Scholarlink

With the advancement of deep learning technology, automatic sleep staging methods based on electroencephalogram (EEG) have garnered widespread attention. Despite previous research on sleep staging achieving high classification performance, several challenges remain unresolved: 1) How to effectively extract and integrate features of sleep at different scales. 2) How to enhance the model’s ability to focus on key features of sleep. 3) How to address the issue of class imbalance in training samples. To tackle these challenges, we incorporate an end-to-end hybrid architecture, CTMASleep, which significantly enhances the model’s ability to extract global and local features by integrating a deep temporal feature extraction module and a contextual temporal fusion module. Additionally, by incorporating a sequence feature reconstruction task, the model’s attention to key feature structures is strengthened. Furthermore, we propose a dynamic class balance loss function to address the class imbalance issue. Experimental results demonstrate that CTMASleep outperforms existing state-of-the-art models on both the Sleep-EDF-20 and Sleep-EDF-78 datasets.

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

CTMASleep: A Multi-task Learning Framework for Single-Channel Sleep Staging

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jiahao Yang,
  • Shaocong Yao,
  • Qian Qiu,
  • Jiahui Zhang,
  • Chuansheng Lin,
  • Jiahui Pan

摘要

With the advancement of deep learning technology, automatic sleep staging methods based on electroencephalogram (EEG) have garnered widespread attention. Despite previous research on sleep staging achieving high classification performance, several challenges remain unresolved: 1) How to effectively extract and integrate features of sleep at different scales. 2) How to enhance the model’s ability to focus on key features of sleep. 3) How to address the issue of class imbalance in training samples. To tackle these challenges, we incorporate an end-to-end hybrid architecture, CTMASleep, which significantly enhances the model’s ability to extract global and local features by integrating a deep temporal feature extraction module and a contextual temporal fusion module. Additionally, by incorporating a sequence feature reconstruction task, the model’s attention to key feature structures is strengthened. Furthermore, we propose a dynamic class balance loss function to address the class imbalance issue. Experimental results demonstrate that CTMASleep outperforms existing state-of-the-art models on both the Sleep-EDF-20 and Sleep-EDF-78 datasets.