Increasing evidence suggests that brain variability plays a number of important functional roles for neural systems. However, the relationship between brain variability and changing cognitive demands remains understudied. In the current study, we demonstrate experimental condition-based modulation in brain variability using functional magnetic resonance imaging. Within a sample of healthy younger and older adults, we found that blood oxygen level–dependent signal variability was an effective discriminator between fixation and external cognitive demand. Across a number of regions, brain variability increased broadly on task compared with fixation, particularly in younger and faster performing adults. Conversely, older and slower performing adults exhibited fewer changes in brain variability within and across experimental conditions and brain regions, indicating a reduction in variability-based neural specificity. Increases in brain variability on task may represent a more complex neural system capable of greater dynamic range between brain states, as well as an enhanced ability to efficiently process varying and unexpected external stimuli. The current results help establish the developmental and performance correlates of state-to-state brain variability-based transitions and offer a new line of inquiry in the study of rest versus task modes in the human brain.
Increasing evidence suggests that brain variability is a useful attribute for neural systems, perhaps reflecting greater network complexity, connectivity, increased dynamic range, improved signal detection thresholds, and superior cognitive processing (Li et al. 2006; Faisal et al. 2008; McIntosh et al. 2008, 2010; Shew et al. 2009; Garrett et al. 2010; Deco et al. 2011; Garrett et al. 2011; Misic et al. 2011; Vakorin et al. 2011). It has been argued that, as a nonlinear dynamical system, the brain functions at the “edge of criticality” between a variety of possible “states” or functional network configurations (Ghosh et al. 2008; Deco et al. 2009, 2011; McIntosh et al. 2010). When variability is too low, there is little capacity for the system to explore these states, yielding the potential for the system to remain rigidly in a single state. With relatively greater variability, and/or upon dynamic fluctuations in variability, the brain is more capable of modulating from state-to-state. It is poorly understood, though, whether brain variability adjusts from one specific brain state to another. A recent magnetoencephalography (MEG) study of children and young adults noted that brain variability was generally greater when viewing upright versus inverted faces (Misic et al. 2010). However, it remains unknown how particular state-to-state transitions might relate to older age and cognitive performance. Our recent work using functional magnetic resonance imaging (fMRI; Garrett et al. 2010, 2011) suggests that in a variety of experimental conditions, older and poorer performing adults exhibit less variable brain activity overall and more dedifferentiation (i.e., more similarity) in signal variability across brain regions relative to younger and better performing adults. If this variability is important for state modulations (Ghosh et al. 2008; Deco et al. 2009, 2011; McIntosh et al. 2010), older and poorer performing adults may also exhibit more subtle state-to-state modulations in signal variability.
Testing this possibility prompts the question of which primary brain/cognitive state transitions to first consider in relation to age and performance. Interestingly, some work suggests that brain variability is the basis for the probabilistic nature of the brain (Knill and Pouget 2004; Ma et al. 2006; Beck et al. 2008), whereby neurons may utilize a Bayesian process that generates optimal responses in the face of external stimuli of varying reliabilities. Essentially, the authors argued that neural variability yields adaptability in the presence of stimulus uncertainty in one's environment. One viable hypothesis, then, is that brain variability would be greater upon external cognitive demand (i.e., greater stimulus uncertainty) relative to conditions in which no changing external stimulus exists (i.e., less stimulus uncertainty), but that this transition may be more subtle in older poorer performing adults due to a relative lack of neural flexibility and adaptability (Garrett et al. 2010, 2011; McIntosh et al. 2010). Accordingly, a natural and highly reliable state-to-state transition to examine is a contrast of rest (internally driven cognition or “default mode”) and task states (externally focused cognition). Using mean signal-based measures, such contrasts are among the most reliable findings, both statistically and topographically, in functional neuroimaging (Damoiseaux et al. 2006; Toro et al. 2008; Spreng et al. 2009; Andrews-Hanna et al. 2010; Grady et al. 2010) and are known to change with older age (e.g., reduced modulation of the default mode during tasks—Lustig et al. 2003; Grady et al. 2006; Persson et al. 2007; expansion of task-positive network—Grady et al. 2010; degradation of functional connectivity between key network nodes—Andrews-Hanna et al. 2007; Grady et al. 2010; Park et al. 2010; Sambataro et al. 2010; correlations between default or task network integrity and behavioral measures in older adults—Andrews-Hanna et al. 2010; Grady et al. 2010). The contrast of rest (primarily involving medial prefrontal cortex, posterior cingulate/precuneus, and lateral parietal regions) and task states (lateral inferior frontal and parietal areas, dorsolateral prefrontal cortex, and sensorimotor regions) results in distinct patterns of activity for each state, and these patterns of activity are often anticorrelated (Fox et al. 2005; Toro et al. 2008; Spreng et al. 2009). However, it is not known whether: 1) fMRI-based brain variability also differs between rest and task states; 2) the same default and task-positive regions modulate in variability as they do with mean signal; and 3) how this varies with age and performance.
