南方医科大学学报

南方医科大学学报 ›› 2023, Vol. 43 ›› Issue (1): 17-28.doi: 10.12122/j.issn.1673-4254.2023.01.03

• • 上一篇    下一篇

癫痫发作预测模型：斯托克韦尔变换的生成对抗与长短时记忆网络半监督方法

廖家慧，李涵懿，詹长安，杨 丰   

  1. 南方医科大学生物医学工程学院，广东 广州 510515
  • 出版日期:2023-01-20 发布日期:2023-02-22

Construction of an epileptic seizure prediction model using a semi-supervised method of generative adversarial and long short term memory network combined with Stockwell transform

LIAO Jiahui, LI Hanyi, ZHAN Chang'an, YANG Feng   

  1. School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China
  • Online:2023-01-20 Published:2023-02-22

RichHTML

25

PDF

299

摘要： 目的 提出一种半监督癫痫发作预测模型（ST-WGAN-GP-Bi-LSTM预测模型），从脑电（EEG）信号的时频分析、无监督特征模型稳定性以及后端分类器设计三个方面提升发作预测性能。方法 对癫痫EEG信号进行斯托克韦尔变换（ST变换）得到时频输入，通过自适应调节分辨率和保留绝对相位，定位癫痫EEG信号的时频成分；当生成数据分布和真实EEG数据分布无重叠时，为了避免JS散度均为常数而导致特征学习失效的问题，采用Wasserstein生成对抗网络作为特征学习模型，以EM距离结合梯度惩罚策略（WGAN-GP）引导的代价函数，约束模型的无监督训练过程，进而生成高阶特征提取器；构建基于双向长短时记忆网络（Bi-LSTM）的时序预测模型，在获取高阶EEG时频特征间时序相关性基础上提升癫痫分类（预测）性能。利用公开数据集CHB-MIT头皮脑电数据集对本文提出的ST-WGAN-GP-Bi-LSTM预测模型进行评估。结果 本文的ST-WGAN-GP-Bi-LSTM预测模型在AUC、灵敏度和特异性指标上分别达到90.40%，83.62%和86.69%。与现有半监督方法相比，将原有的性能指标分别提升17.77%，15.41%和53.66%，并与基于CNN的有监督预测模型性能持平。结论 本方法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。

关键词: 癫痫发作预测；头皮脑电信号；斯托克韦尔变换；生成对抗网络；双向长短期记忆网络

Abstract: Objective To propose a semi-supervised epileptic seizure prediction model (ST-WGAN-GP-Bi-LSTM) to enhance the prediction performance by improving time-frequency analysis of electroencephalogram (EEG) signals, enhancing the stability of the unsupervised feature learning model and improving the design of back-end classifier. Methods Stockwell transform (ST) of the epileptic EEG signals was performed to locate the time-frequency information by adaptive adjustment of the resolution and retaining the absolute phase to obtain the time-frequency inputs. When there was no overlap between the generated data distribution and the real EEG data distribution, to avoid failure of feature learning due to a constant JS divergence, Wasserstein GAN was used as a feature learning model, and the cost function based on EM distance and gradient penalty strategy was adopted to constrain the unsupervised training process to allow the generation of a high-order feature extractor. A temporal prediction model was finally constructed based on a bi-directional long short term memory network (Bi-LSTM), and the classification performance was improved by obtaining the temporal correlation between high-order time-frequency features. The CHB-MIT scalp EEG dataset was used to validate the proposed patient- specific seizure prediction method. Results The AUC, sensitivity, and specificity of the proposed method reached 90.40%, 83.62%, and 86.69%, respectively. Compared with the existing semi-supervised methods, the propose method improved the original performance by 17.77% , 15.41% , and 53.66% . The performance of this method was comparable to that of a supervised prediction model based on CNN. Conclusion The utilization of ST, WGAN-GP, and Bi-LSTM effectively improves the prediction performance of the semi-supervised deep learning model, which can be used for optimization of unsupervised feature extraction in epileptic seizure prediction.

Key words: epileptic seizure prediction; scalp EEG; Stockwell transform; generative adversarial network; bi- directional long short term memory network