Predictive modelling of neuroimaging measures of language processing

随着神经成像研究的发展，我们开始从研究大脑活动的平均模式转向了解个体间差异的来源。在过去的授课期间，我和我的受训者已经确定了词汇和语法加工的神经特征(N400和P600，使用脑电和脑磁图测量)的个体差异与语言、认知和人口统计变量之间的关系。我们的发现表明，头皮记录的大脑活动对细微的个体差异比之前认为的更敏感，包括社会经济地位和阅读习惯等因素。在我们工作的第二个相关主题中，我们已经开始记录参与自由对话的人的大脑活动(而不是更典型的任务，如接受语言处理，或简单、受控地产生单个单词)。在验证了这种方法之后，我们准备进入神经成像研究的新时代，无论是在更自然的条件下研究语言，还是开发包含自然语言数据和相关大脑活动的数据库(语料库)都大有可为。这项工作需要开发和应用新的统计方法来分析复杂的、4维的神经成像数据，并将其与包括机器学习技术在内的大量预测变量相关联。 我们在下一个授权期的主要工作目标将是开发语言处理过程中神经反应的预测模型，使我们能够根据客观的语言、认知和人口统计指标以及之前的语言环境准确地预测大脑活动模式。因此，我们不是假设每个人都会以相同的大脑活动模式做出反应，而是试图建模并因此理解围绕平均活动模式驱动个体差异的因素。这有望更丰富地了解语言是如何在大脑中进行的，并最终通过我们与行业合作的研发努力，获得指导个性化培训(例如，第二语言学习)和治疗(例如，言语语言治疗)的见解。

As neuroimaging research evolves, we are beginning to move from the study of average patterns of brain activity to understanding the sources of inter-individual variance. Over the past granting period my trainees and I have identified relationships between individual differences in neural signatures of lexical and grammatical processing (the N400 and P600, measured using EEG and MEG) and linguistic, cognitive, and demographic variables. Our findings indicate that scalp-recorded brain activity is much more sensitive to fine-grained individual differences than previously thoughtincluding factors such as socioeconomic status and reading habits. In a second, related theme of our work, we have begun to record brain activity from people engaged in free conversation (rather than more typical tasks such as during receptive language processing, or simple, controlled production of single words). Having validated this approach, we are poised to enter a new era of neuroimaging research that holds great promise both in terms of studying language under more naturalistic conditions, and developing a database (corpus) containing both natural language data and associated brain activity. This work has necessitated the development and application of novel statistical approaches to analyzing complex, 4-dimensional neuroimaging data and relating it to a large number of predictive variables, including machine learning techniques. The overarching goal of our work in the next granting period will be to develop predictive models of neural responses during language processing, that allow us to accurately predict brain activity patterns based on a combination of objective linguistic, cognitive, and demographic measures, and preceding language context. Thus rather than assuming everyone will respond with the same pattern of brain activity, we attempt to model and therefore understand the factors that drive individual differences around the average pattern of activity. This holds the promise of a much richer understanding of how language is proceed in the brain, and ultimately insights that will guide individualized training (e.g., second language learning) and therapy (e.g., speech-language therapy) through our industrially-partnered R&D efforts.