Multimodal neuroimaging analysis reveals age‐associated common and discrete cognitive control constructs

Abstract

The aims of this study were to determine which cognitive control functions are most sensitive to cross‐sectional age differences and to identify neural features in different neuroimaging modalities that associated cognitive control function across the adult lifespan. We employed a joint independent component analysis (jICA) approach to obtain common networks among three different brain‐imaging modalities (i.e., structural MRI, resting‐state functional MRI, and diffusion tensor imaging) in relation to the cognitive control function. We differentiated three distinct cognitive constructs: one common (across inhibition, shifting, and updating) and two specific (shifting, updating) factors. These common/specific constructs were transformed from three original performance indexes: (a) stop‐signal reaction time, (b) switch‐cost, and (c) performance sensitivity collected from 156 individuals aged 20 to 78 years old. The current results show that the cross‐sectional age difference is associated with a wide spread of brain degeneration that is not limited to the frontal region. Crucially, these findings suggest there are some common and distinct joined multimodal components that correlate with the psychological constructs of common and discrete cognitive control functions, respectively. To support current findings, other fusion ICA models were also analyzed including, parallel ICA (para‐ICA) and multiset canonical correlation analysis with jICA (mCCA + jICA). Dynamic interactions among these brain features across different brain modalities could serve as possible developmental mechanisms associated with these age effects.

1. INTRODUCTION

Various cognitive functions deteriorate as people age, and cognitive control function is often the most prominent deficit (Salthouse, 1990; Salthouse, Atkinson, & Berish, 2003). However, recent meta‐analyses have indicated that older adults do not necessarily perform worse than younger adults in cognitive control function tasks (e.g., Rey‐Mermet & Gade, 2017; Verhaeghen, 2011; Wager et al., 2005). The implications of these reports suggest that cognitive control function may not be unitary. Recent behavioral and neuropsychological evidence has also indicated that cognitive control may be more accurately characterized as a collection of related but separable abilities (Baddeley, 1996; Collette et al., 2005; Friedman et al., 2006).

Miyake and colleagues have advocated the view of unity versus diversity of cognitive control (Friedman & Miyake, 2017; Miyake et al., 2000; Miyake & Friedman, 2012; see also Duncan, Johnson, Swales, & Freer, 1997). Using latent variable analyses, Miyake et al. (2000) identified three cognitive control1 functions that grouped different sets of experimental tasks: updating (monitoring and changing working memory contents), shifting (flexible changes between task‐sets or goals), and inhibition (overriding habitual or prepotent responses). They discovered that these three functions were moderately correlated (i.e., unity) but were separable (i.e., diversity) at the level of latent variables (Miyake et al., 2000; Miyake & Friedman, 2012). After accounting for the unity (common executive function [EF]), they found that basic processing speed was primarily related to the common EF. This reflects that speed seems to be related to all three of these cognitive control functions because it is genetically related to the common EF that they all share (Friedman et al., 2008).

Further analysis reveals that there is no unique variance left for inhibition, hence leaving no inhibition‐specific factor in common EF. Accordingly, the function of inhibition was then seen as “common” cognitive control constructs and has been shown to impact the performance in all cognitive control tasks (Friedman & Miyake, 2017; Miyake & Friedman, 2012). Recent review articles suggest that there are diverse deficits in cognitive control related to aging (see Anderson & Craik, 2017; also Gratton, Cooper, Fabiani, Carter, & Karayanidis, 2018). However, there is not yet a strong consensus with respect to the brain localization of the multiple cognitive control functions in the unity/diversity framework (see Friedman & Miyake, 2017).

Most neuroimaging studies investigate only which brain region is activated when performing tasks that measure cognitive control (e.g., Collette et al., 2005; Sylvester et al., 2003) or what areas have stronger blood‐oxygen‐level‐dependent (BOLD) signal responses during a more control‐demanding condition. These studies reveal the areas of the brain that reliably activate across participants during task performance but they do not necessarily reveal the neural areas that predict individual differences in performance. Moreover, the literature lacks information about how individual differences in brain activation patterns across different cognitive control tasks relate to behavioral performance in the factors of inhibition, updating, and shifting (Friedman & Miyake, 2017). Friedman and Miyake's group (2017) links individual differences in neuroanatomical measures to these behavioral performance measures that are tested outside an MRI scanner, which indicates that this approach may provide insights into the unity and diversity of cognitive control functions (see also Reineberg, Andrews‐Hanna, Depue, Friedman, & Banich, 2015; Smolker, Depue, Reineberg, Orr, & Banich, 2015; Tamnes et al., 2010). Thus, the primary aim of this study is to apply various behavioral tests that measure the unity and diversity of cognitive control functions in order to study individual differences in cognitive control function across the adult lifespan.

