基于注意力图池化的图卷积网络癫痫发作预测
Epileptic Seizure Prediction by Graph Convolutional Network Based on Graph Pooling of Attention
张倩云 1乔晓艳1
作者信息
- 1. 山西大学 物理电子工程学院,山西 太原 030006
- 折叠
摘要
对癫痫发作及时准确预测可在发作前对患者实施干预措施防止意外伤害.为提高癫痫发作预测的准确率,提出一种注意力图池化的图卷积神经网络模型,用于癫痫发作预测.将多导联脑电数据转换为图结构关系,设计改进的图卷积神经网络模型,通过嵌入注意力图池化,筛选重要节点信息,避免特征冗余,提高模型学习能力和稳健性.在此基础上,分析不同脑电节律、滑动时间窗口以及预测时长对癫痫预测的影响.仿真实验结果表明:采用4 s时间窗口0.5 s滑动步长,在癫痫发作前5 min预测可达到97.03%准确率,95.89%召回率,98.16%特异性和96.12%的F1值.该模型可以提高癫痫发作预测精度,具有较好泛化能力.
Abstract
Timely and accurate prediction of epileptic seizures can take intervention measures to prevent from accidental injury be-fore epileptic seizures.In order to improve the accuracy of epileptic seizure prediction,a graph convolution neural network model,based on graph pooling of attention,is proposed for epileptic seizure prediction.The multi-lead Electroencephalography(EEG)data were converted into a graph structure,and an improved graph convolution neural network model was designed.By adding graph pooling of attention,important node information was screened to avoid feature redundancy and improve the learning ability and ro-bustness of the model.On this basis,the effects of different EEG rhythm,sliding time window and prediction duration on epilepsy prediction were analyzed.The simulation results show that the accuracy,recall rate,specificity and F1 value of the prediction model can achieve separately 97.03%,95.89%,98.16%and 96.12%for 5 minutes before seizures,by using the sliding step size of 0.5 s in the 4 s time window.Therefore,this model can improve the prediction accuracy of epileptic seizure and can be easily generalized.
关键词
癫痫发作预测/脑电信号/图卷积神经网络/注意力图池化Key words
prediction of epileptic seizure/EEG signal/graph convolution neural network/graph pooling of attention引用本文复制引用
基金项目
山西省回国留学人员科研项目(2020-009)
出版年
2024
国家科技期刊平台
NSTL
万方数据