All procedures were approved by the Institutional Review Boards of Yale University and the VA Connecticut Healthcare System. Informed consent was obtained from 20 right-handed healthy control subjects (9 males and 11 females, mean age
SD of 26.6
8.0 years, mean years of education
SD of 17.0 ± 2.9). Mean IQ
SD estimated from performance on the Wechsler Test of Adult Reading (WTAR) (Ginsberg 2003
) was 112.7 ± 11.6. Each subject was assessed with the Structured Clinical Interview for DSM-IV to rule out presence of psychiatric illness (First et al. 2002
). Subjects with significant medical or neurological illnesses were also excluded.
N-back task (Fig. )
Each subject underwent a functional magnetic resonance imaging (fMRI) session composed of four separate runs, with each run consisting of 11 blocks lasting 32–40 s each, with 10 s between each block for rest. Blocks were counterbalanced between runs to minimize potential confounds of fatigue and block order. Subjects were given trials grouped in blocks, with 16–20 trials/block. Each trial consisted of a visual presentation of a letter for 500 msec followed by an interstimulus interval of 1500 msec. Subjects were instructed to press a target button as soon as possible for each trial in which the letter shown was the same as the letter shown N trials previously, and to press a non-target button otherwise. For the 0-back, targets were each occurrence of the letter “a.” Target responses were set to occur with 50% frequency, with accuracy computed as the number of correct responses divided by the number of trials. Initial letters in a sequence with no comparison letters were not counted as trials. Subjects were given blocks of 0-back, 1-back, 2-back, and 4-back, with half of the 2-back blocks occurring immediately after a 4-back block (“2-back/4”), and the remaining 2-back blocks occurring immediately after a control-condition 1-back block (“2-back/1”). The 4-back blocks were used intentionally to exceed subjects’ working memory capacity, with the expectation of subsequent deficits in 2-back performance and associated involvement of limbic circuitry. Prior to the fMRI testing session, each subject was given written and verbal instructions on the task and underwent a practice session composed of blocks of 0-back, 1-back, and 2-back. In order to maximize the subjective failure experience of 4-back condition, subjects were told they may be given 3-back or 4-back blocks, but were not given practice trials at this difficulty level. Subjects were instructed to perform the tasks to the best of their ability, emphasizing accuracy over reaction time when possible. They were informed that this experiment would assist in our understanding of cognitive processes but were otherwise not told about the specific purpose of the experiment until after the fMRI session. Upon completing the fMRI runs, subjects were given a post-test questionnaire that asked them to rate the different N-back conditions on separate 5-point scales of difficulty and frustration (1
not at all difficult/frustrating, 5
very difficult/frustrating). As subjects were not made aware of the distinction between 2-back/1 and 2-back/4 blocks until after the testing session, their ratings did not distinguish between these two conditions.
Fig. 1 Experimental design. a Sample N-back task for N=2. b Block design. Subjects underwent runs in Sequence A and B in counterbalanced fashion FMRI acquisition and preprocessing
Blood oxygen level-dependent (BOLD) fMRI data were collected on a Siemens Trio 3-Tesla scanner using an EPI pulse sequence (TR
2 sec, TE
20 cm) and whole brain coverage (33 axial oblique slices, slice thickness
4.0 mm, gap
0.5 mm, voxel size
). Preprocessing of images was performed using SPM2 (http://www.fil.ion.ucl.ac.uk/spm
). Slice timing correction was performed using the middle slice as the reference, and the slices were re-orientated setting the origin at the anterior commissure and setting the horizontal axis to run along the anterior commissure-posterior commissure line. Motion correction was then performed with INRIalign (Freire et al. 2002
) using the first scan of each run as the reference. For each subject, the motion parameters of each run were visually inspected. One subject exhibited greater than 2.5 mm of translational motion during the experiment. This subject was excluded from further fMRI analysis. Only two of the six motion parameters (movement in y and z directions) exhibited greater than 0.5 mm or 0.5° deviations for any subject during the scanning. These two parameters were entered as covariates of non-interest during model specification. Following motion correction, mean functional scans were normalized to the EPI average brain template from the Montreal Neurological Institute (MNI; supplied in SPM2) using affine and nonlinear warping algorithms, and the resulting transformation matrix was used to normalize all the individual scans in the time series. Functional scans were then re-sampled to a 4
4 mm voxel size. Spatial smoothing was performed using a 10 mm full-width half-maximum Gaussian kernel, and data were subjected to a high-pass temporal filter (.008 Hz).