Smolker et al. (2015) noted that none of the literature has investigated the unity and diversity of cognitive control functions in healthy adult populations, especially the middle‐aged and elderly. Furthermore, neuroimaging aging studies have probed this issue using only a single brain modality, such as anatomical brain images, functional images, or white‐matter tensor‐based images (e.g., Reineberg et al., 2015). Some studies incorporate multimodal brain measures but only focus on either the structure or function of the brain (e.g., Lerman‐Sinkoff et al., 2017; Smolker et al., 2015). For example, Smolker et al. (2015) evaluated multiple neuroanatomical measures, including brain voxels, surface thickness, and white‐matter tensor‐based measures of brain structures. However, they did not use resting‐state brain functional imaging, so their interpretation of the functional brain is limited.

Lerman‐Sinkoff et al. (2017) employed a multimodal approach to identify neural correlates of cognitive control function, but their sample was mainly younger adults (22–35 years old), and their use of behavioral tasks was not able to assess the unity/diversity of the cognitive control function framework (Friedman & Miyake, 2017). Therefore, the aim of this study is to obtain multimodal brain neuroimaging data (i.e., gray matter, white matter, and functional data) and to examine neural correlates of cognitive control function based on the unity/diversity framework across the adult lifespan. The idea of applying multimodal measures by incorporating functional connectivity data is motivated by recent advances in brain imaging research, which indicate that different correlational measures in either structural or functional data could generate “connectivity” between brain regions, the overlaps of connectivity, and the whole brain organizations that are largely different between modalities. Different connectivity measures do not necessarily lead to the same underlying network structure. Therefore, caution is required when interpreting the results of joint network analysis from different imaging modalities (Di et al., 2017).

Several multivariate analysis methodologies have been developed to calculate brain network similarities between different imaging modalities, including joint independent component analysis (jICA), multiset canonical correlation analysis (mCCA), mCCA plus jICA (mCCA + jICA), and parallel ICA (para‐ICA) (e.g. Alexander‐Bloch, Raznahan, Bullmore, & Giedd, 2013; Calhoun, Adali, Hansen, Larsen, & Pekar, 2003; Calhoun, Adalı, & Pekar, 2004; Calhoun, Adali, Stevens, Kiehl, & Pekar, 2005; Clos, Rottschy, Laird, Fox, & Eickhoff, 2014; Di et al., 2017; Lerman‐Sinkoff et al., 2017; Reid et al., 2017; Sui, Adali, Yu, Chen, & Calhoun, 2012; Sui et al., 2011, 2015). In this study, we employed the jICA approach developed by Calhourn (Calhoun et al., 2003; Calhoun, Adalı, & Pekar, 2004; Calhoun, Pekar, & Pearlson, 2004; Calhoun et al., 2005; see also Sui et al., 2011, 2012, 2015) to obtain common networks in relation to cognitive control function among three different imaging modalities: structural MRI (sMRI), resting‐state functional MRI (rfMRI), and diffusion tensor imaging (DTI; dMRI).

This approach has recently been applied to general cognition to search for neuromarkers of multi‐domain cognition in individuals with schizophrenia (Sui et al., 2018). jICA is applied to different imaging modalities and extracts the spatially independent maps for each modality. These maps are coupled by a shared loading parameter: the same mixing coefficient matrix. The advantage of the jICA approach is that it provides complementary information across modalities. For example, rfMRI measures spontaneous fluctuations in hemodynamic signals, sMRI enables us to estimate the type of tissue for each voxel in the brain, and dMRI can provide additional information about the integrity of white matter tracts and structural connectivity.

