基于多视图与注意力机制的睡眠脑电分期
Sleep EEG staging based on multi-view and attention mechanism
李兰亭 1苗敏敏2
作者信息
- 1. 湖州师范学院信息工程学院 湖州 313000
- 2. 湖州师范学院信息工程学院 湖州 313000;浙江省现代农业资源智慧管理与应用研究重点实验室 湖州 313000
- 折叠
摘要
为了更全面地对睡眠脑电进行特征提取,提出一种基于多视图与注意力机制的睡眠脑电分期方法.首先针对原始睡眠脑电信号构造时域和时频域两类视图数据;然后设计融合注意力机制的混合神经网络对多视图数据进行表征学习;接着通过双向长短时记忆(bi-directional long short-term memory,BiLSTM)网络进一步学习睡眠阶段之间的转换规则;最后使用Softmax函数进行睡眠分期,并利用类别加权损失函数解决睡眠数据类别不均衡的问题.实验使用Sleep-EDF数据库中前20名受试者的单通道脑电信号并采用20折交叉验证对模型进行性能评估,睡眠分期准确率达到83.7%,宏平均F1 值达到79.0%,Cohen's Kappa系数达到0.78.与现有方法相比,算法性能提升明显,证明了所提方法的有效性.
Abstract
In order to extract features of sleep EEG more comprehensively,a sleep EEG staging method based on multi-view and attention mechanism is proposed.First,two types of view data,namely time domain and time-frequency domain,are constructed based on the original sleep EEG signal.Then,a hybrid neural network with attention mechanism is designed to perform representation learning on multi-view data.Next,the transition rules between sleep stages are further learned through a bidirectional long short-term memory network.Finally,the Softmax function is used for sleep staging,and the class weighted loss function is utilized to solve the problem of unbalanced sleep data categories.In this experiment,the single-channel EEG signals of the first 20 subjects in the Sleep-EDF database are used,and 20-fold cross-validation is adopted to evaluate the performance of the model.The accuracy of sleep staging reaches 83.7%,the macro-F1-score(MF1)reaches 79.0%,and the Cohen's Kappa coefficient reaches 0.78.Compared with the existing methods,the performance of the algorithm in this paper is significantly improved,which proves the effectiveness of the proposed method.
关键词
睡眠分期/多视图/注意力机制/双向长短时记忆网络/类别加权损失函数Key words
sleep staging/multi-view/attention mechanism/bidirectional long short-term memory networks/class weighted loss function引用本文复制引用
基金项目
国家自然科学基金青年科学基金(62101189)
浙江省自然科学基金(LTGC23F010001)
出版年
2024
NETL
NSTL
万方数据