机器学习对局灶性癫痫发作间期脑电信号模式的分析

Epilepsy is a common chronic neurological disorder, among which focal epilepsy contributes the majority. Early diagnosis is crucial to prompt treatment and optimal prognosis while atypical clinical manifestations and unremarkable interictal electroencephalogram (EEG) findings among focal epilepsy make the diagnosis delayed and inaccurate. Herein we investigate the interictal epileptiform free EEG signal by using state-of-the-art machine learning (ML) technology to detect the specific EEG pattern in newly diagnosed focal epilepsy. Preliminary preparation has done including establishment of a network server for calculation (training the model) and application of the open source ML algorism. In details, bivariate features associated with EEG synchronization will be extracted firstly and aggregated over consecutive time points, to form patterns. Later patient-specific machine learning-based classifiers (support vector machines, logistic regression or convolutional neural networks) will be trained to discriminate interictal EEG patterns in focal epilepsy from that in normal controls. It is a challenge and breakthrough to traditional EEG research. The promising results will benefit the early diagnosis and the detection of the potential epilepsy candidates in population, which have a positive effect on health economic policy of epilepsy.

癫痫是常见的慢性神经系统疾病，其中局灶性癫痫居多。早期的诊断有利于及时治疗以获得最佳预后，然而局灶性癫痫往往由于发作表现的多样化以及难以捕捉具有诊断意义脑电图，导致诊断延误。本研究突破传统脑电研究模式，拟通过时下较为前沿的机器学习探索新诊断局灶性癫痫无癫痫放电的发作间期脑电信号特征。前期准备工作包括搭建处理数据的网络服务器平台以及对机器学习开源算法及代码的测试。我们首先提取与脑电信号同步化相关的双变量，然后将这些变量集合到连续时间维度中以形成个体特异的脑电模式。最后通过这些模式训练机器学习分类器（支持向量机、Logistic回归、卷积神经网络），用以区分正常人与局灶性癫痫患者的间期脑电模式。其成果有助于局灶性癫痫的早期诊断以及对人群中潜在癫痫人群的筛查，从而对癫痫的卫生经济政策产生积极影响。

癫痫是常见的慢性神经系统疾病，其中局灶性癫痫居多。早期的诊断有利于及时治疗以获得最佳预后，然而局灶性癫痫往往由于发作表现的多样化以及难以捕捉具有诊断意义脑电图，导致诊断延误。本研究突破传统脑电研究模式，拟通过时下较为前沿的机器学习探索癫痫无癫痫放电的发作间期脑电信号特征。由于疫情的影响，本研究由患者数据收集转为提取线上开源数据库资料（TUH）。通过筛选合适研究对象，初步收集20例癫痫患者（发作间期正常范围脑电）及20例非癫痫患者脑电（正常范围脑电）数据。通过数据预处理，每个个体提取10个清醒脑电片段（每个持续4秒）。全头部及每个导联的时频特征被提取用于后期机器学习。多层感知机，神经网络，卷积神经网络用于探索单个导联的特征能量谱。基于F1导联的特征能量谱初步发现，多层感知机，神经网络，卷积神经网络预测准确度为55.24%, 59.13% 和66.31%。 结论：虽然初步的研究结果未发现具有较强预测性的学习模型，但该领域潜在巨大空间值得进一步深入研究。