Age-related connectivity changes in fMRI data from children listening to stories

Abstract

The way humans comprehend narrative speech plays an important part in human development and experience. A group of 313 children with ages 5–18 were subjected to a large-scale functional magnetic resonance imaging (fMRI) study in order to investigate the neural correlates of auditory narrative comprehension. The results were analyzed to investigate the age-related brain activity changes involved in the narrative language comprehension circuitry. We found age-related differences in brain activity which may either reflect changes in local neuroplasticity (of the regions involved) in the developing brain or a more global transformation of brain activity related to neuroplasticity. To investigate this issue, Structural Equation Modeling (SEM) was applied to the results obtained from a group independent component analysis (Schmithorst, V.J., Holland, S.K., et al., 2005. Cognitive modules utilized for narrative comprehension in children: a functional magnetic resonance imaging study. NeuroImage) and the age-related differences were examined in terms of changes in path coefficients between brain regions. The group Independent Component Analysis (ICA) had identified five bilateral task-related components comprising the primary auditory cortex, the mid-superior temporal gyrus, the most posterior aspect of the superior temporal gyrus, the hippocampus, the angular gyrus and the medial aspect of the parietal lobule (precuneus/posterior cingulate). Furthermore, a left-lateralized network (sixth component) was also identified comprising the inferior frontal gyrus (including Broca's area), the inferior parietal lobule, and the medial temporal gyrus. The components (brain regions) for the SEM were identified based on the ICA maps and the results are discussed in light of recent neuroimaging studies corroborating the functional segregation of Broca's and Wernicke's areas and the important role played by the right hemisphere in narrative comprehension. The classical Wernicke–Geschwind (WG) model for speech processing is expanded to a two-route model involving a direct route between Broca's and Wernicke's area and an indirect route involving the parietal lobe.

Introduction

Story comprehension skills are an important and integral part of human experience. Annual school performance scores have also shown that story comprehension ability has a profound impact on early childhood development (Lorch et al., 1998). Human communication often involves story-like constructs making structured narration a part of human experience. This component of human experience, the ability to comprehend narration, may also influence the interpersonal as well as the intrapersonal development in children. Comprehension of fictional narratives can be thought of as a series of events that unfold over time without violating causality (Graesser et al., 1980, Barthes, 1981). The rules of cause and effect demand that the goals and the nature of the characters in the stories be integrated in a logically coherent manner (Trabasso and Stein, 1997, Van den Broek, 1997). Thus, the comprehension of narratives is not simply the comprehension of individual sentences in the story but rather a complicated mechanism involving higher-order cognitive processes.

Functional brain imaging methods have recently emerged as a means to investigate connectional anatomy and the dynamic flow of information across neural networks (McIntosh and Gonzales-Lima, 1994, McIntosh et al., 1994). The two basic functional brain imaging methods that measure the neural activity in terms of either electrical/magnetic fields generated (EEG/MEG) or the hemodynamic response of tissue (fMRI), have revealed functional networks involved in different cognitive tasks. However, the conventional fMRI data analysis methods have generally focused on identifying areas of activation under different behavioral conditions without paying much attention to the underlying network characteristics. Activity in one brain region may be involved in more than one functional network and the traditional analysis methods are incapable of separating this within a ‘neural context’. Although the traditional correlation analysis is capable of determining the degree to which two brain regions co-vary, no causal information can be derived from such an analysis. This traditional correlation analysis does not involve directionality and is identified with functional connectivity (FC) in the neuroimaging literature. Furthermore, when more than two brain regions are involved, correlation between two regions may involve direct as well as indirect effects. Indirect effects are due to one brain region influencing another via a third region. Such indirect effects cannot be separated out by a traditional correlation analysis. Similarly, neural interactions also involve many brain regions simultaneously limiting the effectiveness of traditional correlation analysis. Thus, the organizational principles of the nervous system require an advance modeling technique to investigate the neural interactions.

Structural Equation Modeling (SEM) or path analysis is a widely used statistical method primarily used to test hypothesis about causal influences among measured or latent variables. SEM is capable of testing a variety of theoretical models that hypothesize how sets of variables define constructs and how these constructs are related to each other. Unlike functional connectivity, SEM deals with effective connectivity meaning a causal relationship (directionality) can be ascribed to a path coefficient. It utilizes various types of models to capture relationships among variables with the basic goal of providing a quantitative test for a theoretical model hypothesized by a researcher. SEM approaches data from a different perspective than most commonly used statistical methods for modeling individual observations. The emphasis is on the whole variance–covariance data structure rather than individual observations. The parameter estimation in SEM involves minimizing the difference between the observed variance–covariance structure and the one predicted by the implied model. The basic models in SEM consist of regression, path and confirmatory factor models that have been extensively used in psychology, economics and other social sciences. Path coefficients in SEM represent the proportion of activity in one area determined by the other area. SEM also allows for influences, not measured or measurable to be incorporated in a model as residuals. This makes SEM a desirable modeling approach to handle neuronal interactions. A residual describes the combined influence on one region by the regions not included in a model combined with the influence of the brain region upon itself.

To investigate the neural correlates of auditory narrative comprehension in children, we performed a large-scale fMRI study involving 313 children ages 5–18. Simple short stories were presented to the subjects. The simplicity obviated the necessity for higher order integrative processes (e.g., between groups of events) and made the difficulty level suitable for children as young as 5 years. In general, story comprehension appears to recruit networks involving frontal, temporal and cingulate areas that support working language and memory processes. More recently, group Independent Component Analysis (ICA) has been used to identify modes describing activity in a sparsely distributed network (McKeown et al., 1998, Calhoun et al., 2001, Schmithorst and Holland, 2004). ICA is a data-driven method that makes no assumptions about the underlying biology. A number of studies have already been carried out to investigate math processing (Schmithorst and Brown, 2004), alcohol intoxication effects on simulated driving (Calhoun et al., 2004), and music perception (Schmithorst, 2005) using group ICA analysis. The main advantage of this method is that it does not require a prior knowledge of the Hemodynamic Response Function (HRF) which may be allowed to vary across subjects. Significant variance in HRF across subjects may impact the investigation of macro-level processes considerably and may not correlate well with the typical regressor used in the General Linear Model (GLM) (e.g., a square wave convolved with an impulse HRF) (Worsley and Friston, 1995). For the task of auditory narrative processing, in addition to the activation in primary auditory cortex, the group ICA has detected activation in superior temporal gyrus, posterior cingulate gyrus, the hippocampus and the inferior parietal lobules consistent with previous studies (Mazoyer et al., 1993, Fletcher et al., 1995, Gallagher et al., 2000, Schmithorst et al., 2005). This is in addition to the activation in perisylvian language areas recruited for semantic and syntactic processing, namely the angular gyrus and the inferior frontal gyrus (Holland et al., 2001, Fiebach et al., 2005). The present study addresses the related issue of developmental changes among neural networks involved in narrative auditory processing in children. The possibility of combining SEM with group ICA to investigate age-related effective connectivity changes among regions in the developing brain is a novel extension of cognitive neuronal modeling using SEM. This approach may also have clinical relevance in predicting connection deficits associated with language pathologies in individual children relative to a normative reference group.