用于癫痫发作预测的脑电特征深度学习研究

The accurate prediction of epileptic seizure can provide the intractable epileptic patients with the effective treatment method and is the key link of implantable closed-loop epilepsy stimulators. As the core of seizure prediction technology, the feature extraction of EEG signals has been the bottleneck which blocks performance improvement of prediction technology. For the difficulties of the artificial construction of EEG features, the project will employ the deep learning theory to build the deep learning network model via combining extreme learning machine (ELM) and sparse coding to establish the extensible way to automatically learn EEG features, and the structural coding dictionary is designed to further improve the identification ability of EEG features. For the actual situation of the shortage of the labeled EEG samples, the project will make the best of unlabeled EEG samples, and develop semi-supervised learning method of the deep learning network utilizing the advantage of ELM to improve its generalization performance. Additionally, the project will propose the algorithm of metric distance of EEG features through shrinkage expansion adaptive metric learning, and offer the risk value of impending seizure using the metric distances of the EEG features. The launch of this project can provide novel ideas and methods for EEG feature extraction, and promote the development of seizure prediction and the application of deep learning in EEG analysis field. There are important theory significance and clinical application prospect for this project.

对癫痫发作的准确预测可以为顽固性癫痫患者提供有效的治疗方法，也是植入式闭环癫痫刺激器的关键环节。而作为发作预测技术的核心，脑电(EEG)信号的特征提取已成为阻碍预测技术性能提高的瓶颈。针对人工发掘EEG特征的困难，本项目拟基于深度学习理论，结合极端学习机(ELM)与稀疏编码构建EEG特征的深度学习网络模型，建立可扩展的EEG特征全自动学习途径，并通过构造结构型编码字典进一步提升EEG特征的辨识能力；对于有标记EEG样本少的临床现实，本项目拟充分利用无标记EEG样本，借助ELM的优势设计深度网络的半监督学习方法，提高其泛化性能；另外，拟利用自适应收缩膨胀测度学习方法获取EEG特征的测度距离算法，并通过计算测度距离给出癫痫预发作的风险值。本项目的开展将为EEG特征提取提供新思路与新方法，有助于推进发作预测技术的性能提升，也将推动深度学习在EEG分析领域的应用，具有重要的理论意义及临床应用前景。

准确预测癫痫发作可以使医生或患者提前采取有效的临床预防措施，对于建立新的治疗方法和改善患者的生活质量至关重要。脑电信号的特征提取作为发作预测技术的核心，是显著提高预测技术性能的关键。针对人工发掘脑电特征的困难，本项目依据深度学习理论，提出了基于深层卷积神经网络的癫痫发作预测方法，实现了癫痫患者脑电特征的自动学习，在21例癫痫患者的573.07个小时的连续脑电记录上，可以达到87.70%的敏感度和0.25/h的误判率。同时，考虑到癫痫患者脑电数据分布的不平衡，本项目分别提出了基于加权极端学习机与基于EasyEnsemble学习算法的不平衡脑电分类模型。实验表明这两个模型显著提高了癫痫发作检测的敏感度。之后，本项目又探索了脑电纹理特征在癫痫识别中的作用，通过实验验证了脑电纹理特征可以作为检测癫痫发作的指标依据。本项目的研究为脑电信号分析提供了新思路与新方法，有助于推进发作预测技术的性能提升，也将推动深度学习在脑电分析领域的应用，具有重要的理论意义及临床应用前景。