文章摘要

文章信息/Info

Title:: Application of intelligent hybrid systems in precise sleep stage discrimination

作者:: 邵益梓1; 2; 黄柏恺1; 2; 杜利东1; 2; 3; 王鹏1; 2; 3; 李振锋1; 2; 3; 陈贤祥1; 2; 3△; 方震1; 2; 3△; （1.中国科学院空天信息创新研究院，北京100190；2.中国科学院大学电子电气与通信工程学院，北京 100049；3.中国医学科学院 “个性化呼吸慢病管理研究”创新单元，北京 100730）

Author(s):: SHAO Yizi 1; 2; HUANG Bokai 1; 2; DU Lidong 1; 2; 3; WANG Peng 1; 2; 3; LI Zhenfeng 1; 2; 3; CHEN Xianxiang 1; 2; 3 ; FANG Zhen 1; 2; 3; （1.Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China;2. School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, Beijing 100049;3.Personalized Management of Chronic Respiratory Disease, Chinese Academy of Medical Sciences, Beijing 100730）

Keywords:: Sleep stage classification; Hybrid intelligence; Multitask learning; Feature mapping; Interpretability; Sleep graphs correction

摘要:: 针对手工睡眠分期过程繁琐，自动睡眠分期模型存在精度不足或难以解释分类结果的问题，本研究提出了一种基于混合智能的自动睡眠分期模型，结合数据智能和知识智能以实现睡眠分期精度、可解释性和泛化性的平衡。首先，基于典型脑电（electroencephalography，EEG）和眼电（electrooculography，EOG）通道的任意组合，建立了基于U-Net架构的时序全卷积网络和多任务特征映射结构；其次，通过组合不同睡眠图校正方法，探究了知识智能对粗睡眠图的不同作用方式。本模型在ISRUC和Sleep-EDFx数据集上的F1指标分别为0.804、0.780。此外，本研究利用知识智能解决了模型得到的粗睡眠图跳变过多、睡眠阶段转换不合理的问题。结果表明，本研究能够为睡眠医师提供有效的判读辅助，在提高临床睡眠分期效率上具有巨大潜力。

Abstract:: In view of the cumbersome manual sleep staging process and the insufficient accuracy or the difficulty in interpreting classification results of automatic sleep staging model, we proposed an automatic sleep staging model based on mixed intelligence, which combined data intelligence and knowledge intelligence to achieve a balance of sleep staging accuracy, interpretability and generalization. Firstly, based on any combination of typical electroencephalography (EEG) and electrooculography (EOG) channels, a sequential full convolutional network and multi-task feature mapping structure of U-Net architecture were constructed. Secondly, by combining different sleep map correction methods, the different action ways of knowledge intelligence to rough sleep map was explored. The F1 index of this model on the ISRUC and Sleep EDFx datasets were 0.804 and 0.780, respectively. In addition, the knowledge intelligence was used to the excessive jump in the rough sleep map and unreasonable transition of sleep stages. This research can provide an effective interpretive aid for sleep physicians, and has great potential in improving the efficiency of clinical sleep staging.