Scholarlink

<p>Sleep staging plays a crucial role in evaluating sleep quality and diagnosing various neurological and physiological disorders, including insomnia, sleep apnea, and narcolepsy. Conventionally, the annotation of sleep stages is performed manually or through semi-automated techniques based on polysomnography (PSG), requiring trained clinicians to interpret complex bio-signals. This process is inherently tedious, time-consuming, and subject to inter-scorer variability, which affects consistency and scalability in clinical practice. To address these challenges, we propose a Multi-interactive Learning Model for Sleep Staging (MILMSS), which leverages the heterogeneous nature of PSG signals by integrating a stacking-ensemble learning architecture (SEA) within a multi-interaction framework. The SEA module performs robust feature extraction from individual modalities such as EEG, EOG, and EMG, capturing local temporal dependencies and preserving physiological signal characteristics. These features are then passed into the multi-interactive learning block, which models cross-modal interactions and temporal continuity through adaptive attention mechanisms and hierarchical fusion. The model was rigorously evaluated on six publicly available PSG datasets: Sleep Heart Health Study (SHHS), Sleep European Data Format 2013 (S-EDF-13), Sleep-EDF 2018 (S-EDF-18), S-EDF-18 Sleep Cassette recordings + Sleep Telemetry recordings (S-EDF-18-SC + ST), Dreams (DRMS), and ISRUC-Sleep (SG3). Using cross-validation protocols across all datasets, our MILMSS model demonstrated consistently superior performance, achieving classification accuracies of 97.42%, 92.22%, 90.13%, 95.10%, 92.98%, and 96.49%, and corresponding Cohen’s kappa scores of 0.95, 0.89, 0.85, 0.93, 0.90, and 0.95, respectively. Overall, our proposed model offers a highly accurate, scalable, and interpretable solution for automated sleep staging. It holds strong potential for integration into both clinical diagnostic workflows and portable sleep monitoring systems, thereby addressing the growing need for reliable and cost-effective sleep health technologies.</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

A multi-interactive learning model for sleep staging based on polysomnography signals

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月28日

查看全文
  • Suren Kumar Sahu,
  • Santosh Kumar Satapathy,
  • Sudhir Kumar Mohapatra,
  • Getachew Mekuria Habtemariam

摘要

Sleep staging plays a crucial role in evaluating sleep quality and diagnosing various neurological and physiological disorders, including insomnia, sleep apnea, and narcolepsy. Conventionally, the annotation of sleep stages is performed manually or through semi-automated techniques based on polysomnography (PSG), requiring trained clinicians to interpret complex bio-signals. This process is inherently tedious, time-consuming, and subject to inter-scorer variability, which affects consistency and scalability in clinical practice. To address these challenges, we propose a Multi-interactive Learning Model for Sleep Staging (MILMSS), which leverages the heterogeneous nature of PSG signals by integrating a stacking-ensemble learning architecture (SEA) within a multi-interaction framework. The SEA module performs robust feature extraction from individual modalities such as EEG, EOG, and EMG, capturing local temporal dependencies and preserving physiological signal characteristics. These features are then passed into the multi-interactive learning block, which models cross-modal interactions and temporal continuity through adaptive attention mechanisms and hierarchical fusion. The model was rigorously evaluated on six publicly available PSG datasets: Sleep Heart Health Study (SHHS), Sleep European Data Format 2013 (S-EDF-13), Sleep-EDF 2018 (S-EDF-18), S-EDF-18 Sleep Cassette recordings + Sleep Telemetry recordings (S-EDF-18-SC + ST), Dreams (DRMS), and ISRUC-Sleep (SG3). Using cross-validation protocols across all datasets, our MILMSS model demonstrated consistently superior performance, achieving classification accuracies of 97.42%, 92.22%, 90.13%, 95.10%, 92.98%, and 96.49%, and corresponding Cohen’s kappa scores of 0.95, 0.89, 0.85, 0.93, 0.90, and 0.95, respectively. Overall, our proposed model offers a highly accurate, scalable, and interpretable solution for automated sleep staging. It holds strong potential for integration into both clinical diagnostic workflows and portable sleep monitoring systems, thereby addressing the growing need for reliable and cost-effective sleep health technologies.