Interestingly, our previous work suggests near orthogonality between mean- and variability-based spatial patterns (Garrett et al. 2010, 2011); we would thus not expect modulations between rest and task modes to occur in anticorrelated default and task-positive regions. Rather, there may be a unidirectional increase in brain variability upon external cognitive demand (task) to handle increased stimulus uncertainty (Knill and Pouget 2004; Ma et al. 2006; Beck et al. 2008). Older and poorer performing adults may be less able to increase their brain variability levels on task to flexibly adjust to stimulus uncertainty, either because their brain variability levels are too low or too dedifferentiated to permit the occurrence of state transitions (McIntosh et al. 2010; Garrett et al. 2011). Furthermore, if older poorer performing adults can still modulate variability levels to some extent (even if less than younger adults), this may not occur in the same regions as in young high performers. Although older poorer performers are typically less variable, our previous work indicates a small proportion of regions in which this group expresses greater variability than do younger higher performers (Garrett et al. 2010, 2011), indicating an age-based bidirectional signal variability effect. A similar type of age effect may be found in relation to age group differences in internal-to-external state modulations.
Here, we expand our recent work on fMRI-based brain variability (Garrett et al. 2010, 2011) to examine 2 primary hypotheses with regard to brain state modulations: 1) brain variability will increase upon external cognitive demand (Knill and Pouget 2004; Ma et al. 2006; Beck et al. 2008), but not necessarily in regions that typify default and task-positive networks when mean activity changes are assessed and 2) older and poorer performing adults will exhibit smaller internal-to-external state transitions in brain variability, as well as less signal variability overall (Garrett et al. 2010, 2011). We assessed these hypotheses using a series of multivariate models assessed within a resampling framework, and subsequently provide evidence for the split-half reliability of our brain variability data.
Materials and Methods
We initially examined 18 young adults (mean age = 25.79 ± 3.28 years, range 20–30 years, 10 women) and 27 older adults (mean age = 66.46 ± 8.25 years, range 56–85 years, 14 women). Upon analysis of our full sample Partial Least Squares (PLS) model (see Results, first section), we discovered one older adult outlier who was more than 3 standard deviations (SDs) above the older adult mean on every brain score measure; this participant was removed from the sample. All models and results in the current paper thus reflect a total N of 44 (18 young and 26 old). Most participants were right handed (3 in each group were left handed), and all were screened using a detailed health questionnaire to exclude health problems and/or medications that might affect cognitive function and brain activity (e.g., stroke and cardiovascular disease). Structural MRIs were inspected to rule out severe white matter (WM) changes or other abnormalities. There was no relation between age and performance on the mini-mental state examination (Folstein et al. 1975). The present experiment was approved by the Research Ethics Board at Baycrest, and all participants gave informed consent (following the guidelines of the Research Ethics Board at Baycrest and the University of Toronto) and were paid for their participation.
Data of Interest: Fixation and Task Blocks
All fMRI analyses were performed using volumes acquired during fixation and task blocks in a block design (Grady et al. 2010). Visual stimuli were band-pass filtered white noise patches with different center frequencies. We analyzed 4 scanning conditions: 1) fixation (Fix), 2) perceptual matching (PMT), 3) attentional cueing (ATT), and 4) delayed match to sample (DMS). Fixation blocks were interspersed between each task block (e.g., fixation–task–fixation–task, etc.); participants viewed a fixation cross for the entirety of each fixation block. On average, there were 32 fixation blocks per subject, with 10 volumes (20 s) per block. For each task, there were 8 blocks (4 runs, 2 blocks per run) presented in random order; in all tasks, the intertrial interval was 2000 ms. For PMT, a sample stimulus appeared in the upper center portion of the screen along with 3 choice stimuli located in the lower part of the screen (for 4000 ms). Participants were asked to indicate which of the 3 choice stimuli matched the sample stimulus. Six of these trials occurred in each PMT block (total block length [including intertrial intervals of 2000 ms] = 36 s). For ATT, a stimulus appeared for 1500 ms in the upper central part of the screen. Then, an arrow pointing either to the right or to the left appeared (in the lower part of the screen) with the sample stimulus for 1500 ms. The arrow was removed and 500 ms later, 2 stimuli appeared in the right and left locations for 3000 ms. Participants were asked to attend only to the location that had been cued by the arrow and press 1 of 2 buttons to indicate whether or not the cued target stimulus matched the sample. There were 4 trials in each ATT block (total block length = 34 s). Finally, for the DMS task, a sample stimulus was presented for 1500 ms in the upper central portion of the screen followed by a delay of 2500 ms (blank screen). Three choice stimuli were then presented for 3000 ms in the lower portion of the screen and the participants had to press 1 of 3 buttons to indicate which of the 3 stimuli matched the previously seen sample stimulus. There were 4 trials in each DMS block (total block length = 36 s). For a visual depiction of all tasks, see Figure S1 in Garrett et al. (2011).