Analysis of behavioral data
Mean accuracy and median reaction times for each condition were determined for each subject, and subsequently used to generate overall group means. Percent change in 2-back accuracy was calculated as (mean accuracy on 2-back/4 minus mean accuracy on 2-back/1) divided by mean accuracy on 2-back/1. Subjective responses, accuracy scores, and reaction times were analyzed using repeated measures multivariate analysis of variance (MANOVA) (F tests based on Wilks’ Lambda), with task condition (0-back, 1-back, 2-back, and 4-back) as a within-subjects factor. Paired t-tests (two-tailed) were used to examine accuracy and reaction time differences between the 2-back/1 and 2-back/4 conditions. Otherwise, post-hoc comparisons of task conditions were made only when the main analysis was significant at the p
Analysis of fMRI data
fMRI data were analyzed using the general linear model as implemented in SPM2
. For first-level individual subject analyses, fixed effects multiple linear regression time series analysis was implemented to model task condition effects, generating images of the parameter estimates (beta images) for each condition, as well as contrast images generated by subtracting beta images for specific conditions. For second-level group analyses, a random-effects model was applied to the individual subject beta or contrast images derived from the first-level analyses to determine the location and extent of brain activations. Exploratory voxel-wise one-sample t-tests of whole brain contrast images were conducted using a p
.05 voxel-level probability threshold and an extent threshold of 3 voxels, with correction for multiple comparisons based on the false discovery rate (FDR) (Genovese et al. 2002
). Cluster-level analysis was based on these same parameters. Within each significant (corrected p
.05) cluster, the percent of each region activated and the corresponding number of activated voxels were determined using the MSU utility (http://www.ihb.spb.ru/~pet_lab/MSU/MSUMain.html
). All voxel coordinates are given in MNI space.
Region of interest analysis
Mean activation of voxels for each task condition was determined for a priori regions of interest (ROIs) including left and right DLPFC (BA46), amygdala, and VMPFC (BA25). Analyses were conducted separately for each hemisphere in consideration of the verbal nature of the working memory task and studies reporting lateralized cognitive load effects or emotional-cognitive interactions(Altamura et al. 2007
; Erk et al. 2007
; Low et al. 2009
; Sandrini et al. 2008
; Siegle et al. 2006
). Because of the relatively large area encompassed by these anatomical regions, “functional” ROI masks were generated by identifying voxels within each anatomical region that exhibited supra-threshold activation or deactivation in any of the task conditions. This procedure prevented attenuation of task-related signals associated with inclusion of inactive voxels in the calculation of ROI means. The first step of this ROI procedure involved conducting second-level group analyses for the contrasts of each task condition with the 0-back condition (1-back minus 0-back, 2-back/1 minus 0-back, 2-back/4 minus 0-back, 4-back minus 0-back, and the reverse contrast for each of the above). The resulting activation maps were masked with anatomical ROIs generated by the WFU Pickatlas (Maldjian et al. 2003
) (dilated x1), based on the Talairach Daemon database (Lancaster et al. 1997
). For each of the above contrasts, voxel values within the masks were thresholded at p
.001, uncorrected for multiple comparisons, with an extent of 3 voxels. For each ROI, the supra-threshold voxels generated from each task condition contrast were combined, forming a single functional ROI mask that represented the union of the activated voxels from each contrast. This ROI mask thus included only voxels that were activated or deactivated in at least one of the task conditions, but was otherwise unbiased in terms of direction of activation. The resulting six functional ROI masks (right and left BA46, amygdala, and BA25) were overlaid on the individual subject contrast images for each condition vs. the 0-back control condition, and the mean activation of all voxels within each of these ROIs was tabulated for each subject. Values from 2-back/1 and 2-back/4 were initially pooled to form a single 2-back condition. Two-way (condition x hemisphere) repeated measures MANOVA (F tests based on Wilks’ Lambda) was used to analyze the subjects’ mean activations in the 1-back, 2-back, and 4-back conditions for each ROI, and to test for hypothesized differences in regional brain activity between the 2-back/1 and 2-back/4 conditions. Where appropriate, follow-up paired t-tests (two-tailed) were used to further parse the results.Across subjects, mean activity values derived from contrast images in bilateral amygdala for the 4-back and 2-back (pooled 2-back/1 and 2-back/4) conditions (minus 0-back) were tested for correlation with percent change in accuracy from 2-back/1 to 2-back/4. Similar voxel-wise correlational analyses were also performed within SPM2 for Talairach-based bilateral amygdala ROI masks using small-volume correction, an extent threshold of 3 voxels, and a significance threshold of p
.05, corrected for multiple comparisons based on FDR. All correlations in this study were performed using two-tailed assumptions.
