基于CNN-BiLSTM的自动睡眠分期方法

针对目前睡眠分期存在的依赖人工特征提取、无法识别长时关联数据中的时序模式、模型对EEG时序数据分期不准确等问题,提出一种基于CNN-BiLSTM的自动睡眠分期方法.将原始数据通过改进MSMOTE算法进行过采样形成类平衡数据,再通过CNN表达其高级特征,并馈送至BiLSTM中挖掘各睡眠阶段间的依赖关系,实现睡眠数据分期特征的自动学习和睡眠周期判定.在Sleep-EDF公开数据集上的实验结果表明,CNN-BiLSTM模型的分类准确率为92.21%.同时引入改进的MSMOTE过采样技术缓解因数据不平衡所导致的少数类睡眠期判定不准确问题.在原始数据集类不平衡的情况下,实现了睡眠数据自动分期,有效提高了睡眠分期模型的准确率,具有一定的实用价值.

An Automatic Sleep Staging Method Based on CNN-BiLSTM

To solve the problems, including the manual dependence in the extraction of sleep staging features, difficult identification to the timing pattern in long-term correlated data, and the inaccuracy of EEG timing data staging in the model, and so on, an automatic sleep staging method based on CNN-BiLSTM was proposed. Firstly, the original data was over-sampled with improving the MSMOTE algorithm to form the class equilibrium data. And then the advanced features were expressed by CNN and fed to BiLSTM to explore the dependency relationship between sleep stages, so as to realize the automatic learning and sleep cycle determination of sleep data staging characteristics. The experimental results on the Sleep-EDF open data set show that the classification accuracy of the CNN-BiLSTM model can reach 92.21%. The improved over-sampling technique of MSMOTE can alleviate the problem of inaccuracy in the determination of sleep stage. In the case of unbalanced class of original data set, automatic sleep data staging is realized, which can effectively improve the accuracy of sleep staging model, possessing a certain practical value.