Scholarlink

<p>Sleep staging is a critical process in the study and evaluation of sleep architecture, with its accuracy directly influencing the effectiveness of sleep disorder diagnosis. In temporal sleep staging, a common challenge arises from class imbalance—particularly, the significantly smaller number of N1 stage samples compared to other stages. Existing approaches have primarily focused on oversampling and expanding N1 stage data to address this issue. However, such methods may disrupt the temporal structure of the original signals and potentially lead to overfitting. To overcome these limitations, this paper proposes a Multi-Modal Spatio-Temporal Network for Imbalanced Sleep Staging (MSTN-ISNet), designed to perform data augmentation and enhancement for underrepresented sleep stages, thereby effectively alleviating the problem of insufficient training data for minority classes. First, we introduce a spatio-temporal statistics-driven time-frequency data augmentation module to enhance the representation of minority class samples. Next, we develop a deep interactive cross-modal feature encoding module to extract multi-modal spatio-temporal dynamic features from polysomnography (PSG) signals. Finally, a Markov chain entropy-weighted intelligent decision module is employed to calculate state transition probabilities among sleep stages, capturing the dynamic transitions between stages. Furthermore, to address the issue of class confusion, we implement strategies that reward minority classes, calculate inter-stage similarity, and utilize a weighted cross-entropy loss function to enhance the classification performance for minority stages. We evaluate the model on three benchmark datasets. Experimental results demonstrate classification accuracies of 90.8%, 90.4%, and 77.4%, respectively, outperforming existing state-of-the-art (SOTA) models. Furthermore, classification accuracies for N1-stage classification reached 85.4%, 92.5%, and 84.4%, respectively, exhibiting outstanding performance in minority class identification and surpassing all current advanced models.</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

MSTN-ISNet: A probability-guided spatio-temporal decoding for alleviating imbalance in sleep stage classification

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Dongrui Gao,
  • Shibing Li,
  • Haokai Zhang,
  • Qian Wu,
  • Zongyao Peng,
  • Shihong Liu,
  • Liping Xie,
  • Shaofei Ying,
  • Jiaxin Xie

摘要

Sleep staging is a critical process in the study and evaluation of sleep architecture, with its accuracy directly influencing the effectiveness of sleep disorder diagnosis. In temporal sleep staging, a common challenge arises from class imbalance—particularly, the significantly smaller number of N1 stage samples compared to other stages. Existing approaches have primarily focused on oversampling and expanding N1 stage data to address this issue. However, such methods may disrupt the temporal structure of the original signals and potentially lead to overfitting. To overcome these limitations, this paper proposes a Multi-Modal Spatio-Temporal Network for Imbalanced Sleep Staging (MSTN-ISNet), designed to perform data augmentation and enhancement for underrepresented sleep stages, thereby effectively alleviating the problem of insufficient training data for minority classes. First, we introduce a spatio-temporal statistics-driven time-frequency data augmentation module to enhance the representation of minority class samples. Next, we develop a deep interactive cross-modal feature encoding module to extract multi-modal spatio-temporal dynamic features from polysomnography (PSG) signals. Finally, a Markov chain entropy-weighted intelligent decision module is employed to calculate state transition probabilities among sleep stages, capturing the dynamic transitions between stages. Furthermore, to address the issue of class confusion, we implement strategies that reward minority classes, calculate inter-stage similarity, and utilize a weighted cross-entropy loss function to enhance the classification performance for minority stages. We evaluate the model on three benchmark datasets. Experimental results demonstrate classification accuracies of 90.8%, 90.4%, and 77.4%, respectively, outperforming existing state-of-the-art (SOTA) models. Furthermore, classification accuracies for N1-stage classification reached 85.4%, 92.5%, and 84.4%, respectively, exhibiting outstanding performance in minority class identification and surpassing all current advanced models.