) plays a critical role in diagnosing cardiac diseases by analyzing essential information from the ECG waves. Its convenience has been appreciated by many people. However, like other forms of medical data, ECG gives rise to privacy concerns when distributed and analyzed. Generating synthetic data offers a potential strategy for addressing the privacy issues when sharing sensitive and confidential health information. Recent advancements in diffusion models have opened up exciting possibilities for generating synthetic data that closely resembles real-world data without compromising individual privacy. In this paper, we have developed an innovative electrocardiogram data generation model that harnesses the power of conditional diffusion models. Our approach seamlessly integrates several advanced techniques to effectively capture both long-term temporal dependencies and short-range signal information by incorporating structured state space model and patch-transformer. To enhance the model’s ability of distinguishing subtle features of different diseases, we have mapped 71 different diagnosis features into 71 different high-dimensional subspaces. Furthermore, we have leveraged the capabilities of cross-transformer to capture information from dual perspectives and employed multi-scale convolutional technology to process data at various resolutions. We performed the experiments on the PTB-XL dataset to generate synthetic 12-lead 10s electrocardiograms conditioned on more than 70 different ECG diagnostic statements. We evaluated the results of our experiment using both quantitative and qualitative criteria to provide a comprehensive analysis of our model’s performance.</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

Diffusion-based high dimensional subspace mapping for ECG generation with structured state space models

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Baofeng Zhu,
  • Chengbao Peng,
  • Xia Zhang,
  • Jiren Liu

摘要

Electrocardiogram (ECG) plays a critical role in diagnosing cardiac diseases by analyzing essential information from the ECG waves. Its convenience has been appreciated by many people. However, like other forms of medical data, ECG gives rise to privacy concerns when distributed and analyzed. Generating synthetic data offers a potential strategy for addressing the privacy issues when sharing sensitive and confidential health information. Recent advancements in diffusion models have opened up exciting possibilities for generating synthetic data that closely resembles real-world data without compromising individual privacy. In this paper, we have developed an innovative electrocardiogram data generation model that harnesses the power of conditional diffusion models. Our approach seamlessly integrates several advanced techniques to effectively capture both long-term temporal dependencies and short-range signal information by incorporating structured state space model and patch-transformer. To enhance the model’s ability of distinguishing subtle features of different diseases, we have mapped 71 different diagnosis features into 71 different high-dimensional subspaces. Furthermore, we have leveraged the capabilities of cross-transformer to capture information from dual perspectives and employed multi-scale convolutional technology to process data at various resolutions. We performed the experiments on the PTB-XL dataset to generate synthetic 12-lead 10s electrocardiograms conditioned on more than 70 different ECG diagnostic statements. We evaluated the results of our experiment using both quantitative and qualitative criteria to provide a comprehensive analysis of our model’s performance.