Scholarlink

<p>Sleep staging is important for diagnosing and managing sleep disorders. However, artificial intelligence-based models still struggle with explainability, sequence plausibility, and personalization across subjects. We propose <i>LogicSleep</i>, a neurosymbolic framework that integrates transition rules with deep learning models for explainable, physiologically valid, and subject-adaptive sleep staging. <i>LogicSleep</i> extracts multi-level evidence from raw polysomnography via a CNN-graph U-Net backbone, providing explainable patterns beyond classification. Sequence plausibility is enforced by combining Conditional Random Fields with Deep Learning with Differentiable Logics (DL2) rules, which discourage implausible transitions, promote physiologically valid ones, and regularize dwell times; the same rules also guide subject-specific test-time adaptation via entropy minimization. Across different datasets (i.e., ISRUC_S1, ISRUC_S3, and MASS_SS3 datasets), under both within- and cross-dataset settings, <i>LogicSleep</i> consistently outperforms strong baselines and improves sequence-level metrics. Ablations confirm complementary contributions of each component. <i>LogicSleep</i> advances sleep staging toward physiologically plausible, personalized, and explainable predictions.</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

LogicSleep: a neurosymbolic-guided framework for explainable and personalized sleep staging

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Jingying Ma,
  • Qika Lin,
  • Feng Wu,
  • Yucheng Xing,
  • Ziyu Jia,
  • Mengling Feng

摘要

Sleep staging is important for diagnosing and managing sleep disorders. However, artificial intelligence-based models still struggle with explainability, sequence plausibility, and personalization across subjects. We propose LogicSleep, a neurosymbolic framework that integrates transition rules with deep learning models for explainable, physiologically valid, and subject-adaptive sleep staging. LogicSleep extracts multi-level evidence from raw polysomnography via a CNN-graph U-Net backbone, providing explainable patterns beyond classification. Sequence plausibility is enforced by combining Conditional Random Fields with Deep Learning with Differentiable Logics (DL2) rules, which discourage implausible transitions, promote physiologically valid ones, and regularize dwell times; the same rules also guide subject-specific test-time adaptation via entropy minimization. Across different datasets (i.e., ISRUC_S1, ISRUC_S3, and MASS_SS3 datasets), under both within- and cross-dataset settings, LogicSleep consistently outperforms strong baselines and improves sequence-level metrics. Ablations confirm complementary contributions of each component. LogicSleep advances sleep staging toward physiologically plausible, personalized, and explainable predictions.