Before scanning, participants were tested in a mock scanner to determine within-subject accuracy thresholds for each task. These thresholds indicated the difference in center frequency between stimulus columns necessary for accuracy in each task to be ∼80% during the scanning session. Although accuracy was roughly equated across subjects, mean reaction time and response variability still varied greatly across subjects/groups (see Supplementary Table S1).
MRI Scanning and Preprocessing
Brain images were acquired with a Siemens Trio 3T magnet. We first obtained a T1-weighted anatomical volume using a SPoiled Gradient Recalled (SPGR) sequence (time echo [TE] = 2.6 ms, time repetition [TR] = 2000 ms, field of view [FOV] = 256 mm, slice thickness = 1 mm) for co-registration with the functional images, and to ensure that there were no significant brain abnormalities in any participants. We then obtained T2* functional images (TE = 30 ms, TR = 2000 ms, flip angle = 70°, FOV = 200 mm) using echo-planar imaging acquisition. Each functional sequence consisted of 28 five-mm thick axial slices, positioned to image the whole brain. Images were registered to a nonlinear group average anatomical image (Kovacevic et al. 2005; Chen et al. 2006; Levine et al. 2008) intended to serve as an unbiased anatomical template. We have employed this template in previous studies of both young and older adults (Garrett et al. 2010; Grady et al. 2010). Functional data were slice-time corrected using Analysis of Functional Neuroimages software (http://afni.nimh.nih.gov/afni) and motion corrected using Automated Image Registration software (http://bishopw.loni.ucla.edu/AIR5/) by registering all functional volumes to the 100th volume within run. By averaging all functional volumes within a motion-corrected run, we calculated mean functional volumes. For each run, the mean functional volume was registered with each subject's structural volume using rigid body transformation. After appropriate transform concatenations, from initial volume to the 100th volume within run, from mean run volume to structural volume, and from structural volume into our template space, we obtained a direct nonlinear transform from each initial fMRI volume into template space. We then applied the FSL/FNIRT registration algorithm to find a nonlinear transform between our template and MNI 152_T1 provided with FSL software (www.fmrib.ox.ac.uk/fsl). Data were smoothed using an 8-mm Gaussian kernel.
We also performed a number of subsequent preprocessing steps intended to further reduce data artefacts (see Garrett et al. 2010, 2011). We first corrected functional volumes in the Common Template space for artefacts (e.g., motion effects, high frequency components, susceptibility, scanner drift, ventricle activation, etc.) via Independent Component Analysis (ICA) within run, within person, as implemented in FSL/Melodic (Beckmann and Smith 2004). Resulting components were classified manually as artefacts according to several primary criteria: 1) the presence of motion (e.g., ringing effects and large time series spikes), 2) susceptibility and flow artifacts (e.g., sinus activation), 3) ventricle activation, 4) low frequency signal drift, 5) high power in high-frequency ranges unlikely to represent neural activity (above frequencies of 0.13 Hz), and 6) spatial distribution (e.g., spatial maps highly concentrated in very few voxels; see Kovacevic et al. 2010). Next, a training set was obtained by manually classifying artefact components from a small set of runs (∼20) within randomly selected subjects (and equally for younger and older adults). The training set was then used in a quadratic classifier to automatically separate components from all runs into artifact and nonartifact categories in those data not manually classified. Components identified as artefact were then filtered out from the corresponding fMRI runs. Following ICA denoising, voxel time series were further adjusted by regressing out motion correction parameters, and WM and cerebrospinal fluid (CSF) time series using in-house MATLAB code. For WM and CSF regression, we extracted time series from unsmoothed data within small regions of interest (ROIs) in the corpus callosum and ventricles of the Common Template, respectively. ROIs were selected such that they were deep within each structure of interest (corpus callosum and ventricles) to avoid partial volume effects. The choice of a one 4-mm3 voxel within corpus callosum for WM and same size voxel within one lateral ventricle for CSF was based on our experience in having excellent registration of these structures across all ages. To localize regions from our functional output, we submitted MNI coordinates to the Anatomy Toolbox in SPM8, which applies probabilistic algorithms to determine the cytoarchitectonic labeling of MNI coordinates (Eickhoff et al. 2005, 2007).
To prepare our reaction time (RT) data for examination, we set a lower bound (150 ms) for responses for each task on the basis of minimal legitimate RTs (MacDonald et al. 2006; Dixon et al. 2007). We then initially trimmed extremely slow trials relative to the rest of the sample on each task (≥4000 ms). Final upper bounds were established for each task by dropping all trials more than three SDs from within-person means. The number of trials dropped across all participants and tasks was negligible (179/5040 total trials). For each task, missing values due to outlier trials were then imputed via regression imputation (as implemented in SPSS 18.0). We subsequently calculated mean RTs and intraindividual SDs (ISDs) of RTs for each participant on each task according to previous research (Hultsch et al. 2000, 2008; MacDonald et al. 2006; Dixon et al. 2007). Significant relations between various variables (e.g., age and practice effects) and meanRTs are typically observed, and meanRTs are often positively correlated with raw SD values. To remove these potential confounds prior to ISD calculation, we residualized the effects of age, block, trial, and all interactions from all RT trials separately for each task. Finally, we computed within-task within-person ISDs from these trial-level residuals.
Calculation of Blood Oxygen Level–Dependent Signal SDs
To calculate blood oxygen level–dependent signal SDs (SDBOLD) for each voxel, person, and experimental condition, we first normalized all blocks within condition such that the overall 4D mean across brain and block was 100 to correct for possible low-frequency artifacts. For each voxel within person and condition, we then subtracted block means and concatenated across all blocks (for further details, see Garrett et al. 2010). Finally, we calculated voxel SDs across this concatenated mean-block corrected time series. Our previous work demonstrated that our more extensive preprocessing pipeline (ICA, motion, WM, and CSF parameter regression block normalization) reduced the average voxel SD by half compared to standard preprocessing steps, while simultaneously doubling the age predictive power of brain variability (Garrett et al. 2010). Thus, more extensive removal of possible low- and high-frequency “junk noise” sources in our sample likely serves to improve the predictivity of the SD–brain signal.
PLS Analyses of Experimental Conditions by Age Group and Subsequent Relations to Cognitive Performance
To examine multivariate relations between brain variability and experimental conditions in the context of young and older adult groups, we utilized PLS analyses (Task PLS; McIntosh et al. 1996; McIntosh and Lobaugh 2004; Krishnan et al. 2011). Task PLS begins with a covariance matrix between the experimental conditions and each voxel's signal (“signal” here refers to voxel SD); covariances are calculated across subjects, within group for each task. Each group's covariance matrices are then “stacked” (i.e., horizontally concatenated) into a single covariance matrix. This final covariance matrix is decomposed using singular value decomposition to produce orthogonal latent variables (LVs) that optimally represent relations between brain voxels and experimental conditions in the context of our 2 groups. Each LV contains a spatial activity pattern depicting the brain regions that, as a whole, show the strongest relation to (e.g., are covariant with) the task contrast identified by the LV. Each brain voxel has a weight, known as a salience, which is proportional to the covariance of activity with the task contrast. To obtain summary measures of each participant's expression of an LV's pattern, we calculated within-person “brain scores” by multiplying each voxel's salience or weight by the BOLD SD in that voxel, and summing over all brain voxels for each participant. Significance of detected relations between multivariate spatial patterns and experimental conditions was assessed using 1000 permutation tests of the singular value corresponding to each LV. A subsequent bootstrapping procedure revealed the robustness of voxel saliences across 1000 bootstrapped resamples of our data (Efron and Tibshirani 1993). By dividing each voxel's mean salience by its bootstrapped standard error, we obtained “bootstrap ratios” as normalized estimates of robustness. We thresholded bootstrap ratios at a value of ≥3.00, which approximates a 99% confidence interval. Further, to compare experimental conditions/groups within an LV, confidence intervals (95%) for the average brain scores in each condition were calculated from the bootstrap, and differences in activity between conditions/groups were determined via a lack of overlap in these confidence intervals.
Following our PLS analysis of experimental conditions by age group, we subsequently tested linear and nonlinear relations between this PLS pattern (within-subjects) and age and cognitive performance (between-subjects) using repeated measures analysis of variance (RMANOVA) in SPSS 18.0 (SPSS Inc.).
Differences in BOLD Variability across Conditions by Age Group
First, using Task PLS (McIntosh et al. 1996; McIntosh and Lobaugh 2004; Krishnan et al. 2011), we examined whether multivariate patterns of voxel SDs would differentiate our experimental conditions across age groups. We found a very strong relation between brain and experimental condition within a single significant LV (singular value = 2.57, permuted P < 0.0001) that expressed reliable increases in brain variability when transitioning from fixation to task. Bootstrapped estimated confidence intervals (see Fig. 1a) around derived brain scores revealed that the degree of transition in variability from fixation to task was reduced in the old group. Within groups, the young group showed marked increases in brain variability on task compared with fixation (see blue bars in Fig. 1a; for the spatial pattern, see Fig. 1b). Notably, the old group showed a more modest increase in variability between fixation and tasks (PMT, ATT) and not at all between fixation and DMS (see red bars in Fig. 1a; PMT and DMS also did not differ). We then formally tested this apparent interaction between brain score for each condition (within subjects) and age group (between subjects) using RMANOVA and found it was significant (multivariate F3,40 = 4.23, P = 0.01, partial η2 = 0.24; a significant main effect of brain score (i.e., condition) was also present (multivariate F3,40 = 6.78, P = 0.001, partial η2 = 0.34)). Within-subject contrasts of the interaction revealed a significant linear trend (F1,42 = 6.81, P = 0.01, partial η2 = 0.14) and a marginal cubic trend (F1,42 = 3.30, P = 0.08, partial η2 = 0.07).
The bootstrapped spatial pattern revealed a widely distributed set of regions, every one of which increased in variability from fixation to task (see Fig. 1b). Among other smaller clusters (e.g., middle occipital gyrus and superior frontal gyrus), a remarkably large cluster resulted (4732 voxels; peak = lingual gyrus; other prominent effects were present in ventromedial prefrontal cortex, cerebellum), indicating the generalized and expansive nature of this phenomenon, particularly in the young group. For peak MNI coordinates, bootstrap ratios, and cluster sizes for each reliable cluster, see Supplementary Table S2.
Comparison of Young and Old Group Spatial Patterns
To more clearly differentiate how our age groups contributed to the overall spatial pattern determined in our whole-sample model (Fig. 1), we subsequently examined our young and older groups in separate Task PLS analyses. Because the overall PLS model we ran with both groups revealed a single spatial pattern per LV, it remained unclear whether the regions that shift between experimental conditions might differ by age group. Running separate models for young and older groups thus allowed separate visualizations of each group's respective spatial map, which we could then compare.
Results for the young group again demonstrated a clear fixation versus task contrast (singular value = 2.25, permuted P < 0.0001) that mirrored the whole-sample model results above. A large subset of regions (e.g., lingual gyrus, cerebellum, ventromedial, middle, and superior frontal gyri; see Fig. 2a and Supplementary Table S2) showed greater brain variability on task compared with fixation. Our examination of the older group also revealed a significant contrast of experimental conditions that mirrored our overall sample results (singular value = 1.60, permuted P < 0.0001), in which brain variability increased from fixation to PMT and ATT, but not reliably from fixation to DMS (bootstrapped error bars overlapped; see Fig. 2b). However, this contrast was represented in a substantially smaller set of regions relative to the young/high group (e.g., lingual and inferior frontal gyri; see Fig. 2b and Supplementary Table S2). Thresholded voxels that comprised this effect for the old group represented only 29% (1091 total voxels) as many as for the young group (3674 total voxels).
To better gauge similarities and differences between spatial patterns, we overlaid the spatial maps noted for young and older groups from Figure 2. In Figure 3, it is clear that there is relatively little overlap between group maps, suggesting that the regions that reliably increase in variability from fixation to task in the young group are largely absent in the old group, and vice versa. We confirmed this visual interpretation upon calculation of the correlation between the bootstrap ratio spatial maps (see Fig. 2a,b) for each age group (r = 0.14). A correlation between saliences (unthresholded weights) of the respective spatial maps was also weak (r = 0.28).
Modulation in BOLD Variability across Conditions by Level of Cognitive Performance
Next, we tested whether an interaction was also present between brain scores for each condition (derived from our full PLS model above) and cognitive performance using RMANOVA. For our brain score × meanRT model, we found a substantial interaction (multivariate F3,40 = 5.57, P = 0.003, partial η2 = 0.30). Within-subject contrasts revealed significant linear (F1,41 = 5.10, P = 0.029, partial η2 = 0.11) and cubic trends (F1,41 = 8.54, P = 0.006, partial η2 = 0.17). Because meanRT is a continuous measure, we employed typical methods of visualizing discrete (i.e., brain scores for each condition) by continuous variable interactions (Cohen and Cohen 1983; Aiken and West 1991). Accordingly, we calculated point estimates based on ±1 SD from the whole-sample meanRT value and plotted the interaction with condition brain scores (see Fig. 4). This interaction highlighted that faster performers showed greater increases in brain variability from fixation to task than did slower performers, and the form of this interaction was highly similar to that noted between brain score and age group in Figure 1a. No reliable interaction existed between brain score and ISDRT (P = 0.17), although these data showed an effect similar to that seen in Figure 4. Furthermore, although we attempted to establish within-person accuracy levels at 80% prior to scanning, subtle age group differences in accuracy did result (older adults had slightly lower DMS and overall accuracy; see Supplementary Table S1); regardless, accuracy did not interact with brain scores (P = 0.18). Given age group differences in education (see Supplementary Table S1), we also related education to brain scores; we found no reliable effect.
Reliable Measurement of Block Design-Based Modulations in BOLD variability
Although we have taken various strides to accurately measure modulations in SDBOLD levels (see Materials and Methods), one potential remaining concern regarding the comparison of brain variability between experimental conditions in our (or any other) block design is whether the variability level of any given condition is “condition-pure.” Because our blocks alternated fixation–task–fixation–task throughout, it is possible that the variability in any given fixation block may be impacted by the signal dynamics of the immediately preceding task block (and vice versa). If true, a preceding block should have the greatest impact on variance within the first portion the succeeding block (i.e., due to BOLD signal spillover from the preceding block) rather than within later portions of the succeeding block. To test this, we calculated split-half reliability on within-person whole brain SDBOLD values from concatenated first and second block halves for each experimental condition. Results indicated that split-half reliability was near unity (r range across tasks = 0.91–0.97), with no clear difference in SD magnitude between the block halves (see Fig. 5). We subsequently examined split-half reliability for each age group separately; both young (r = 0.96, 0.93, 0.97, and 0.94) and old groups were similarly reliable (r = 0.97, 0.91, 0.92, and 0.96) for Fix, PMT, ATT, and DMS tasks, respectively. Furthermore, when we examined split-half reliability only for robust voxels from our full sample model (i.e., those voxels visible in Fig. 1b), values were virtually identical to our whole brain estimates shown in Figure 5 (full sample [r = 0.97, 0.90, 0.94, 0.94]; young group [r = 0.97, 0.93, 0.97, 0.92]; old group [r = 0.97, 0.89, 0.94, 0.96]). Thus, we find no evidence overall or within age groups that SDBOLD levels are markedly impacted by signal spillover.
In the current study, we sought to characterize how variability in BOLD activity differs by experimental condition across levels of age and cognitive performance. In particular, we were interested whether primary state-to-state increases in variability would occur from fixation (internal cognition) to task (external cognitive demands). We confirmed the presence of a substantial contrast that suggested reliable increases in brain variability when transitioning from fixation to task, especially in young (and in fast performing) adults. Regardless of group, a unidirectional pattern of variability existed across brain regions; when external cognitive demands were placed on participants, voxel variability only increased.
To more precisely investigate young and old group differences, we examined these groups in separate models to compare their respective spatial maps. For the young group, an expected fixation versus task contrast emerged in which variability only increased across brain regions. For the old group, a reliable experimental condition contrast also emerged, although this was represented in relatively few brain regions that bore little visual or statistical resemblance to the young pattern. This suggests some level of bidirectionality in our results (as seen in our previous work; Garrett et al. 2010, 2011), in which younger adults modulated variability in an expansive set of regions, and older adults modulated in a largely different and reduced set of regions. Interestingly, older adults did not reliably modulate in variability level from fixation to DMS. As this may be the first study to examine internal to external state modulation in brain variability levels, the reasons for a lack of fixation to DMS modulation are not clear. Whether this effect is due to cognitive processes, task design parameters, or other effects will require future work. Nonetheless, it remains clear that older group modulation from fixation to task is reduced compared with younger adults, both in magnitude and spatial representation.
Our findings expand on recent MEG work establishing task-to-task brain variability transitions in children and young adults (Misic et al. 2010), and on work examining relations between brain variability, age, and performance (Garrett et al. 2010, 2011) to now help establish the developmental and performance correlates of transitions in variability from internal-to-external cognitive states. The focus of our previous papers (Garrett et al. 2010, 2011) was not on detecting differences in state-to-state transitions (i.e., between rest and task, as in the present case) but rather on whether younger and better performing adults were relatively more or less variable in BOLD at fixation or during task performance. The current study extends these results to examine a highly topical area of study, that of differences between fixation and task modes in the human brain.
Notably, our results were achieved while equating task accuracy for each task at ∼80% prior to scanning, within subject. Thus, our reported age and performance-based differences in brain variability would likely only increase in situations in which task difficulty is not equated between persons; greater task difficulty for older adults may yield further decreases in variability modulation between fixation and task. At the same time, if we did not equate accuracy, group differences in fixation to task modulations could have simply been due to increased task difficulty in older and poorer performers (see Stern et al. 2005; Reuter-Lorenz and Cappell 2008). Regardless, even though accuracy was largely matched, speeded performance still varied across subjects (see Supplementary Table S1) and at least for meanRT, strongly predicted brain variability modulations between fixation and task.
Examining Brain Variability in the Context of Rest and Task States
Using mean–brain measures, the contrast of rest versus task typically results in distinct patterns of activity for both conditions, which have come to be known primarily as the default and task-positive modes, respectively. In addition, default and task modes are often anticorrelated and are among the most reliable findings in functional neuroimaging (Damoiseaux et al. 2006; Toro et al. 2008; Spreng et al. 2009; Andrews-Hanna et al. 2010; Grady et al. 2010). Using the same sample and tasks as employed here, Grady et al. (2010) previously demonstrated clear mean signal-based fixation versus task contrasts, with a typical and expected default (e.g., medial prefrontal and parietal cortices) versus task-positive network pattern (e.g., anterior superior and inferior parietal regions; inferior frontal gyri) in which fixation differed greatly from tasks but tasks differed very little. We also showed various differences with age in these patterns (e.g., an expanded task-positive network). In the present study, a clear fixation versus task contrast existed with our measure of brain variability. However, this contrast highlighted a spatial pattern (e.g., a large portion of the bilateral cerebellum) that was marked not by the anticorrelated default and task-related patterns typically found with mean signal-based analyses, but instead by a unidirectional increase in brain variability (see Figs. 1 and 2). Remarkably, no region decreased in variability on task, irrespective of the model we ran. Thus, it appears the neural manifestation of variability-based transitions from fixation to task could be conceived of as a magnitude increase that tempers from young to older age and with poorer cognitive performance (combined with a qualitative age-related shift in spatial foci), rather than a give-and-take between opposing networks (as for mean signal). These marked differences between SD- and mean-based results converge with our previous findings that the 2 brain measures produce nonoverlapping, yet complementary results (Garrett et al. 2010, 2011).
Although our SD–brain spatial pattern did not resemble typical mean signal-based rest versus task regions, certain regions thought to be part of the default and task-positive networks did show variability changes from fixation to task. For example, particularly in young adults, we noted reliable increases in brain variability on task in inferior and dorsolateral prefrontal regions, both of which also typically show increased mean activity on task as part of the task-positive network; however, large portions of the task-related sensorimotor cortex were not present (e.g., Fox et al. 2005; Toro et al. 2008). Interestingly, we also found that a well-known default mode region, the ventromedial prefrontal cortex, increased in brain variability on task despite typical mean-based decreases in activity on task in this region. Interestingly, this suggests that the vmPFC signal is more stable during fixation, even though increased trial-to-trial neural variability can be a function of greater coherence between regions (Fox et al. 2006), and the vmPFC typically serves as an important region in the default network (Raichle et al. 2001; Greicius et al. 2003; Andrews-Hanna et al. 2007). However, perhaps the most critical hub in the default mode network, the posterior cingulate cortex/precuneus (Achard et al. 2006; Hagmann et al. 2008; Buckner et al. 2009), was largely absent in our pattern (only a small area of the ventral precuneus was present). Disentangling these intriguing relations between brain variability and default and task-positive regions will require future exploration, ideally within the context of a broad-based comparison of relations between functional connectivity, mean signal, and brain variability.
Why an Increase in Brain Variability during Internal-to-External State Transitions?
There are a number of well-documented reasons why brain variability is functional for neural systems, such as greater network complexity (McIntosh et al. 2008), connectivity (Fox et al. 2006), network robustness (Basalyga and Salinas 2006), improved signal detection thresholds (Li et al. 2006; McDonnell and Abbott 2009), optimized learning (Mandelblat-Cerf et al. 2009), and better cognitive performance (McIntosh et al. 2008; Garrett et al. 2010, 2011). But, how do we interpret a greater increase in brain variability on task? First, this increase could indicate a more sophisticated neural system capable of greater dynamic range. Greater dynamic range is important for the adaptability and efficiency of neural systems because it allows a greater range of response to a greater range of stimuli. Interestingly, cellular research indicates that dynamic range (and information transfer within networks) is best achieved when cellular excitation and inhibition are balanced, and response variability is optimally high; upon imbalance and a lack of response variability, dynamic range suffers (Shew et al. 2009, 2011). In the present case, our young and fast performers appeared to exhibit the greatest dynamic range of BOLD signal (i.e., greater signal variability), within and across voxels and when making transitions from rest to task.
Second, variability may increase on task due to greater stimulus uncertainty. Some researchers (Knill and Pouget 2004; Ma et al. 2006; Beck et al. 2008) argue that variability is critical for the nervous system to operate in an optimal probabilistic Bayesian manner. Effectively, neural variability yields adaptability across levels of uncertainty in one's environment. The authors contend that if neurons fired in exactly the same manner every time a specific stimulus was encountered, one would be less able to adapt to different circumstances that involve that stimulus. In a probabilistic manner, the brain evaluates the reliability of incoming signals and optimally chooses from a distribution of options, resulting in reliable firing in the presence of environmental uncertainty. In our study, participants were faced with a series of different stimuli that were altered trial-to-trial and block-to-block according to task demands. We found that BOLD variability only increased when shifting from fixation (in which relatively little environmental uncertainty exists) to an externally oriented state (in which unusual cognitive demands were placed on participants with stimuli not normally encountered). Because our external world is naturally made up of moment-to-moment changes in stimulus information (e.g., as we walk, interact, and experience), it is plausible that our neural responses should be accordingly variable when processing this environmental information, with reliance on probabilistic population coding to prepare, adjust, and handle levels of stimulus uncertainty that may occur.
Finally, another possible reason for increases in variability on task is that brain variability may provide the kinetic energy for networks to achieve a variety of possible functional states, and that this energy is more vital on task. When variability is too low, there is little capacity for the system to achieve new states, yielding the potential for the system to remain rigidly and maladaptively in a single state (Ghosh et al. 2008; Deco et al. 2009, 2011; McIntosh et al. 2010). When comparing fixation to task then, the requirement for kinetic energy may increase on task due to constantly changing external stimuli that, to be effectively processed, require a multitude of functional states. Because our participants did not know from moment to moment exactly what they would be shown on task (compared with relatively predictable fixation blocks), the optimal number of functional states may have been ideally greater in order to handle a varying cognitive load, while allowing for potential adjustments in neural processing and strategic approach as the tasks progressed. Our young and fast performers may be most functional and adaptable in this way, as evidenced by significantly greater variability on each task relative to older, slower adults.
New Evidence for the Dedifferentiation of the Aging Brain
In a recent paper (Garrett et al. 2011), we found a lack of brain variability differences across regions in older poorer cognitive performers relative to younger higher performers, and we interpreted this as novel evidence for brain dedifferentiation. Dedifferentiation theory (Baltes and Lindenberger 1997; Park and Reuter-Lorenz 2009; Park et al. 2010; Carp et al. 2011) suggests that older brains may become less functionally distinct during task performance (i.e., through reductions in the selectivity and specificity of neural processes). Mean–brain-based findings often indicate that young adults typically exhibit more cortical selectivity (distinctiveness) from task to task, whereas older adults show more similar patterns across conditions (e.g., Carp et al. 2010, 2011; Park et al. 2010). Our previous results (Garrett et al. 2011) indicate age- and performance-based dedifferentiation of signal variability across a variety of regions. Our current results expand this finding in a new direction to suggest that older and slower brains are not only more dedifferentiated across spatial regions with regard to brain variability, they are also more dedifferentiated across internal and external cognitive states (i.e., older adult levels of brain variability do not vary across these states as much as in young adults). Even when examined in a separate analysis unconstrained by age group differences, relatively few regions increased in variability from fixation to task (or between tasks) for the old group (see Fig. 2b). These results indicate that older brains simply become more rigid and uniform, within and across regions, and within and across experimental conditions. Brain variability measures thus represent an alternative view of cortical selectivity and dedifferentiation in the study of aging and cognition; older brains not only become less cortically selective within a given cognitive state but they are also less cortically responsive to changes between states.
Key Future Direction of Interest: Parametric Task Manipulation
One key future direction for work on condition-based modulations in brain variability (and in relation to age and performance) is to test whether brain variability can be manipulated parametrically within a variety of cognitive domains. That is, does brain variability adjust to the precise level of externally driven cognitive demand or does brain variability primarily adjust to a generalized task-relevant state (i.e., similar to our current results, which utilized task accuracy levels preset to ∼80% within person)? If brain variability does adjust to specific levels of cognitive demand, the sensitivity and utility of brain variability in cognitive neuroscience and aging research would increase dramatically. Ideally, a parametric manipulation of brain variability would also be combined with a thorough examination of co-occurring manipulations of functional connectivity (e.g., correlation, graph theory, mutual information). Recent EEG work on children and young adults suggests that brain variability covaries broadly with functional connectivity (Misic et al. 2011; Vakorin et al. 2011), which may thus provide a useful neural context within which to interpret the experimental malleability of brain variability.
As is typical for mean signal, we established here that BOLD variability was an effective discriminator of internal cognitive processes and externally driven cognitive demands. We found a broad increase in brain variability on task (compared with fixation) that reduced in overall magnitude and spatial expression with older age and slower cognitive performance. Notably, this may reflect reduced dynamic range and a lack of ability to efficiently process varying and unexpected external stimuli. Older poorer performing brains also appeared more dedifferentiated, both within and across brain regions and experimental conditions. The current results continue to support the idea that moment-to-moment brain variability represents an important “signal” within what is typically considered measurement-related “noise” in fMRI (Garrett et al. 2010, 2011), and may open a new line of inquiry into variability-based default versus task modes in the human brain.
Supplementary material can be found at: http://www.cercor.oxfordjournals.org/
Canadian Institutes of Health Research (grant MOP14036 to C.L.G. and grant MOP13026 to A.R.M.); JS McDonnell Foundation (to A.R.M). Canada Research Chairs program, Ontario Research Fund, and Canadian Foundation for Innovation (C.L.G.).
Conflict of Interest : None declared.