基于迁移学习和空洞卷积的癫痫状态识别方法

Epileptic States Recognition using Transfer Learning and Dilated CNN

癫痫患者脑电信号的自动检测和发作诊断对临床治疗癫痫具有重要意义。针对训练数据有限及训练与测试数据分布不一致的难点,采用领域间联合知识迁移学习方法,实现小训练数据量下的癫痫状态识别。首先对脑电信号进行4层小波包分解,提取小波包分解系数作为特征,通过边缘分布和联合分布迭代调整,完成源域和目标域特征之间的知识迁移,训练空洞卷积神经网络作为分类器,完成目标域癫痫状态识别。分别在波士顿儿童医院CHB-MIT脑电数据集(22例被试,共计790 h)和波恩大学癫痫脑电数据集(5组,每组100个片段,每段23.6s)上进行算法验证,实验结果表明,所提出的方法对复杂癫痫状态的平均识别准确度、敏感性、特异性在CHB-MIT数据集上达到96.8%、96.1%、96.4%;在波恩数据集上,平均识别准确率为96.9%,有效提高了癫痫状态识别综合性能,实现了癫痫发作稳定可靠检测。

The automatic detection and seizure diagnosis of EEG signals in patients with epilepsy is of great significance for clinical treatment of epilepsy.Aiming at solving the problem in the conventional method that the labeled training data volume is insufficient and the classification accuracy of seizure is low due to the inconsistent distribution of training and test data,a joint knowledge transfer method between domains was proposed in this work.Firstly,the EEG signal was decomposed by four-layer wavelet packet,and the wavelet packet decomposition coefficients of 16 frequency bands were extracted as features.The marginal and joint distribution iterative adaptation were used to complete the knowledge transfer between the source and target domain.The dilated convolutional neural network was trained to complete the target domain recognition.In this study,the algorithms were estimated on two public EEG datasets including CHB-MIT dataset(22 patients,790 hours’ recording) and Bonn dataset(5 groups,one hundred 23.6 s episodes in each group).The experimental results showed that the average recognition accuracy,sensitivity and specificity of the proposed method for different epilepsy states was 96.8%,96.1%,96.4% on CHB-MIT dataset respectively.The average recognition accuracy was 96.9% on the Boon dataset,which effectively improved the comprehensive performance of seizure detection and achieve reliable detection results.