结合连续小波变换与生成对抗网络的癫痫发作预测  

Epileptic Seizure Prediction Based on Continuous Wavelet Transform and Generative Adversarial Network

作　　者：廖家慧 杨丰[1] 詹长安[1] 张利云 Liao Jiahui;Yang Feng;Zhan Chang'an;Zhang Liyun(School of Biomedical Engineering,Southern Medical University,Guangzhou 510515,China)

机构地区：[1]南方医科大学生物医学工程学院,广州510515

出　　处：《中国生物医学工程学报》2023年第2期168-179,共12页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(61771233)。

摘　　要：目前半监督深度学习模型已成功用于脑电信号(EEG)的癫痫发作预测,但该模型在EEG预处理方式与半监督模型稳定性等方面还有提升空间。本研究提出一种结合连续小波变换(CWT)与基于梯度惩罚的Wasserstein生成对抗网络(WGAN-GP)的改进方法(CWT-WGAN-GP)。首先对未标记的EEG信号进行CWT获得时频图,并结合特定患者的EEG数据集训练WGAN-GP模型,生成高性能的特征提取器;其次,以经过训练的WGAN-GP的判别器为特征提取器、两个全连接网络层为分类器(预测器),用少量有标记的EEG信号CWT时频图完成分类模型训练;最后,WGAN-GP的判别器与稳定的全连接网络组成半监督深度学习预测模型,用于癫痫发作预测。用CHBMIT头皮脑电数据集中所筛选的13例患者数据,评估改进的半监督癫痫发作预测模型,并与现有半监督方法相比。该方法在灵敏度、特异性、准确率和AUC指标上分别达到82.69%,67.48%,82.08%和84.03%,将原有的性能指标分别提升14.48%,34.45%,7.87%和11.4%;CWT-WGAN-GP的预测性能与现有方法的差异具有统计学意义(P<0.05)。CWT与WGAN-GP模型相结合能有效地改善半监督深度学习模型预测性能,在癫痫发作预测中发挥无监督特征提取的优化作用。Nowadays,semi-supervised deep learning model has been successfully applied in epileptic seizure forecasting based on electroencephalogram(EEG),however,there is still room for improvement in EEG preprocessing method and stability of semi-supervised model.This paper proposed an improved solution which combined continuous wavelet transform(CWT)and Wasserstein generative adversarial network based gradient penalty(WGAN-GP).Firstly,CWT was performed on unlabeled EEG to acquire spectrograms,and WGANGP was trained using the EEG dataset of specific patients to get a high-performance feature extractor.Then the trained discriminator of WGAN-GP was used as a feature extractor and two fully connected layers were used as classifier.A small amount of spectrogram of CWT of labeled EEG were used to complete the training of classifier model.Finally,the discriminator of WGAN-GP and the fully-connected network constituted a semi-supervised deep learning prediction model,carrying out epileptic seizure forecasting.The proposed semi-supervised patientspecific seizure forecasting method was evaluated by CHB-MIT scalp EEG dataset and compare the performance with present semi-supervised method.The sensitivity,specificity,accuracy and AUC of the proposed method reached 82.69%,67.48%,82.08%and 84.03%,respectively,improving the original performance by 14.48%,34.45%,7.87%and 11.4%.Compared to the present semi-supervised method,the difference of the prediction performance of CWT-WGAN-GP and current methods is showing significance(P<0.05).The result showed that the combination of CWT and WGAN-GP effectively improved the prediction performance of semisupervised deep learning model,and played an optimized role of unsupervised feature extraction in the epileptic seizure prediction.

关 键 词：癫痫发作预测 头皮脑电信号 深度学习 连续小波变换 基于梯度策略的Wasserstein生成对抗网络 

分 类 号：R318[医药卫生—生物医学工程]