J Shanghai Jiaotong Univ Sci ›› 2025, Vol. 30 ›› Issue (4): 720-732.doi: 10.1007/s12204-024-2734-z

• • 上一篇    下一篇

基于Transformer对比学习的自动睡眠分期方法

  

  1. 1. 上海交通大学 电子信息与电气工程学院,上海200240;2. 上海市儿童医院,上海200240;3.上海交通大学 机械与动力工程学院,上海200240
  • 收稿日期:2023-05-16 接受日期:2023-06-27 发布日期:2025-07-31

Transformer-Based Contrastive Learning Method for Automated Sleep Stages Classification

马进1,任泽1,张彤彤1,丁颍2,陆熠磊1,彭颖红3   

  1. 1. School of Electronic, Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China; 2. Children’s Hospital of Shanghai, Shanghai 200062, China; 3. School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
  • Received:2023-05-16 Accepted:2023-06-27 Published:2025-07-31

摘要: 自动睡眠分期由于其在分析整晚多导睡眠(PSG)信号方面具有高效性,能够有效支持临床专家对睡眠障碍进行诊疗。然而,现有的研究主要集中在与实际临床数据不相同的公共数据集上。为了缩小理论模型与实际临床实践之间的差距,提出了一种新的深度学习模型,将视觉Transformer与监督对比学习相结合,实现有效的睡眠阶段分期。实验结果表明,该模型能够更有效地对多通道PSG信号进行分期。在两个公开的睡眠数据库上该模型平均F1得分分别为79.2%和76.5%,优于之前的研究,表明了该模型强大的能力,在儿童小数据库上的平均准确率也达到了88.6%。提出的模型不仅在公共数据库上进行了验证,而且在提供的临床数据库上进行了验证,以严格评估其在临床实践中的使用情况。

关键词: 睡眠分期,视觉Transformer,对比学习,多导睡眠信号

Abstract: Automated sleep stages classification facilitates clinical experts in conducting treatment for sleep disorders, as it is more time-efficient concerning the analysis of whole-night polysomnography (PSG). However, most of the existing research only focused on public databases with channel systems incompatible with the current clinical measurements. To narrow the gap between theoretical models and real clinical practice, we propose a novel deep learning model, by combining the vision transformer with supervised contrastive learning, realizing the efficient sleep stages classification. Experimental results show that the model facilitates an easier classification of multi-channel PSG signals. The mean F1-scores of 79.2% and 76.5% on two public databases outperform the previous studies, showing the model’s great capability, and the performance of the proposed method on the children’s small database also presents a high mean accuracy of 88.6%. Our proposed model is validated not only on the public databases but the provided clinical database to strictly evaluate its clinical usage in practice.

Key words: sleep stages classification, vision transformer, contrastive learning, polysomnography (PSG) signal

中图分类号: