Maturation of task-induced brain activation and long range functional connectivity in adolescence revealed by multivariate pattern classification

Abstract

The present study uses multivariate pattern classification analysis to examine maturation in task-induced brain activation and in functional connectivity during adolescence. The multivariate approach allowed accurate discrimination of adolescent boys of respectively 13, 17 and 21 years old based on brain activation during a gonogo task, whereas the univariate statistical analyses showed no or only very few, small age-related clusters. Developmental differences in task activation were spatially distributed throughout the brain, indicating differences in the responsiveness of a wide range of task-related and default mode regions. Moreover, these distributed age-distinctive patterns generalized from a simple gonogo task to a cognitively and motivationally very different gambling task, and vice versa. This suggests that functional brain maturation in adolescence is driven by common processes across cognitive tasks as opposed to task-specific processes. Although we confirmed previous reports of age-related differences in functional connectivity, particularly for long range connections (> 60 mm), these differences were not specific to brain regions that showed maturation of task-induced responsiveness. Together with the task-independency of brain activation maturation, this result suggests that brain connectivity changes in the course of adolescence affect brain functionality at a basic level. This basic change is manifest in a range of tasks, from the simplest gonogo task to a complex gambling task.

Introduction

Adolescence is a developmental period, between childhood and adulthood, of profound changes in body and behavior. These changes also include increased risk taking behavior, leading to rising numbers of accidents and early death, and an increased incidence of behavioral disorders in adolescence (Casey et al., 2008, Dahl, 2004, Steinberg, 2008). Although the neurobiological mechanisms underlying these changes are not yet well understood, it is likely that prolonged myelination and synaptic pruning play a role. Post-mortem studies have shown that these neural maturation processes continue well into young adulthood (Glantz et al., 2007, Huttenlocher and Dabholkar, 1997, Yakovlev and Lecours, 1967). Synaptic pruning is thought to underlie the decrease in gray matter density and cortical thickness observed in MRI studies, particularly in more complex association cortices (Giedd, 2004, Gogtay et al., 2004, Sowell et al., 2004). Myelination, on the other hand, is commonly associated with the linear increase in white matter volume seen in structural MRI studies (Giedd, 2004) and the protracted maturation of white matter bundles measured with diffusion weighted MRI (Lebel et al., 2008).

These anatomical changes are paralleled by wide-spread changes in functional brain organization. Brain-wide patterns of low-frequency temporal correlations in the fMRI signal during rest have shown a shift from childhood to adulthood in the strength of functional connectivity from short range toward long range connections (Fair et al., 2009, Kelly et al., 2009). Graph Theory analyses of whole brain networks have shown that network efficiency does not change with age, but that in childhood brain networks are organized more locally, based on anatomical proximity, whereas adult brain networks comprise regions spread over different brain lobes (Fair et al., 2009, Power et al., 2010, Vogel et al., 2010).

Many developmental fMRI studies have shown that task-related brain activity also undergoes changes during the transition from childhood to adulthood (for reviews, see Berl et al., 2006, Durston and Casey, 2006, Luna et al., 2010). The most consistent finding for cognitive tasks in the time window of adolescence is a maturation-related increase of BOLD response strength in the brain structures that are associated with task performance in adults (e.g. Crone et al., 2006, Keulers et al., 2011, Rubia et al., 2006). In other brain regions, hypothesized to be uncorrelated to task performance, the magnitude of brain activation tends to decrease or even disappear with age (e.g. Brown et al., 2005, Durston et al., 2006, Rubia et al., 2006). These age-related differences in response strength of particular brain areas during task execution are thought to be the consequence of brain-wide maturational changes in the way of interactions between neuron populations (e.g. Luna et al., 2010, Vogel et al., 2010).

A question fundamental to understanding functional brain maturation is then how the differences in task-induced activation relate to the changes in functional brain organization. It is well-established that there is a close correspondence between functional connectivity patterns, as evidenced from low-frequency BOLD fluctuations during rest as well as task performance, and the patterns of brain activity that emerge contingent upon task execution (e.g. Calhoun et al., 2008, Fox et al., 2006b, Smith et al., 2009). Specifically, the distributed functional organization of cortical areas that emerges during adolescent development largely overlaps with the clusters of commonly activated and deactivated brain areas that are consistently found during performance of a wide range of cognitive tasks (Cole and Schneider, 2007, Dosenbach et al., 2006, Dosenbach et al., 2007, Sridharan et al., 2008). Moreover, in adults there is a close relationship between the strength of functional connectivity between two brain sites during rest and their response strength modulation over trials during the execution of cognitive tasks (Mennes et al., 2010). Based on this close correspondence we hypothesized that particularly those regions that change their task-induced response strength with age will also show maturation of functional connectivity in the same age range.