jICA is based on a strong hypothesis that different modalities can be fused to yield the same mixing coefficient matrix (e.g., Calhoun et al., 2006). However, despite its popularity, some may question the validity of using the same mixing matrix across the three different modalities. To generalize the current findings on the jICA approach, we also used two other types of fusion models: mCCA + jICA and para‐ICA. These models yield different mixing coefficient matrices for each modality. Each of these analysis approaches has advantages and limitations (e.g., Calhoun & Sui, 2016; Mohammed, Taha, & Faragallah, 2014). Since this methodology is still in its infancy, we provide detailed results from other fusion models in Supporting Information Files S1 and S2 for the mCCA + jICA and para‐ICA approaches. This information allows for a comprehensive overview of aging‐associated cognitive control functions across different brain imaging modalities. As for the cognitive control behavioral indices, this study adopts the unity and diversity model (Miyake et al., 2000; Miyake & Friedman, 2012) to differentiate cognitive control function into at least three distinct constructs: a common factor (unity) and two specific factors (diversity: shifting and updating).

Three cognitive control constructs were transformed from the original cognitive control indexes: (a) stop‐signal reaction time (SSRT) derived from a stop‐signal task (Verbruggen & Logan, 2009), (b) switch‐cost derived from a task‐switching paradigm (Whitson, Karayanidis, & Michie, 2012), and (c) performance sensitivity (d’) derived from a 2‐back task (Kirchner, 1958) collected from 156 individuals aged 20 to 78 years. The statistical approach of a fusion independent‐component analysis (ICA) identifies some joint independent components (ICs) across all individuals and then correlates these ICs with age and each of the cognitive control constructs by using robust regression. The goal was to determine the cognitive control constructs that are sensitive to cross‐sectional age differences and the fusion components that are correlated to common or specific cognitive control constructs across different neuroimaging modalities. We also aimed to elucidate how individual differences in brain structures and functions may affect age‐related differences in cognitive control function.

2. MATERIALS AND METHODS

2.1. Participants

We recruited 183 participants from southern Taiwan by advertisements on the Internet and bulletin boards. All participants were assessed by the Montreal Cognitive Assessment (MoCA) to screen for probable dementia (Nasreddine et al., 2005) and the Beck Depression Inventory II (BDI‐II; Beck, Steer, & Brown, 1996) to screen for depression. Individuals with MoCA scores <22 (n = 4) and BDI‐II scores >13 (n = 7) were excluded from data analysis. Sixteen participants were further excluded because of technical MRI problems or incomplete data. The screening criteria for imaging quality control were based on head motion parameters and frame‐wise displacement (FD). None of the remaining 156 participants' max head motion exceeded 2.5 mm, or means FD exceeded 0.25. We also visual inspection of all images after normalization and coregistration steps to make sure there was no bad warping. The remaining 156 participants subjectively reported were all right‐handed and without a history of current psychological disorders, neurological disease. The mean age of the 156 participants (74 females) was 47.84 ± 1.34 (standard error; SE) years (range 20–78). The mean BDI‐II score was 5.14 ± 0.32 (SE). See Table 1 for participants' age range distribution and demographic information (see also Supporting Information Figure S1 for histograms of the individual age and the behavioral task Z‐scores). All participants provided written informed consent, and the study protocol was approved by the Research Ethics Committee of the National Cheng Kung University, Tainan, Taiwan, R.O.C. All participants were rewarded with 1,500 new Taiwan dollars (NTD) after completing the experiment, including neuropsychological tests, cognitive tests, and neuroimaging acquisition.

Table 1.

Demographic information and neuropsychological assessment scores of 156 participants

Age range (year)	n	Female	Female%	MoCA mean(SD)	BDI_II mean(SD)	Age mean(SD)
20–30	34	14	41.18	28.59(0.96)	4.97(3.66)	24.27(2.93)
30.01–40	18	9	50.00	27.39(2.00)	6.39(4.38)	33.64(2.82)
40.01–50	25	11	44.00	26.60(2.18)	6.16(3.94)	45.05(3.13)
50.01–60	34	20	58.82	26.79(1.87)	5.29(4.06)	55.34(2.97)
60.01–70	32	17	53.13	27.03(1.79)	4.38(4.30)	64.78(2.38)
70.01–80	13	3	23.08	26.23(1.96)	3.38(3.64)	73.20(2.46)

Note. SD: standard deviation; MoCA: montreal cognitive assessment; BDI‐II: Beck depression inventory II.

