基于多模态融合策略和注意力机制的睡眠自动分期模型  

Automatic sleep staging model based on multimodal fusion strategy and attention mechanism

作　　者：陈丽娟 王磊 沙宪政[3] 常世杰[3] 陈勇[1] CHEN Lijuan;WANG Lei;SHA Xianzheng;CHANG Shijie;CHEN Yong(Nantong First People's Hospital,Nantong 226001,China;Liaoning Medical Device Test Institute,Liaoning Inspection,Examination and Certification Centre,Shenyang 110036,China;School of Intelligent Medicine,China Medical University,Shenyang 110122)

机构地区：[1]南通市第一人民医院,南通226001 [2]辽宁省检验检测认证中心辽宁省医疗器械检验检测院,沈阳110036 [3]中国医科大学智能医学学院,沈阳110122

出　　处：《生物医学工程研究》2025年第1期24-30,共7页Journal Of Biomedical Engineering Research

基　　金：南通市卫生健康委员会科研项目(QNZ2022006)。

摘　　要：针对现有睡眠分期研究仅围绕单通道脑电(EEG)数据,无法有效利用睡眠状态转换规则的问题,本研究基于多模态融合策略与注意力机制,提出了一种自动睡眠分期模型。首先,构建表征学习模块捕捉多模态睡眠信号特征,挖掘特征通道间的关系;然后,设计多通道融合策略加强对特征的校准学习,并融合多模态信号间的互补信息;最后,将融合后的特征输入上下文通道依赖学习模块,利用注意力机制学习睡眠信号的上下文关系,以获得精准的睡眠分期结果。结果表明,该模型在Sleep-EDF-20、Sleep-EDF-78和蒙特利尔睡眠研究档案(MASS)三个公共数据集上的准确率分别为85.9%、85.2%和88.5%,宏平均F1分数(MF1)分别为80.8%、80.0%和82.1%。本研究模型的准确率和鲁棒性优于其他模型,可为睡眠分期提供技术参考。To solve the problem that existing studies on sleep staging only focus on single channel electroencephalogram(EEG)data,cannot effectively use sleep state transition rules,we proposed an automatic sleep staging model based on multimodal feature fusion and attention mechanism.Firstly,the representation learning module was constructed to capture the characteristics of multimodal sleep signals and explore the relationships between feature channels.Subsequently,a multichannel fusion strategy was designed to enhance the calibration learning of features and integrate complementary sleep information from multimodal signals.Finally,the fused features were input into the context channel dependency learning module,where attention mechanism was utilized to learn the contextual relationships within sleep signals,to achieve precise sleep staging outcomes.The results showed that the accuracy of this model on the three public datasets Sleep-EDF-20,Sleep-EDF-78 and Montreal arohive of sleep studies(MASS)was 85.9%,85.2% and 88.5%,respectively,and the macro average F1 score(MF1)was 80.8%,80.0% and 82.1%,respectively.The accuracy and robustness of this model are superior to the other models,which can provide technical reference for sleep staging.

关 键 词：睡眠分期 多模态信号 深度学习 特征融合 编码器 分类网络 

分 类 号：R318[医药卫生—生物医学工程] TP391[医药卫生—基础医学] TP183[自动化与计算机技术—计算机应用技术]