【作者】 吴畏；

【导师】 高上凯；

【作者基本信息】 清华大学 ， 生物医学工程， 2012， 博士

【摘要】 脑电是记录脑功能活动的重要工具，它适合用于在大空间尺度下研究脑功能的机制。为有效提取多导联脑电信号中丰富的时间和空间信息，本论文在统计建模的框架下对脑电信号的时空建模方法进行了深入系统的研究。脑电节律信号和事件相关电位（event-related potential，ERP）是两大类具有不同特点的脑电信号，本文采用不同的方案对它们建立时空模型。对于脑电节律信号，本文提出了一种分层贝叶斯模型来对其分实验条件和实验次数进行时空建模，并通过模型约束实现了不同实验条件和实验次数下数据信息的共享。用于模型参数估计的变分贝叶斯算法可通过稀疏贝叶斯学习方法自动推断模型中成分信号的个数，避免了模型对数据的过拟合。由于ERP在单次实验中信噪比很低，为对其增强，本文提出混合效应模型分别对ERP成分和背景脑电成分进行建模。此外，模型还对ERP成分在不同实验下的幅度和不应期建模，以揭示ERP成分随实验次数变化的动态过程。由于该模型提供了更多的ERP特征信息，具有很好的实用价值。在以上时空模型的基础上，本文提出了三种空域滤波器学习算法用于脑-机接口系统中进行特征提取。其中，OVR-CSP和R-CSP算法推广了经典的共空域模式算法，使其适用于对多类别以或包含野点的脑电数据进行分类；SIM算法通过极大化ERP成分的信噪比，有效增强了ERP成分。这些算法可以较大程度上提高脑-机接口性能，具有较好的应用前景。作为对时空建模方法的补充，本文还提出了一种联合学习空域滤波器、频域滤波器和分类器的算法ISSPL用于对脑-机接口想象运动脑电的分类。ISSPL算法采用最大间隔学习的思想对频域滤波器和分类器同时进行优化，保证了算法的泛化能力，避免了频域滤波器谱系数的高维度带来的过拟合问题。本文通过仿真实验数据和真实脑电数据分析，验证了以上各模型的合理性和各算法的有效性。与同类算法相比，本文提出的各算法在数据分析时均表现出更优越的性能。更多还原

【Abstract】 Electroencephalography (EEG) is an important tool for recordingfunctional brain activity. It is suited for investigating the mechanism ofbrain functions at large spatial scales. For the purpose of effectivelyextracting rich spatial and temporal information from multichannel EEGsignals, this thesis undertakes a comprehensive and in-depth study ofspatio-temporal modeling approaches for EEG within a statistical modelingframework.Oscillatory signals and event-related potentials (ERPs) are two generalcategories of EEG signals. According to their distinct characteristics, thisthesis employs different strategies to establish their respectivespatio-temporal models.For oscillatory signals, a hierarchical Bayesian model is introduced forlearning their spatio-temporal patterns in a condition-and trial-wise fashion.Data information within multiple trials and conditions are allowed to beshared among one another via proper model constraints. The variationalBayes algorithm, which is developed for model inference, is capable ofautomatically inferring the number of components in the model via sparseBayesian learning, avoiding potential overfit to the data.For ERP signals, since their signal-to-noise ratio is rather low in a single trial, to enhance its estimation a mixed-effect statistical model isproposed to model ERP components and background EEG components in aseparate manner. Moreover, to reveal ERP dynamics across trials, theinter-trial amplitude and latency variability of the ERP components is alsoexplicitly parameterized in the model. The fact that the model offers moreinformation regarding ERPs suggests its high value for practicalapplications.Based on the above proposed spatio-temporal models, three spatialfiltering learning algorithms are designed for feature extraction inbrain-computer interfaces (BCIs). Specifically, the OVR-CSP and R-CSPalgorithms generalize the classic common spatial patterns (CSP) algorithm,extending its use to situations where there are multiple classes or there areoutliers in the data. The SIM algorithm enhances the ERP components bymaximizing their signal-to-noise ratio. All these algorithms cansignificantly improve the performance of current BCI systems, indicatingtheir high potential for a broad range of applications.As a supplement to the spatio-temporal modeling framework, thisthesis also develops an algorithm, termed ISSPL, for joint spatial filtering,spectral filtering, and classification of motor imagery EEG data. In theISSPL algorithm, spectral filters are optimized in conjunction with theclassifier via maximal margin learning, overcoming the potentialoverfitting issue due to the high-dimensionality of the spectral coefficients. The validity of all the proposed models is discussed at full length. Theeffectiveness of all the algorithms is shown through the analysis ofsimulated data and real EEG recordings. The results demonstrate that, incomparison with other contemporary algorithms, the algorithms describedin this thesis yield superior performance in all cases.更多还原