2.2. Cognitive control tasks

2.2.1. Inhibition: Stop‐signal task

Participants were instructed to stare at the stimulus shown on the monitor's screen and press the “z” or “/” button corresponding to the target “O” or “X” with their left and right index finger, respectively (Figure 1). The screen's background was white and the target stimulus was black. The target stimulus “O” or “X” was presented in the center of the screen for 100 ms (it was 2 cm at a visual angle of 0.64⁰). Participants were instructed to respond to the stimulus as quickly and as accurately as possible. There was a “beep” sound (1,000 Hz for 100 ms) in the background, and participants were asked to ignore this sound. In the second block of practice, participants were instructed to stop their reaction immediately when they heard an auditory stop signal. This was presented for 300 ms at a frequency of 500 Hz following the stimulus onset. They were told not to slow down their reaction to wait for the stop signal to occur.

Figure 1.

(a) Stop signal task paradigm; (b) task‐switching paradigm; (c) 2‐back test paradigm [Color figure can be viewed at http://wileyonlinelibrary.com]

The formal experiment commenced after this practice and consisted of five repeats of 140 trials (40 stop‐trials and 100 go‐trials); all of the settings and rules were as described for the second practice block. The stop‐signal delay (SSD) varied depending on the participants' response to the stop‐trials, and the SSD for each stop‐trial was selected from one of the two interleaved staircases—each starting with SSD values either 150 or 350 ms. If the participants successfully stopped, then the SSD would increase 50 ms in the next stop‐trial; otherwise, there was a decrease of 50 ms in the next stop‐trial if they failed to stop (SSD range, 0–800 ms). Note that the range of 0–800 ms of the SSD was a default lower‐ and upper‐bound. There were no trials that reached 700 ms. The mean of the SSD was 377 ms with a SE of 12 ms. The staircase procedure ensured that subject's likelihood of stopping occurred around 50% of the time. The inter‐stimulus interval (ISI) varied from 1,300 to 4,800 ms, and the stop‐signal reaction time (SSRT) was calculated by subtracting the median SSD from the median RT of the go trials (Band et al., 2003). The completion time was approximately 30 min including instruction and practice time.

2.2.2. Shifting: Task‐switching paradigm

This task‐switching paradigm was adapted and modified from an original version developed by Karayanidis, Whitson, Heathcote, and Michie (2011). The stimuli were generated using presentation software and presented on a 17‐in. monitor (1,024 × 768 resolution). A cued‐target task‐switching paradigm was used (see Figure 1). There were two cues: a cold color (e.g., blue, green) and a hot color (e.g., red, orange) associated with a “number” classification or a “letter” classification task, respectively (see Figure 1). The exact cue color was never repeated in successive trials to reduce the effects of repeating a physically identical cue. The target stimuli were drawn in a white color (background screen's color was in black). For the non‐informative cuing conditions (i.e., non‐informed task conditions), the cues were drawn in a grey color (background screen's color was in black) so as not to indicate the forthcoming task type. The target stimuli were drawn in colors, either hot (red/orange) or cold (blue/green) ones (similar to informative cue colors). Stimuli consisted of a neutral pair (e.g., #丙 or 丙%) or an incongruently mapped bivalent Chinese letter–Arabic number pair (e.g., 甲4 or 4甲). Chinese letters consisted of eight Chinese letters (first‐half:「甲」、「乙」、「丙」、「丁」; second‐half:「戊」、「己」、「庚」、 「辛」) derived from the Ten Celestial Stem system (i.e., Tiangan). Participants were asked to respond using the right and left index fingers mapped to odd/even or first‐half/second‐half for Arabic‐number and Chinese‐letter tasks, respectively. Hand‐task mapping and cue‐task mapping were counterbalanced across participants. On mixed‐task blocks, switch probability was 50% with no more than four mixed‐repeat or switch trials in succession.

