[1]额吉纳,喻文杰,费凌霞,等.多通道时空特征提取的癫痫发作预测模型[J].中国医学物理学杂志,2025,42(2):213-219.[doi:DOI:10.3969/j.issn.1005-202X.2025.02.011]
　E Jina,YU Wenjie,FEI Lingxia,et al.Epileptic seizure prediction model based on multichannel spatiotemporal feature extraction[J].Chinese Journal of Medical Physics,2025,42(2):213-219.[doi:DOI:10.3969/j.issn.1005-202X.2025.02.011]
点击复制

多通道时空特征提取的癫痫发作预测模型()
分享到：

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

卷:
42
期数:
2025年第2期
页码:
213-219
栏目:
医学信号处理与医学仪器
出版日期:
2025-01-20

文章信息/Info

Title:
Epileptic seizure prediction model based on multichannel spatiotemporal feature extraction
文章编号:
1005-202X（2025）02-0213-07
作者:
额吉纳1; 喻文杰1; 费凌霞2; 庄君2; 梁国华1; 杨丰1
1.南方医科大学生物医学工程学院， 广东 广州 510515； 2.广东三九脑科医院癫痫内科， 广东 广州 510520
Author(s):
E Jina1;  YU Wenjie1;  FEI Lingxia2;  ZHUANG Jun2;  LIANG Guohua1;  YANG Feng1
1. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China 2. Department of Epilepsy, Guangdong Sanjiu Brain Hospital, Guangzhou 510520, China
关键词:
癫痫发作预测; 多通道头皮脑电信号; 斯托克韦尔变换; 自适应特征提取; 双向邻域长短期记忆网络
Keywords:
Keywords: epileptic seizure prediction multichannel scalp EEG signal Stockwell transform adaptive feature extraction bidirectional neighborhood long short-term memory network
分类号:
R318；TP391
DOI:
DOI:10.3969/j.issn.1005-202X.2025.02.011
文献标志码:
A
摘要:
提出一种基于多通道时空特征提取的癫痫发作预测模型。该模型首先对原始多通道脑电图（EEG）信号进行斯托克韦尔变换，提取时频成分。针对癫痫发作前期和间期时频特征差异不明显的问题，设计了由ConvNeXt网络、SENet和金字塔池化模块组成的自适应特征提取模块，增强对每个EEG通道内关键时频特征的捕获能力。同时，构建基于Bi-NLSTM的预测模型，增强多通道高阶EEG时频特征之间的时空依赖性，进一步提高癫痫分类性能。在CHB-MIT数据集上，该模型的准确率、灵敏度、特异性和AUC分别达到96.0%、97.8%、96.8%和0.987，每小时假阳性率降至0.038，优于现有主流方法。消融实验验证各组件对提升模型性能的实际效果。本方法通过优化局部时频特征提取和增强多通道时空依赖性，有效提升癫痫发作预测的整体性能。
Abstract:
Abstract: A novel epileptic seizure prediction prediction model based on multichannel temporal and spatial feature extractions is presented. The model applies Stockwell transform to the original multichannel electroencephalogram (EEG) signals for extracting time-frequency components. To address the issue of insignificant difference between preseizure and interseizure time-frequency features, an adaptive feature module composing of ConvNeXt, SENet and pyramid pooling module is designed to enhance the ability of capturing key time-frequency features in each EEG channel. Meanwhile, a prediction model based on Bi-NLSTM is constructed to improve the spatiotemporal dependence between the time-frequency features of multichannel high-order EEG for further promoting the epilepsy classification performance. On the CHB-MIT dataset, the model has an accuracy, sensitivity, specificity and AUC of 96.0%, 97.8%, 96.8% and 0.987, respectively, and the false positive rate per hour decreased to 0.038, outperforming the existing mainstream methods. In addition, the effect of each component on the model performance is verified by ablation study. The proposed method improves the overall performance for seizure prediction effectively by optimizing local time-frequency feature extraction and enhancing multichannel spatiotemporal dependence. Keywords: epileptic seizure prediction multichannel scalp EEG signal Stockwell transform adaptive feature extraction bidirectional neighborhood long short-term memory network

备注/Memo

备注/Memo:
【收稿日期】2024-10-17 【基金项目】国家自然科学基金（61771233） 【作者简介】额吉纳，硕士研究生，研究方向：癫痫发作预测，E-mail: 934250561@qq.com 【通信作者】杨丰，博士，教授，博士生导师，E-mail: yangf@smu.edu.cn
更新日期/Last Update: 2025-01-22