抑郁症已然成为了全社会、全世界难以忽视的健康负担。其困境在于致病机理尚不明确，治疗手段有效率低，而且治疗周期漫长。在这样的背景下，通过某种手段在治疗之前对患者是否适合当前治疗方式进行鉴别，以提高治疗成功率，减少不必要的副作用，减轻患者的时间和经济负担，成为了当下亟待研究的重要问题。脑电图（electroencephalogram，EEG）的高时间分辨率和无创、便捷性使得它越来越多地被用做预测抑郁症患者治疗效果的生物标志物。本文以静息态EEG为研究对象，提出并实现了两种脑电数据驱动的疗效预测分类模型，用于判断抑郁症患者服用抗抑郁药物舍曲林8周后是否起效。为了保证模型预测的效果和稳定性，本文采用了一项大规模多中心、随机对照临床试验的公开数据集。为实现研究的可重复性并提高处理效率，本文研究了现有的自动化预处理流水线，并在其基础上优化了自动化预处理流程，衡量了预处理的效果，为后序分析奠定了基础。为提取EEG的通道间相关信息，本文提出了判别式EEG隐空间回归模型（Discriminant EEG Latent SpacE Regression，DELSER），该方法以基线期静息态EEG的协方差矩阵为输入，输出对治疗结果的预测。在此基础上，本文进一步采用了将具有预测能力的睁眼α频带EEG进一步划分子带的滤波器组策略。结果表明，这一策略使分类准确率提升了3.4%，从64.1%提高到67.5%。结合滤波器组的DELSER预测能力也强于传统量化EEG特征如Cordance，频带功率等。对不同数据规模的比较探究结果表明，数据规模越大模型准确率和稳定性越高。为将特征提取和模型训练结合，并减少对先验知识的依赖，本文进一步提出了端到端疗效预测神经网络模型CC-EEGNet。该网络以EEGNet为主干，通过加入局部通道卷积和跨频域通道卷积模块实现了对于静息态EEG的更强特征提取能力。结果表明，该网络相比EEGNet、DeepConvNet、ShallowConvNet等网络的分类准确率至少提升了7.8%。该模型的分类效果还显著优于共空间模式（Common Spatial Pattern,CSP）特征提取算法。此外，更大的数据规模也有利于准确率的提升。本文面向抑郁症药物治疗的预后，以大规模脑电数据集为基础，提出了两种数据驱动预测模型，探索了有效的预测性能提升途径，为抑郁症个性化精准治疗的推进提供了有益的尝试。

Depression has become a health burden that cannot be ignored by the whole society and the world. The dilemma is that the pathogenic mechanism is not yet clear, the effective rate of the treatment is low, and the treatment cycle is long. In this context, it has become an urgent research to identify whether the patient is suitable for the current treatment by some means before treatment, so as to improve the success rate of treatment, reduce unnecessary side effects, and reduce the time and economic burden of patients. question. The high temporal resolution, non-invasiveness and convenience of electroencephalogram (EEG) make it more and more used as a biomarker to predict the treatment effect of depression patients. In this paper, resting-state EEG was used as the research object, and two EEG data-driven curative effect prediction classification models were proposed and implemented, which were used to judge whether the antidepressant drug sertraline was effective after 8 weeks in patients with depression.In order to ensure the effect and stability of the model prediction, this paper uses a public dataset of a large-scale multi-center, randomized controlled clinical trial. In order to achieve the repeatability of the research and improve the processing efficiency, this paper studies the existing automated preprocessing pipeline, optimizes the automated preprocessing process based on it, measures the effect of preprocessing, and lays the foundation for subsequent analysis.In order to extract the inter-channel correlation information of EEG, this paper proposes a discriminant EEG latent space regression model (DELSER). predict. On this basis, this paper further adopts a filter bank strategy to further divide the eye-opening α-band EEG with predictive ability into subbands. The results show that this strategy improves the classification accuracy by 3.4%, from 64.1% to 67.5%. The predictive ability of DELSER combined with filter banks is also stronger than traditional quantized EEG (QEEG) features such as Cordance, Band Power, etc. The comparison and exploration results of different data scales show that the larger the data scale, the higher the accuracy and stability of the model.In order to combine feature extraction and model training and reduce the dependence on prior knowledge, this paper further proposes CC-EEGNet, an end-to-end therapeutic effect prediction neural network model based on resting-state EEG. The network uses EEGNet as the backbone, and achieves stronger feature extraction capabilities for resting EEG by adding local channel convolution and cross-frequency domain channel convolution modules. The results show that the classification accuracy of the network is at least 7.8% higher than that of EEGNet, DeepConvNet, ShallowConvNet and other networks. The classification performance of the model is also significantly better than that of the Common Spatial Pattern (CSP) feature extraction algorithm. In addition, larger data scale is also conducive to the improvement of accuracy.In this paper, for the prognosis of depression drug treatment, based on large-scale EEG data sets, two data-driven prediction models are proposed, and an effective way to improve the prediction performance is explored, which provides a useful attempt for the advancement of personalized precision treatment of depression.