Scholarlink

<p>Sleep stage classification is crucial in sleep medicine, but manual scoring is time-consuming, and automated solutions often struggle with complex sleep patterns. This study introduces a novel approach combining multi-scale temporal fusion with sequential XGBoost (eXtreme Gradient Boosting) processing. The approach analyzes polysomnographic data at multiple time scales (30, 15, and 5&#xa0;s) while incorporating temporal context through sequential processing. Developed from direct clinical experience in sleep scoring, where experts evaluate both brief events and broader stage transitions, the method prioritizes practical applicability in sleep medicine settings. The method was validated on a clinical dataset (224 polysomnographic recordings) and the Sleep-EDF expanded database, achieving accuracy rates of 81.5% (Cohen's kappa [κ] = 0.742) on clinical data and up to 91.5% (κ = 0.826) on Sleep-EDF data. The sequential processing notably enhanced non-rapid eye movement stage 1 (N1) detection, with F1-score improvements ranging from 27 to 61% across datasets. For datasets of approximately 170 recordings, model training and validation requires up to 12&#xa0;min on standard hardware. The results suggest this combined approach shows promise as a practical tool for automated sleep staging, though further research is needed to improve its performance and validate its utility across diverse clinical settings.</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

Automated sleep staging using sequential XGBoost and multi-scale temporal fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月10日

查看全文
  • Jiří Kuchyňka,
  • Oldřich Vyšata

摘要

Sleep stage classification is crucial in sleep medicine, but manual scoring is time-consuming, and automated solutions often struggle with complex sleep patterns. This study introduces a novel approach combining multi-scale temporal fusion with sequential XGBoost (eXtreme Gradient Boosting) processing. The approach analyzes polysomnographic data at multiple time scales (30, 15, and 5 s) while incorporating temporal context through sequential processing. Developed from direct clinical experience in sleep scoring, where experts evaluate both brief events and broader stage transitions, the method prioritizes practical applicability in sleep medicine settings. The method was validated on a clinical dataset (224 polysomnographic recordings) and the Sleep-EDF expanded database, achieving accuracy rates of 81.5% (Cohen's kappa [κ] = 0.742) on clinical data and up to 91.5% (κ = 0.826) on Sleep-EDF data. The sequential processing notably enhanced non-rapid eye movement stage 1 (N1) detection, with F1-score improvements ranging from 27 to 61% across datasets. For datasets of approximately 170 recordings, model training and validation requires up to 12 min on standard hardware. The results suggest this combined approach shows promise as a practical tool for automated sleep staging, though further research is needed to improve its performance and validate its utility across diverse clinical settings.