Abstract:

As clinical complex neuropsychiatric syndromes, schizophrenia and depression often have some similar clinical manifestations. A kind of fast and objective method of distinguishing them is needed. EEG, as a high time resolution and non-invasive method, which can directly detect and track dynamics of brain electrical activities, used as a potential bio-marker for distinguishing two disorders. It classifies the two disorders by training a Convolutional Neural Network（CNN） model using their rest state EEG. Firstly, the rest EEG spectrums are converted to gray scale images as input into the model. Secondly, the CNN model is used to extract features automatically and classify disorders. Thirdly, it uses cross validation to evaluate the model. Finally, data augmentation is applied to EEG data to improve the model performance. The classification accuracy, sensitivity and specificity of the model can reach up to 87.50%, 84.09% and 91.67%, respectively. The results show that original resting state EEG spectrums can be used, without manual feature selection, to distinguish schizophrenia and depression with high accuracy. EEG may be a promising bio-marker for distinguishing them. This study may provide reference value for the clinical diagnose between the two disorders.

Key words: Convolutional Neural Network（CNN）, schizophrenia, depression, rest state EEG, data augmentation

摘要：

精神分裂症和抑郁症作为临床精神科中的两种重性精神疾病综合征，两者的发病机制至今未明，因此临床上对于二者的区别常常依靠量表和医生的临床经验进行判断。脑电图（EEG）作为一种高时间分辨率的非侵入性诊断方法，可以作为两种疾病的一种潜在生物标记物。通过采集两种疾病临床静息态（睁眼和闭眼）的脑电数据，训练卷积神经网络（CNN）模型来对两种疾病进行分类，以试图找到一种区分两种疾病的客观手段。研究采集了70名精神分裂症和70名抑郁症患者静息态的EEG，通过分析提出了一种新的数据输入形式，即利用快速傅里叶变换将时域下的EEG数据转换到频域，再将频域数据转换成灰度图的形式输入卷积神经网络。不同于以往传统机器学习的方法，深度学习可以对数据进行自动的特征提取，而不需要人为筛选特征。而后采用交叉验证的方法对模型进行评估，并将数据增强的方法应用于EEG，来提高模型的性能。通过数据增强的方法可以使模型对两种疾病的分类准确度达到87.50%，敏感度和特异性分别达到84.09%和91.67%。实验结果表明卷积神经网络结合静息态脑电数据可以对精神分裂症和抑郁症进行较好的区分，EEG可以作为两种疾病的一种潜在生物标记物。

关键词: 卷积神经网络（CNN）, 精神分裂症, 抑郁症, 静息态脑电, 数据增强