Each trial consisted of a cue and a target. A cue was presented for 600 ms followed by a cue‐target interval of 1,000 ms and a target presented for 5,000 ms or until a response was made. The interval between a response to the following target was 1,600 ms. Participants were instructed to respond as quickly and as accurately as possible. Each error was followed by immediate auditory feedback, and the next trial was delayed by 1,000 ms. The mean reaction time (RT) and error rate feedback were provided after each block of trials. Prior to the formal experiment, there was a practice session containing six types of blocks: (a) one single‐task block of numbers for 16 trials; (b) one single‐task block of Chinese letters for 16 trials; (c) two mixed informed task blocks for 32 trials per block; and (d) two mixed non‐informed task blocks for 32 trials per block. The subsequent formal experiment consisted of 12 blocks: (a) two single‐task blocks of numbers for 70 trials per block; (b) two single‐task blocks of Chinese letters for 70 trials per block; (c) four mixed informed task blocks for 70 trials per block; and (d) four mixed non‐informed task blocks for 70 trials per block. The entire experiment lasted for about 30–40 min. The switch cost is calculated by subtracting the mean RT of the repeat trials in the mixed‐task blocks from the mean RT of the switch trials in the mixed‐task blocks (Draheim, Hicks, & Engle, 2016).

2.2.3. Updating: 2‐back task

Participants performed the 2‐back working memory test developed by Jaeggi, Buschkuehl, Jonides, and Perrig (2008). A three by three grid was presented to the participants in every trial, and one of the grid squares was filled with blue (Figure 1). The blue square would show up randomly in any position within the three by three grid. Participants were asked to memorize the prior two blue grid location and compare them with the current picture's blue grid location. If they were at the same place, then participants should press the “F” button with their left index finger; if they were different, participants should press the “J” button with their right index finger. The stimulus was shown for 500 ms, and participants had a 2000‐ms inter‐stimulus interval (ISI) to respond. Participants completed one block of practice with feedback as well as three blocks of the formal experiment (21 trials per block). The entire experiment lasts for about 20–30 min.

The performance sensitivity (d’) was calculated based on the hit rate (H) and false‐alarm (F) rate via the formula d’ = Z(H) − Z(F) (note: Z refers to the z score of the normal distribution). The sensitivity index, d’, indicates the degree to which a participant could discriminate a true signal from noise. Importantly, in order to let the three cognitive control indexes have the same functional meaning while delineating their relationship with age or brain structure/function, we let d’ become negative values such that the higher d’ value (i.e., less negative) would indicate worse performance; the performance trend is similar for SSRT and switch costs.

2.2.4. Computing common and specific cognitive control constructs

Following Miyake and Friedman's (2012) procedures, we calculated Z value for each participant's performance on each of the three tasks individually: stop‐signal task (inhibition/common construct), task‐switching (shifting construct), and 2‐back task (working memory updating). The three z scores were averaged for each participant to create a composite score reflecting common cognitive control construct. For each specific (diversity) construct, we regressed the targeted tasks performance against the other two tasks' performance yielding a specific residual. For example, for a shifting‐specific residual, we regressed the switch cost against SSRT and 2‐back d’; for the working memory updating residual, we regressed the 2‐back d’ against switch cost and SSRT.

2.3. Neuroimaging acquisition and analysis

2.3.1. Image acquisition

MRI images were acquired on a GE MR750 3 T scanner (GE Healthcare, Waukesha, WI) in the Mind Research Imaging center at the National Cheng Kung University. High‐resolution structural images were acquired with fast‐SPGR consisting of 166 axial slices (TR/TE/flip angle 7.6 ms/3.3 ms/12°; field of view [FOV] 22.4 × 22.4 cm²; matrices 224 × 224; slice thickness 1 mm). The entire process lasted 218 s. The resting‐state functional images were collected using an interleaved T2* weighted gradient‐echo planar imaging (EPI) pulse sequence (TR/TE/flip angle, 2,000 ms/30 ms/77°; matrices, 64 × 64; FOV, 22 × 22 cm²; slice thickness, 4 mm; voxel size, 3.4375 × 3.4375 × 4 mm). These slices covered the entire brain of each participant, and the scan time was 490 s ([number of samples + number of dummy scan] × TR = [240 + 5] × 2 = 490 s) per subject. A total of 245 volumes were acquired; the first five were dummy scans and were discarded to avoid T1 equilibrium effects. During the resting‐state functional scans, the participants were instructed to remain awake with their eyes open and stare at the white cross as shown on the screen (each scan lasted for 8 min for each participant).