For each run of each subject’s testing session, a representative amygdala time series was extracted using the first eigenvariate of the voxels in bilateral Talairach-defined amygdala. For each subject, a voxel-wise multiple regression analysis was performed, regressing each voxel’s time series on the amygdala time series as well as several covariates of non-interest, including the time series of global mean brain activity and y- and z- motion parameters. Resulting beta images for each subject, reflecting voxel-wise correlations (i.e. functional connectivity) with amygdala activity, were then passed forward to second level random effects group analysis. Exploratory voxel-wise t-maps were generated for positive and negative linear regression slopes using a probability threshold of p
.05 (FDR-corrected) and an extent threshold of 3 voxels. To test for hypothesized functional connectivity between the amygdala and specific ROIs, small volume corrections were applied using ROI masks for DLPFC (Talairach-defined BA 9 and 46) and VMPFC (Talairach-defined BA25) generated using WFU Pickatlas (dilated x1). A similar procedure was utilized to explore functional connectivity between VMPFC and DLPFC, using VMPFC as the seed region. Talairach-defined regions were used here as opposed to the “active voxel” approach used in the ROI analysis due to the emphasis of functional connectivity on relationships over time as opposed to mean activity.
Psychophysiologic interaction (PPI) analysis
In order to explore whether the functional relationship between amygdala and DLPFC varied between the difficult 4-back and control 1-back conditions, PPI analysis was conducted using SPM2 based on Gitelman, et al. (Gitelman et al. 2003
) Specifically, we tested whether the slope of the regression line relating amygdala activity to DLPFC activity in the individual subjects’ time-series data differed between the 4-back and 1-back conditions, and whether this slope difference correlated with the percent change in 2-back accuracy across subjects. PPI-analysis was conducted for each individual subject at the first level, with PPI beta images then being passed forward to second level random effects group analysis. The left and right Talairach-defined amygdalae were each used as source regions in separate analyses. In the first level PPI analysis, the time series for the first eigenvariate of the voxels in the amygdala was extracted, representing the “physiological variable.” This hemodynamic time series was deconvolved to estimate the underlying neural time series used in the calculation of the PPI interaction term (Gitelman et al. 2003
). The “psychological variable” was a two-condition contrast vector representing the timing of the 1-back and 4-back conditions. The psychophysiologic (PPI) term was then defined as the product of the amygdala’s neural time series and the two-condition contrast vector, reconvolved with the default hemodynamic response function. To generate whole brain PPI maps, the time series for each voxel was regressed on the amygdala’s hemodynamic time series, the contrast vector, and the PPI product term in a multiple regression model. Beta images for the PPI term from each subject, reflecting the degree to which the slope of the regression line describing the relationship of each voxel’s time series with the amygdala’s time series differed between task conditions, were then carried forward to second level analyses. A whole-brain t-map was generated using voxel-wise one-sample t-tests of these PPI slope differences and thresholded using an FDR-corrected probability threshold of p
.05 and an extent threshold of 3 voxels. Because of our interest in the relationship between amygdala and DLPFC, the PPI t-map was interrogated with the same left and right DLPFC ROI masks used in the functional connectivity analysis described above. In order to test the relationship between amygdala-DLPFC interaction and behavioral effects of 4-back exposure, voxel-wise correlational analyses were performed between the PPI terms and percent decline in 2-back accuracy within the left and right DLPFC ROI masks, again using a voxel-wise probability threshold of p
.05 with FDR and small volume corrections.In order to derive the separate slope terms for the 4-back and 1-back conditions, separate task condition vectors for the 1-back and 4-back (i.e. vectors consisting of ones and zeros rather than ones and negative ones) were generated in addition to the single two-condition contrast vector. A time series was also extracted using the first eigenvariate of the voxels for DLPFC regions showing a significant PPI with amygdala. Linear regressions were then performed for each task condition vector separately, regressing the DLPFC time series on the amygdala time series, the single task condition vector, and the PPI product. Modeled in this way, the PPI product represents the slope of the DLPFC-amygdala correlation for the specified task condition, deriving separate slopes for the 1-back and 4-back conditions at the first level for each subject. These slopes were then compared across conditions in a second level random effects paired t
-test analysis. A similar procedure was used with the original two-condition contrast to extract the amygdala-DLPFC PPI values for the 4-back minus 1-back contrast. To assist in conceptualizing the above analysis, the six subjects showing the greatest percent decline in 2-back accuracy were compared to the six subjects showing the least decline using the Mann-Whitney U statistic.As a final step, 4-back accuracy, amygdala activity, and amygdala-DLPFC PPI product were entered into a multiple regression model predicting change in 2-back accuracy.