基于压缩感知机理的EEG信号癫痫波形自动检测与识别方法研究

The diagnosis and prevention of epilepsy is a challenging problem. The essential issue of this problem is the detection and recognition of characteristic waveform in electroencephalogram (EEG) signals. Based on the compressive sensing method ,this project will make an investigation on epilepsy characteristic waveform detection and recognition with the removal of artifacts. The creative contribution will be made as follows: 1) We will first construct a multi-morphological component dictionary for background wave, epilepsy intrinsic wave and artifact's wave, then in the morphological component sparse representation framework ,we will establish compressive sensing(CS) and reconstruction model for EEG. with the removal of artifacts. 2) Using the artifact removal CS model, we will propose a series of methods for epilepsy intrinsic wave pursuit and detection. 3) Three class features such as morphological component sparse representation coefficient, epilepsy intrinsic waveform atom and CS measurements will be integrated to be the distinguishing characteristics, and a two- stage feature selection and recognition theory will be proposed under semi-supervised learning and relevance feedback framework using clinical diagnosis and medical knowledge. Not only the new methods of epilepsy intrinsic wave detection and feature extraction will be proposed, but also the application of EEG analysis in the fields such as medical, military and physical training will be extensionally put forward.

癫痫病诊断和预防是一个富有挑战性的问题，脑电（EEG）的癫痫特征波形检测和识别是关键。本项目研究EEG信号去伪迹压缩感知的癫痫波形追踪检测和识别。创新性在于：通过EEG信号本源背景波形、非平稳的癫痫特征波以及伪迹等形态分量特征波形的形态核字典构造与形态分量稀疏表示机理研究，建立形态分量稀疏表示的EEG信号去伪迹压缩感知重构模型；研究去伪迹压缩重构EEG的癫痫特征波形追踪模型，提出一套EEG癫痫波形检测的新方法；结合形态分量稀疏表示系数、特征波形表示原子和压缩感知投影测量等三方面的特征信息，提出结合临床诊断和医学知识半监督学习下相关反馈的特征提取和识别两步方案。项目不仅提出一套EEG癫痫波形检测和特征提取的新方法，而且将推动EEG 信号分析在医学、军事、体育等领域的广阔应用前景。

EEG是一维的多通道非平稳(non-stationary)信号，它是临床脑疾病诊断、神经生理学、脑科学等研究的重要途径。由于EEG信号的非平稳性以及大量背景波形与伪迹的影响，对EEG信号进行自动分析是一个富有挑战性的问题。本课题利用EEG多通道信号间的时空关联性先验，着重研究EEG信号的自适应联合表示、压缩感知重建与检测分析等关键问题，提出了EEG多通道信号联合表示的图结构稀疏表示与结构化低秩表示模型，继而构建了图稀疏正则的压缩感知重建、联合稀疏与低秩先验的压缩感知重建等算法。并建立了群稀疏表示的鉴别字典学习模型，建立了基于ICA的EEG盲源分离算法，可有效检测与判别癫痫特征波形，分析局部场电位和具体行为关系。进一步将相关理论方法推广应用于视视频、高光谱等具有类似多通道结构的数据。圆满完成了项目计划书规定的研究任务，发表高质量论文数和专利数均满足验收指标。本课题取得的研究成果为癫痈特征波的临床检测提供了技术支撑，在脑电监护、癫痫等脑疾病的控制与治疗上具有重要意义。