Diffusion tensor imaging (DTI) was obtained with a spin‐echo‐echo planar sequence (TR/TE = 5,500 ms/62~64 ms, 50 directions with b = 1,000 s/mm², 100 × 100 matrices, slice thickness = 2.5 mm, voxel size = 2.5 × 2.5 × 2.5 mm, number of slices = 50, FOV = 25 cm, NEX = 3). Reverse DTI was also acquired for top‐up correction in the DTI preprocessing. The acquisition parameters for the reverse DTI were identical to the DTI—the only difference was that there were six directions.

2.3.2. Image preprocessing

Structural MRI

T1 images were segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid using the standard unified segmentation model (Ashburner & Friston, 2005) in SPM 8 (https://www.fil.ion.ucl.ac.uk/spm/software/spm8/). After an initial affine registration of the GM DARTEL templates to the tissue probability maps in Montreal Neurological Institute (MNI) space, DARTEL nonlinear warping of GM images was performed to the standard DARTEL GM template in MNI space (Ashburner, 2007). This was employed in the modulation step (use “non‐linear only” option) to ensure that relative volumes of GM were preserved following the spatial normalization procedure. The modulated, normalized GM images (representing GMV, voxel size of 1.5 × 1.5 × 1.5 mm) were then co‐registered with fMRI contrast via SPM 8 (for the purpose of the subsequent fusion ICA processing) resulting in a final 101 × 116 × 96 matrix with a voxel size of 2 × 2 × 2 mm. Finally, images were smoothed using SPM 8 with a 6 mm³ full width at the half‐maximum Gaussian kernel.

Resting‐state functional MRI

The fMRI data were preprocessed with SPM 8 and the Data Processing & Analysis for Brain Imaging toolbox (DPABI; Yan, Wang, Zuo, & Zang, 2016) implemented in Matlab (The MathWorks, Inc., Natick, MA). The EPI images were slice‐time corrected and realigned to correct for head motion using a rigid‐body transformation. Nuisance time series (24 motion parameters, ventricle and WM signals) were regressed out. The T1 image was co‐registered to the mean EPI image, and the T1 image was co‐registered and normalized to the MNI template. The co‐registration parameters of the T1 were applied to all functional volumes. Images were then resliced to 2  ×  2 ×  2 mm resulting in a data cube of 101 × 116 × 96 voxels. The functional data were spatially smoothed with a 6‐mm Gaussian kernel. For the resting‐state fMRI (rfMRI), we extracted the voxel‐wise mean amplitude of low‐frequency fluctuations (ALFF; Zou et al., 2008)2 to generate a map for each participant. The ALFF calculation computed the fast Fourier transform (FFT) of each voxel time series taking the square root of the power spectrum to obtain amplitudes. Amplitudes were averaged in 0.01–0.08 Hz (see Zou et al., 2008 for details).

Diffusion MRI

All diffusion‐weighted imaging (DWI) data processing and analyses used the FMRIB Software Library (FSL v5.0.9; http://www.fmrib.ox.ac.uk/fsl; Smith et al., 2004). The DWIs were first converted from the DICOM to NIFTI format via the MRIcron dcm2nii tool (https://www.nitrc.org/projects/mricron/). TOPUP (Andersson, Skare, & Ashburner, 2003; Smith et al., 2004) and EDDY (Andersson & Sotiropoulos, 2016) were used to clean the DWIs of artifacts caused by susceptibility‐induced distortions, eddy currents, and head motion. A single image without diffusion weighting (b0; b value = 0 s/mm²) was extracted from the concatenated data, and non‐brain tissue was removed via the FMRIB brain extraction tool (Smith, 2002) to create a brain mask for subsequent analyses. DTIFIT (Behrens et al., 2003) was applied to fit a tensor model (forming diffusion tensor imaging, DTI) at each voxel of the data (Smith et al., 2004) to derive fractional anisotropy (FA), mean diffusivity (MD), and radial diffusivity (RD) measurements. To perform tract‐based investigations into the DTI measurements, we used tract‐based spatial statistics (TBSS) in FSL (Smith et al., 2006). The FA images were slightly eroded, and end slices were zeroed to remove likely outliers from the diffusion tensor fitting. The images were then nonlinearly aligned to each other, and the most representative image was then identified. This target image was subsequently affine‐transformed to a 1 mm MNI space. FA images were transformed to 1 mm MNI space by combining the nonlinear and affine registration, and then co‐registered with fMRI contrast via SPM8 resulting in a final 101 × 116 × 96 matrix with a voxel size of 2 × 2 × 2 mm. Finally, the images were smoothed using SPM 8 with a 6 mm³ full width at the half‐maximum Gaussian kernel.

