结构化张量稀疏表示在癫痫发作预测中的研究

Effective epileptic seizure prediction has great significance for controlling the epilepsy and alleviating the suffering of patients, also helps to reveal the pathogenesis of epilepsy and provides the foundation for the development of new treatments such as electrical stimulator. Since the artificial construction of EEG features is difficult and the conjoint analysis of multiple channels of EEG signals has become inevitable, the project will employ the theory of tensor and combine Holo-Hilbert spectrum analysis as well as sparse representation to build structure tensor sparse representation model, which is used for adaptive EEG signal classification and seizure prediction. Also, a sparse dictionary learning algorithm with discriminant constraint is proposed to increase the ability of identify and representation of the dictionary, which further raises the accuracy of the prediction performance. For the number of labeled EEG samples is too lack to meet the requirement of dictionary learning, the transductive classification is combined for designing the semi-supervised learning method to make use of unlabeled samples and improve the generalization ability of the system. This project will have important theoretical and clinical significance, which will promote the application of tensor in the field of EEG signal analysis, provide the novel ideas and methods for seizure prediction and help to improve the accuracy and robustness of the prediction system for clinical application.

有效的癫痫发作预测对于及时控制疾病，减轻患者痛苦具有十分重要的意义，也有助于揭示癫痫发作的内在机制，为研制新的癫痫治疗方法如脑部电刺激器等提供依据。针对人工挖掘脑电 (EEG) 信号特征的困难以及多导联EEG联合分析的必然趋势，本项目拟基于张量理论，结合全息希尔伯特谱分析与稀疏表示构建结构化张量稀疏表示模型，进行自适应EEG信号分类识别与癫痫发作预测；并提出具有判别约束性的稀疏编码字典学习方案，增加字典的辨识和表示能力，进一步提高预测的准确性；针对有标记EEG样本数量较少而不满足字典学习的问题，结合直推式算法设计EEG数据的半监督学习方案，充分利用无标记样本，提高系统的泛化能力。本项目的研究将推动张量在EEG信号分析领域的应用，为癫痫发作预测提供新的思路与方法，有助于提高预测系统的准确性与鲁棒性，具有重要的理论与临床应用意义。

准确有效的癫痫发作检测和预测技术对于及时控制癫痫疾病，减轻患者痛苦具有重要意义，也有助于探索新的癫痫诊疗方案。利用脑电信号(EEG)进行特征提取是发作检测和预测技术的核心。针对脑电信号提取特征的困难以及多导联脑电信号的联合分析，本项目提出了有效的脑电特征提取和分类算法进行癫痫发作的检测和预测。首先，基于张量分析理论，将改进的Stockwell变换与张量Tucker分解相结合，构建脑电信号的张量表示模型并提取核心张量特征，进行癫痫发作预测。该预测方法在21例患者的605小时的颅内脑电数据库上取得了97.62%的平均灵敏度与0.25/h的误判率，具有较好的预测准确率。同时，将距离测度引入脑电信号特征提取，提出基于张量分解的张量距离（TD）与基于小波变换的推土机距离（EMD）特征，进行癫痫发作检测。结果表明张量距离和推土机距离可以有效的度量不同时期脑电信号的差异性，具有较好的癫痫检测效果。此外，本项目引入深度学习网络，提出了基于双向LSTM深度网络和深层卷积神经网络的癫痫发作检测和预测算法，将脑电信号中隐藏的有效特征逐层自动提取，有效提高了检测和预测的系统性能。并且，在稀疏表示的基础上将核函数方法与概率协作表示相结合，进行癫痫脑电信号的分类与识别，取得了良好的识别效果。本项目为脑电信号的自动分析与处理提供了新的思路与方法，有助于推动癫痫发作检测和预测技术的性能提升，具有重要的理论价值与临床应用意义。