Liu Xinyu, Chen Bingxu, Zhang Haoran, Cui Yi, Wang Bin, Wang Yingtan, Zhao Tong, Yan Yuxiang, Sha Sha, Ren Yanping, Zhang Ling, Zhao Xixi, Wang Gang
Beijing Key Laboratory of Mental Disorders, National Clinical Research Center for Mental Disorders & National Center for Mental Disorders, Beijing Anding Hospital, Capital Medical University, Beijing, China; Advanced Innovation Center for Human Brain Protection, Capital Medical University, Beijing, China.
Faculty of Information Technology, Beijing University of Technology, Beijing, China.
J Affect Disord. 2025 Nov 15;389:119599. doi: 10.1016/j.jad.2025.119599. Epub 2025 Jun 10.
To enhance the differentiation between unipolar depression (UPD) and bipolar depression (BPD), this study integrates machine learning and deep learning models with electroencephalography (EEG) data and clinical features. Utilizing Python for data preprocessing and feature extraction, we analyzed 370 patients diagnosed with either UPD or BPD. The experimental design featured 5-fold cross-validation and leave-one-out cross-validation to assess model generalization and optimize hyperparameters through grid search. Models included Support Vector Machine, Random Forest, and neural networks such as Fully Connected Neural Network(FCNN), Recurrent Neural Network, Long Short-Term Memory networks, and Transformers. Evaluation metrics highlighted FCNN's superior performance with 76 % accuracy, 80 % sensitivity, 73 % specificity, and 76 % F1-score. Results underscore the importance of EEG biomarkers, particularly beta band activity, in distinguishing between the two conditions. The study demonstrates the potential of deep learning in identifying complex mental disorder patterns and advocates for a shift towards data-driven diagnostics in mood disorders. Future research should aim to enhance model interpretability, integrate multimodal data, and develop advanced feature extraction techniques to further precision psychiatry.
为了增强单相抑郁（UPD）和双相抑郁（BPD）之间的鉴别，本研究将机器学习和深度学习模型与脑电图（EEG）数据及临床特征相结合。利用Python进行数据预处理和特征提取，我们分析了370例被诊断为UPD或BPD的患者。实验设计采用5折交叉验证和留一法交叉验证，以评估模型的泛化能力，并通过网格搜索优化超参数。模型包括支持向量机、随机森林以及神经网络，如全连接神经网络（FCNN）、循环神经网络、长短期记忆网络和Transformer。评估指标显示，FCNN表现出色，准确率为76%，灵敏度为80%，特异性为73%，F1分数为76%。结果强调了EEG生物标志物，特别是β波段活动，在区分这两种情况中的重要性。该研究证明了深度学习在识别复杂精神障碍模式方面的潜力，并主张在情绪障碍中转向数据驱动的诊断。未来的研究应致力于提高模型的可解释性，整合多模态数据，并开发先进的特征提取技术，以进一步推动精准精神病学的发展。
