基于贝叶斯相关向量机的脑电睡眠分期

Classification of EEG sleep stage based on Bayesian relevance vector machine

针对支持向量机(SVM)计算复杂度高和参数不易确定的局限性,提出一种基于稀疏贝叶斯相关向量机(RVM)的脑电数据睡眠分期方法.给出二分类RVM的参数推理和优化,并确定了二叉树多分类RVM模型.基于8例健康成年人的MIT/BIH睡眠脑电实测数据,根据已有的专家人工睡眠分期注释,首先提取清醒期和睡眠各期脑电数据的样本熵值作为特征向量样本,然后利用二叉树多分类器法构建贝叶斯RVM睡眠分期模型,输入清醒期和各睡眠期样本进行训练和测试,最终实现各睡眠分期的模式分类.结果表明:在两种径向基核函数下,基于RVM的睡眠分期识别准确率最高达到89.00%,高于SVM方法(87.67%),且较SVM需要更少的支持向量数目及更短的测试时间,即RVM比传统的SVM具有更优的分类能力和更高的计算效率,是一种有效的睡眠分期识别方法.

To overcome the disadvantages of complicated calculation and uncertain parameter selection of support vector machine （SVM） , a new algorithm based on sparse Bayesian relevance vector machine （RVM） was proposed to classify electroencephalography （EEG） sleep stage. Inference and optimization of parameters of the binary classification RVM were given, and binary tree RVM multi-class model was established. According to the known sleep stage annotations by experts, sample entropy （SampEn） features of each sleep stage were extracted from the EEG sleep signals of eight healthy volunteers without any medication in MIT/BIH database. Then the sleep stage types were identified through multi-lay RVM pattern recognition classifier on binary tree categorization by training and testing samples of sleep and awake period. The results show that the maximal identification rate of RVM can reach 89.00% , which is better than that of the SVM （87.67%）. The number of relevance vectors and test time of RVM are both less than those of SVM, which means that the RVM method is an effective tool for sleep stage classification with better classification accuracy and computation efficiency.