Mindfulness video game improves connectivity of the fronto-parietal attentional network in adolescents: A multi-modal imaging study

Mindfulness training has been shown to improve attention and change the underlying brain substrates in adults. Most mindfulness training programs involve a myriad of techniques, and it is difficult to attribute changes to any particular aspect of the program. Here, we created a video game, Tenacity, which models a specific mindfulness technique – focused attention on one’s breathing – and assessed its potential to train an attentional network in adolescents. A combined analysis of resting state functional connectivity (rs-FC) and diffusion tensor imaging (DTI) yielded convergent results – change in communication within the left fronto-parietal network after two weeks of playing Tenacity compared to a control game. Rs-FC analysis showed greater connectivity between left dorsolateral prefrontal cortex (dlPFC) and left inferior parietal cortex (IPC) in the Tenacity group. Importantly, changes in left dlPFC – IPC rs-FC and changes in structural connectivity of the white matter tract that connects these regions –left superior longitudinal fasiculus (SLF) – were associated with changes in performance on an attention task. Finally, changes in left dlPFC – IPC rs-FC correlated with the change in left SLF structural connectivity as measured by fractional anisotropy (FA) in the Tenacity group only.


Supplementary Information
Stimuli were presented with E-prime 2.0 software on a desktop computer screen. The task consisted of photographs of happy and fearful faces with words HAPPY and FEAR written on top of the photographs in red ink, see [1] for details. The stimuli were presented sequentially for 1,000 ms each, and a fixation cross was displayed between the stimuli for a variable duration (ISI =3,000; 4,000; or 5,000 ms). There were 148 trials, divided equally between congruent and incongruent trials.
Brain Imaging data acquisition: MRI data was acquired on a General Electric 3T MR750 MRI scanner (Waukesha, WI).
Resting-state functional MRI data was acquired with a series of sagittal T2*-weighted echo- Functional MRI resting-state data analyses were performed using AFNI [3] analysis package, unless otherwise indicated. Reconstructed echo-planar image volumes were first corrected for motion using rigid-body realignment (3dvolreg) and corrected for slice-timing differences (3Tshift). The first 3 images (6s) were removed to allow magnetization to reach equilibrium. Data were then corrected for B0-field distortions using customized in-house software that calls the FMRIB Software Library, FSL [4] functions PRELUDE and FUGUE.
Images were then aligned to the T1-weighted structural image using an affine transformation and a local Pearson correlation cost function [5]. T1-weighted structural data were aligned to the MNI template using nonlinear warping with ANTS [6]. This warp was then applied to the preprocessed fMRI data, and resampled to 2mm isotropic resolution. Automated segmentation (FSL's FAST) of the T1-weighted structural image was used to define masks of the WM and CSF [4,7,8].
The two signal intensity time-courses resulting from averaging the fMRI data within the eroded WM and CSF masks and their first derivatives (computed by backwards difference) were taken as signals of no-interest (i.e., spurious fluctuations unlikely to be of neuronal origin) and removed from the functional data along with the six rigid-body motion registration parameters [9,10]. Time points where the sum-squared difference (ssd) of consecutive points of the 6 motion realignment parameters exceeded 0.25 mm were censored and ignored in this nuisance regression. The functional images were temporally band-pass filtered between 0.01 Hz and 0.1 Hz, and spatially smoothed with a 3-dimensional Gaussian kernel (FWHM = 6 mm).
Functional connectivity was computed using a seed-to-voxel connectivity approach [11]. The left dlPFC seed region of interest [-42, 16, 28] was defined from an fMRI meta-analysis performed on 47 neuroimaging studies involving conflict resolution [12]. The region was one of the largest clusters reported for Stroop-like tasks (similar to the ECT). The dlPFC is also a critical region that has been found to change functionally in response to focused attention meditation [13]. The preprocessed fMRI data were averaged over the seed region of interest, and then regressed against all voxels in the brain. Time points with excessive motion (ssd > 0.25 mm) were censored. Functional connectivity maps were corrected for multiple comparisons using a cluster-threshold approach. The spatial smoothness of the preprocessed fMRI data was estimated using AFNI's 3dFWHMx [3]. The resulting estimated FWHM was used in a Monte Image pre-processing Brain tissue masks were extracted from b=0 images using the brain extraction tool of FSL [4]. The distortions introduced by eddy currents were corrected using a Gaussian process model based correction implemented in the 'eddy' tool of FSL [15]. A multi-compartment tissue model named neurite orientation dispersion and density imaging (NODDI) was fit to the corrected DWI signal for each voxel in the brain using a three stage (grid search, gradient descent and Markov Chain Monte Carlo) fitting procedure [16]. The intrinsic parallel diffusivity 8 was set to 1.7 x 10 -9 m 2 ·s -1 in the estimation procedure. From the estimated model extra-cellular diffusion tensors were reconstructed allowing us to extract the traditional diffusion tensor image (DTI) measures such as the fractional anisotropy (FA) and mean diffusivity (MD). The NODDI model itself offers neurite density, orientation dispersion and free-water fraction maps.
Unbiased study-specific coordinate system Unbiased global template space was estimated as shown in Figure 1. While estimating an unbiased atlas (coordinate system) is well investigated in cross-sectional imaging studies, there are fewer validation studies using a longitudinal design. The additional bias which we must restrict in a longitudinal study is the interpolation asymmetry that can arise when selecting only one of the time points as a temporal representative in generating the population/study level coordinate system as described in recent works [17]. The subject-specific average that is temporally unbiased was first estimated. The subject specific averages were then used to generate an unbiased population level average template space as shown in Figure 1, where each wavy black line represents a combination of affine and non-linear diffeomorphic transformations. These transformations and the spatial averages were estimated iteratively until convergence. We employed DTI-TK which is an open source and extensively validated image registration toolbox [18,19] for estimating the transformations using the diffusion tensors. We note that our processing offers improvement over [20,21]. [20] also recognizes the need for avoiding subject-specific averages and so heuristically generates a common template space. [21] does not generate a global coordinate system and instead involves a non-study specific stereotaxic space such as ICBM-152 which can introduce additional unwanted biases [22].

Processing for extracting regions of interest
The JHU-ICBM atlas defines 48 deep white matter regions of interest (ROIs) [23]. To obtain the diffusion measures in these ROIs, the JHU-ICBM FA template was registered to the unbiased global average (shown in Figure 1 and 2). The individual ROIs were then inverse warped into individual subject space. This is possible because the non-linear transformations 11 estimated using ANTS [24] are invertible up to numerical accuracy levels. Then to account for any registration imperfections the individual ROIs were thresholded on each diffusion measure map. For FA maps voxels with FA<0.2 were removed. For all the other measures first a standard deviation ( ) of the measure each ROI is estimated and all the voxels above 2 were removed. After these refinements of the ROIs the mean diffusion measure in each of those ROIs were used as "outcome" measures in our analyses.