基于卷积神经网络和迁移学习的癫痫状态识别

Epileptic Seizure Recognition Using Convolutional Neural Networks and Transfer Learning

随着对癫痫状态神经元电活动研究的不断深入,针对癫痫患者的电磁刺激疗法备受关注,自动准确地识别癫痫发作状态是及时准确地实施电磁刺激的关键.因此,构建了一种由通用模型向个性化模型迁移的癫痫发作状态识别方法.首先,基于多个病患的脑电数据,采用一维卷积神经网络建立癫痫状态识别的通用模型,学习不同病患癫痫发作时脑电状态的共性特征,以实现对不同病患癫痫发作状态的通用识别;其次,基于单个病患的脑电数据,通过迁移学习将通用模型的参数迁移到个性化模型之中以简化模型训练、加速收敛,讨论了通用模型参数向个性化模型迁移的全面迁移方式和卷积层参数迁移方式的性能.使用CHB-MIT数据库中17例病患的长程脑电记录数据对算法进行验证,最终所有病患个性化模型的平均准确率达到了91.04%.基于个性化模型对病患的长程脑电记录进行癫痫发作起止时间判断,模型对癫痫发作和结束状态的检出率达到了96.43%和89.29%.结果表明,该模型发挥了深度学习无需手动提取、选择特征的优势,为癫痫状态识别方法用于癫痫治疗方案的开发提供了参考与依据.

With the deepening of research on the electrical activity of epileptic neurons,electromagnetic stimulation therapy for epileptic patients has attracted considerable attention.Automatic and accurate identification of epileptic seizure status is the key to the timely and accurate implementation of electromagnetic stimulation.In this study,a novel patient-specific seizure state recognition technique based on convolutional neural network(CNN)and transfer learning is proposed.First,on the basis of the electroencephalogram(EEG)recordings from multiple patients,the one-dimensional CNN is used to establish a general model for epileptic seizure state recognition.The general model is used to learn the common characteristics of EEG during seizures in different patients to achieve general recognition of seizure states.Then,on the basis of the EEG recordings from individual patients,the parameters of the general model are transferred to the personalized model using transfer learning to simplify model training and accelerate convergence.The model performance of the overall migration and convolution layer parameter migration modes of universal model parameters to the personalized model is also discussed.Finally,the algorithm is applied to long-term scalp EEG recordings of 17 patients in the CHB-MIT database.The average accuracy of all patient personalized models reaches 91.04%.On the basis of the personalized model,the patients’long-term EEG recordings are used to judge the onset and end of seizures.The detection rates of the onset and end of seizure states reach 96.43%and 89.29%,respectively,in the test dataset.Thus,the EEG-based seizure state recognition model using CNN and transfer learn-ing could be used in the development of treatment programs for patients with epilepsy.