This general relationship between changes in functional activation and in connectivity in the course of adolescent maturation may not hold for all task-induced activity, however. During adolescence, maturation is assumed to be associated with higher, more complex cognitive functions. Particularly well-studied are, for instance, response inhibition (e.g. Rubia et al., 2006, Rubia et al., 2007, Velanova et al., 2009), working memory ability (e.g. Crone et al., 2006, Geier et al., 2009) and risk taking behavior (e.g. Keulers et al., 2011, Steinberg et al., 2008, Van Leijenhorst et al., 2010), among others. For each of these abilities maturation may be associated with changes in specific brain structures involved in the ability and at a specific time window. In line with this, Cohen et al. (2010) were able to predict age from functional activation maps depicting the activation differences between successful stop trials and go trials, but not from the activation associated with go trials compared to rest. This suggests that maturation related differences in brain activity may be function specific.

In the present study, we investigated the patterns of maturation in both functional connectivity estimated from low-frequency BOLD signal fluctuations and functional activation related to task execution. Brain-wide maturation patterns will be quantified with multivariate pattern classification (MVP) tools (for reviews, see Haynes and Rees, 2006, Pereira et al., 2009). These tools use machine learning principles to find the pattern in a multidimensional space (e.g. the voxels of the images) that distinguishes two classes of examples. MVP analysis was introduced in the neuroimaging literature to “read” the voxel response patterns in visual cortex associated with particular stimulus categories (e.g. Cox and Savoy, 2003, Haxby et al., 2001), but has since then also been used across subjects, for instance, to find the distributed pattern that characterizes neuropathology (e.g. Fan et al., 2007, Sun et al., 2009). Here we will use MVP to classify functional brain maps according to age. Sensitivity of MVP to developmental patterns has recently been demonstrated (Cohen et al., 2010, Dosenbach et al., 2010). MVP analysis is generally more sensitive than the traditional univariate general linear model (GLM) analysis (e.g. De Martino et al., 2008). Moreover, MVP allows to directly test the generalizability of a distinctive pattern learned from a particular set of examples to a new set of example data, to establish whether they carry the same age-discriminative information. This feature allows us to directly test the functional specificity of maturation patterns by training a classifier on functional maps derived from a particular task and testing on data acquired during another task.

First of all, we established the sensitivity of the MVP classification to characterize maturation processes that take place during adolescence. We made use of brain activation data during the performance of a gonogo task, obtained from participants in three age groups, respectively 13, 17 and 21 years of age. The gonogo task focuses on neural processes related to the inhibition of a preponent response and has been frequently used in neurodevelopmental studies (e.g. Durston et al., 2006, Rubia et al., 2006, Tamm et al., 2002). In the task used here, participants had to press a response key whenever a letter was presented, but had to withhold their response when the infrequent nogo stimulus was presented (letter “A”). Next to the MVP classification analysis, we applied the traditional univariate voxel-wise analysis to establish age-related functional differences between the different age groups.

Secondly, we studied the functional specificity of the maturation patterns found in the task-related brain activity. The intrinsic generalization step of MVP classification allowed us to address this question in a direct way. Different cognitive tasks intend to steer specific neurocognitive processes. By using data from one particular task for training and data from a completely different task for testing, it can be directly investigated whether the pattern learned is associated with a specific cognitive function unique to the training task. Therefore, we also used fMRI data obtained from the same sample of participants while they performed a challenging gambling task (Keulers et al., 2011) and investigated whether the classifier trained on the data from the simple gonogo task was successful in classifying participants according to age from their functional activation maps during the complex gambling task, and vice versa.

Thirdly, we studied the relationship of task-related maturation patterns to the differences in short and long range low-frequency functional connectivities that have been described from childhood to early adulthood (Fair et al., 2007a, Fair et al., 2008, Fair et al., 2009, Kelly et al., 2009). We determined the strength of connectedness (Bullmore and Sporns, 2009) of gray matter voxels based on low frequency signal fluctuations in the gonogo data. Although low-frequency data obtained during task performance yield small differences in the distribution of functional connectivity compared to data obtained during rest, the overall lay-out of networks between the data is very similar (Dosenbach et al., 2010, Fair et al., 2007b). We investigated whether voxels that contributed to task-related age classification in the MVP differed from other voxels in their functional connectedness with the rest of the brain. In addition, in a whole brain exploratory analysis we entered the short and long range connectivity strength maps into a MVP classification analysis to allow a visualization of the maturation in functional connections and to compare these with maturational patterns in brain activity during task execution.