Scholarlink

<p>The classification of sleep stages is considered essential for the interpretation of sleep architecture, the diagnosis of sleep disorders, and the evaluation of therapeutic outcomes. Conventional manual sleep staging is considered labor-intensive and susceptible to subjective bias. The accuracy and efficiency of sleep stage classification are enhanced by the automatic sleep staging model, which offers a reliable tool for clinical sleep research and the diagnosis of sleep disorders. An automated sleep stage classification model, which incorporates residual learning and a multi-level attention mechanism, is proposed in this study. The model used a residual network to fuse channel and spatial attention network to extract and reweight features and then used a self-attention mechanism and gated recurrent network to capture the important features of sleep stage conversion. Finally, the Softmax layer is utilized to complete the sleep staging task. In the meantime, a weighted cross-entropy loss function is applied during the training phase to mitigate the effects of class imbalance. Cross-validation was performed on the Sleep-EDF and Sleep-EDFx datasets to assess the model’s performance. The model yielded overall accuracies of 89.1% and 85.2% for the respective datasets. Cohen’s Kappa values of 0.85 and 0.80 were obtained, while the Macroaverage-F1 scores achieved 84.4% and 80.5%, respectively. It is suggested by the experimental results that the model demonstrates considerable potential for sleep assessment and diagnosis and can alleviate the workload of clinicians.</p> Graphical abstract <p>This study employs a sequence-to-sequence model framework, utilizing a convolution network to extract multi-level features from signals and a GRU network to capture sequence features. The proposed RCSA module solves the gradient degradation problem, recalibrates channel features and enhances the extraction of spatial features. A self-attention mechanism is added after GRU to emphasize important sequence features, achieving a recognition accuracy of 89.1% on the SleepEDF dataset.</p> <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

MLAR-SleepNet: a automatic sleep staging model based on residual and multi-level attention network

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • Keji Zhang,
  • Dechun Zhao,
  • Yuchen Shen,
  • Jin Liu,
  • Lu Qin

摘要

The classification of sleep stages is considered essential for the interpretation of sleep architecture, the diagnosis of sleep disorders, and the evaluation of therapeutic outcomes. Conventional manual sleep staging is considered labor-intensive and susceptible to subjective bias. The accuracy and efficiency of sleep stage classification are enhanced by the automatic sleep staging model, which offers a reliable tool for clinical sleep research and the diagnosis of sleep disorders. An automated sleep stage classification model, which incorporates residual learning and a multi-level attention mechanism, is proposed in this study. The model used a residual network to fuse channel and spatial attention network to extract and reweight features and then used a self-attention mechanism and gated recurrent network to capture the important features of sleep stage conversion. Finally, the Softmax layer is utilized to complete the sleep staging task. In the meantime, a weighted cross-entropy loss function is applied during the training phase to mitigate the effects of class imbalance. Cross-validation was performed on the Sleep-EDF and Sleep-EDFx datasets to assess the model’s performance. The model yielded overall accuracies of 89.1% and 85.2% for the respective datasets. Cohen’s Kappa values of 0.85 and 0.80 were obtained, while the Macroaverage-F1 scores achieved 84.4% and 80.5%, respectively. It is suggested by the experimental results that the model demonstrates considerable potential for sleep assessment and diagnosis and can alleviate the workload of clinicians.

Graphical abstract

This study employs a sequence-to-sequence model framework, utilizing a convolution network to extract multi-level features from signals and a GRU network to capture sequence features. The proposed RCSA module solves the gradient degradation problem, recalibrates channel features and enhances the extraction of spatial features. A self-attention mechanism is added after GRU to emphasize important sequence features, achieving a recognition accuracy of 89.1% on the SleepEDF dataset.