摘要
癫痫发作自动检测对于癫痫的诊断、监测和干预治疗具有重要意义。针对癫痫发作期脑电(EEG)数据有限以及不同患者EEG分布差异大等问题,提出一种基于判别性流形正则与域分布适配的跨被试癫痫发作检测方法。首先将待测试患者的EEG和其他患者已标注类别的EEG分别作为目标域和源域数据,并提取EEG小波包分解系数的均值、方差和样本熵作为特征;然后,采用含有源域样本类别信息的流形正则和类内间距最小化约束,进行领域分布适配,并利用条件分布距离和边缘分布距离的相对偏差,对分布适配权重加以动态调整;最后,利用空间投影后的源域样本训练随机森林分类器,实现对癫痫EEG的模式分类和发作检测。利用CHB-MIT数据库中24例患者的头皮脑电数据,验证所提方法的检测性能,并与现有的域适应算法相比较。所提出方法达到的平均检测灵敏度和准确率分别为94.94%和95.66%,比采用二阶统计量对齐的CORAL算法提高了15.07%和9.98%,比只进行均衡分布适配的BDA算法提高了3.90%和2.52%。判别性流形正则与域分布适配相结合能够减小不同患者脑电信号之间的分布差异,并有效利用源域数据流形结构和标签中的判别信息,为跨被试癫痫发作检测研究提供一种新思路。
Automatic detection of seizure is of great significance for epilepsy diagnosis,monitoring,and intervention treatment.Aiming to address the problems that ictal electroencephalogram(EEG)is limited and data distributions among different patients are significantly different,a cross-subject seizure detection method was proposed based on discriminative manifold regularization and domain distribution adaptation in this work.First,the EEG of patients to be tested and the labeled EEG of other patients were used as target and source domain data,respectively,and the EEG features such as the mean,variance and sample entropy of the wavelet packet decomposition coefficients were extracted.Then,a manifold regularization containing category information of source samples and an intra-class distance minimization constraint were introduced into domain distribution adaptation.Meanwhile,the relative deviation between conditional distribution distance and marginal distribution distance was adopted to dynamically adjust the distribution weight.Finally,pattern classification and seizure detection of target data were realized using the random forest classifier trained by source domain samples after space projection.The detection performance of the proposed method was validated using scalp EEG data from 24 patients in the CHB-MIT database and compared with existing domain adaptation algorithms.The average detection sensitivity and accuracy achieved by the proposed method were 94.94%and 95.66%,respectively,which were 15.07%and 9.98%higher than the CORAL algorithm using second-order statistic alignment,and 3.90%and 2.52%higher than the BDA algorithm that only performs balanced distribution adaptation.In conclusion,the combination of discriminative manifold regularization and domain distribution adaptation reduced the distribution differences between EEG signals from different patients and effectively utilized the discriminative information in the manifold structure and labels of the source domain data,providing a new idea for the research of cross-subject epileptic seizure detection.
作者
张艳丽
邱文龙
周卫东
Zhang Yanli;Qiu Wenlong;Zhou Weidong(School of Information and Electronic Engineering,Shandong Technology and Business University,Yantai 264005,Shandong,China;School of Integrated Circuits,Shandong University,Jinan 250101,China)
出处
《中国生物医学工程学报》
CSCD
北大核心
2024年第6期693-701,共9页
Chinese Journal of Biomedical Engineering
基金
国家自然科学基金(61801269,62271291)。
关键词
脑电信号
癫痫发作检测
跨被试
域分布适配
流形正则
electroencephalogram(EEG)
epileptic seizure detection
cross-subject
domain distribution adaptation
manifold regularization
作者简介
通信作者:张艳丽,E-mail:yrmzhang@sdtbu.edu.cn。
