事件相关电位的提取和效应连接分析方法研究

发布时间：2018-03-22 07:18

本文选题：事件相关电位　切入点：稀疏学习　出处：《清华大学》2016年博士论文　论文类型：学位论文

【摘要】：事件相关电位(event related potential,ERP)是由特定感官刺激、认知任务或者运动事件等诱发的典型电生理响应,是认知科学、心理学等领域中常用到一种脑电组分。近年来,随着对于大脑功能研究的深入,人们认识到不同大脑区域间动态的连接对解释大脑工作机制具有重要意义。因此,许多与ERP相关的研究不仅要求测量实验效应引起的幅度差异、潜伏期差异等,也要求评价实验过程中连接性的变化。本文研究了ERP的提取和效应连接分析方法,主要包括三方面内容:单一实验条件下ERP信号的建模和提取方法、多实验条件下ERP联合建模和提取方法、基于时变多元自回归模型的效应连接分析方法。单一实验条件下ERP建模方法利用了ERP信号的相位锁定特性。ERP信号按试次拼接后可以用一组与波形无关的谐波表示。因此多通道ERP信号就可以由一个空间-频率矩阵来表示。针对该模型,本文提出了ERP空间-频率模式估计算法。该算法可以循环更新求解ERP空间-频率矩阵和噪声协方差矩阵,并根据最大化ERP信噪比的原则对空间-频率矩阵进行进行分解,获得ERP信号的空间模式、频率模式和对应的空间滤波器。该方法可以自动确定ERP成分个数,防止模型过拟合。多实验条件下ERP信号联合建模方法假设同一受试在不同实验条件下同一来源的ERP成分具有相同的空间模式,但是其幅度、潜伏期可能存在差异。为了求解该模型中各参数的后验分布,本文提出了贝叶斯ERP成分估计算法。该方法利用变分贝叶斯原理,迭代更新ERP的空间模式、时域波形、幅度因子、潜伏期差异和噪声协方差矩阵的后验分布。该方法对于噪声协方差矩阵形式没有特定要求,ERP成分个数也能够自动确定,具有较高的实用价值。时变效应连接分析关键在于准确估计时变多元自回归模型。本文针对ERP信号特点提出了一种时变多元自回归模型,它假设模型系数的一阶差分值是稀疏的。根据变分贝叶斯原理设计的算法能够利用数据自动学习得到模型参数并准确地估计出多元自回归模型。模型阶数也可以根据变分下界确定。该方法中关键参数可以自动确定,因此具有很好的应用价值。本文提出的三种算法可以实现ERP信号的有效提取和时变效应连接的准确估计,为ERP在认知神经科学、心理学等领域更广泛地应用提供了新的工具。
[Abstract]:Event related potentialERP is a typical electrophysiological response induced by specific sensory stimuli, cognitive tasks or motor events. It is often used in cognitive science, psychology and other fields. With the development of the study of brain function, it is recognized that dynamic connections between different brain regions are of great significance in explaining the mechanism of brain work. Therefore, many studies related to ERP not only require measuring the amplitude differences caused by experimental effects. The difference of latency also requires evaluating the changes of connectivity in the course of experiment. This paper studies the methods of ERP extraction and effect connection analysis, which mainly include three aspects: modeling and extracting method of ERP signal under a single experimental condition. The joint modeling and extraction method of ERP under multiple experimental conditions, The method of effect connection analysis based on time-varying multivariate autoregressive model. Under a single experimental condition, the ERP modeling method utilizes the phase locking characteristic of ERP signal. The signal can be represented by a set of wave independent harmonics after splicing according to the test times. So the multichannel ERP signal can be represented by a space-frequency matrix. In this paper, a ERP space-frequency pattern estimation algorithm is proposed, which can be used to iteratively solve the ERP space-frequency matrix and noise covariance matrix, and decompose the space-frequency matrix according to the principle of maximizing the ERP signal-to-noise ratio. The spatial mode, frequency mode and corresponding spatial filter of ERP signal are obtained. This method can automatically determine the number of ERP components. The joint modeling method of ERP signals assumes that the same subjects have the same spatial pattern of ERP components from the same source under different experimental conditions, but their amplitude is the same. In order to solve the posterior distribution of each parameter in the model, a Bayesian ERP component estimation algorithm is proposed in this paper. The method iteratively updates the spatial mode, time-domain waveform, amplitude factor of ERP by using variational Bayesian principle. The latent period difference and the posterior distribution of noise covariance matrix. This method can determine the number of ERP components automatically without any specific requirements for the form of noise covariance matrix. This paper presents a time-varying multivariate autoregressive model according to the characteristics of ERP signals, and the key point of time-varying effect connection analysis lies in the accurate estimation of time-varying multivariate autoregressive model. It assumes that the first order difference value of the model coefficient is sparse. The algorithm designed according to variational Bayesian principle can automatically learn the model parameters and estimate the multivariate autoregressive model accurately. The model order can also be estimated by the model order. The key parameters in the method can be determined automatically according to the lower bound of the variation. The three algorithms proposed in this paper can effectively extract ERP signals and accurately estimate the time-varying effects, and provide a new tool for the wider application of ERP in cognitive neuroscience and psychology.
【学位授予单位】：清华大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：R318

【相似文献】