2.4. Joint ICA analysis

We used the volumetric GM images (sMRI), ALFF maps (rfMRI), and FA maps (dMRI) as fusion input. Data were preprocessed with the Fusion ICA Toolbox (FIT, Calhoun et al., 2006 http://mialab.mrn.org/software/fit/index.html) implemented in Matlab. After feature extraction, the 3D image of each participant was reshaped into a one‐dimensional non‐zero vector and stacked individually to form a matrix with dimensions of 156 × [number of voxels] for each imaging modality. The feature matrix was then normalized to have the same average sum‐of‐squares (computed across all participants and all voxels for each imaging modality). This normalization was needed because the raw data had different ranges of values. A single normalization factor was used for each data type; thus, following normalization, the relative scaling within a given data type was preserved, but the units between data types were the same (in a least‐squares sense). After normalization, the data were processed via dimension reduction ➔ joint ICA ➔ component analysis. The component number was estimated using the modified minimum description length (MDL) criteria (Li, Adalı, & Calhoun, 2007). We choose M = 8 for the following analysis. Principal component analysis (PCA) was used to reduce the dimensionality of the data. The infomax algorithm (Bell & Sejnowski, 1995) decomposed the reduced feature matrix to maximally independent component images and subject‐specific mixing (loading) parameters.

For visualizing the spatial maps of sMRI, rfMRI, and dMRI, each component was transformed into a Z map by dividing its SD across all voxels. The use of Z‐scores has linearly transformed data values and standardized distributions, so that each component has a mean of zero and a SD of 1. Hence, a negative value does not indicate any kind of “deactivation,” it just indicates that the original value in the map. Particularly, the negative value represents it was below the mean before it was standardized. We, therefore, can only infer about its relative place in the distribution (Itahashi et al., 2015). The spatial maps were then threshold with ∣Z∣ > 2.5. To display rfMRI and sMRI maps (see Figures in the Results Section), each component's clusters were converted from MNI coordinates to Talairach coordinates and entered into a database to provide anatomic and functional labels for the right (R) and left (L) hemispheres. The volume of activated voxels in each area is provided in cubic centimeters (cc). The maximum Z value and its coordinates are provided in Supporting Information tables for each area. To display the dMRI clusters, we used the Johns Hopkins white‐matter tractography atlas provided in FSL from which 20 tract structures were identified—these were mostly large bundles (Hua et al., 2008). These structures include the anterior thalamic radiation (ATR) L/R, cingulum/cingulate gyrus (CG) L/R, cingulum/hippocampus (CH) L/R, corticospinal tract (CST) L/R, forceps major (Fmaj), forceps minor (Fmin), inferior fronto‐occipital fasciculus (IFF) L/R, inferior longitudinal fasciculus (ILF) L/R, superior longitudinal fasciculus (SLF) L/R, superior longitudinal fasciculus/temporal (SLFT) L/R, and uncinate fasciculus (UF) L/R.

2.5. Statistical analysis and component selection

2.5.1. Robust regression and component selection

Regression analyses were performed between each set of mixing coefficients of the components and (a) age as well as (b) each of the three cognitive control constructs (one common and two specific residuals) using robust regression with 3,000 bootstrap iterations (NCSS, LLC. Kaysville, Utah; http://ncss.com/software/ncss) to reduce the problems caused by multiple comparisons and outliers. Only the components that were significantly correlated with either age or cognitive control scores were subsequently interpreted.

2.5.2. Interpretation of identified positive and negative contribution to cognitive control constructs

In our findings, we demonstrated both positive and negative patterns/contributions in these identified components. It is clear that fusion ICA spatial maps are often oriented to show strong positive clusters, but sometimes the main clusters are shown as negative areas, and/or ICs contain both strong negatives and positives values. Negative values indicate that the voxel/cluster/region appears with negative intensity. For the fusion of different imaging modalities, the output is a set of weights corresponding to the contribution of each spatial and temporal feature in the model. A weight of zero indicates that a particular feature is not used in the model, and the largest weight values (both positive and negative) indicate features that are most informative in identifying the component type. In this study, the multimodal brain imaging data and cognitive control constructs were weighted, thus even for identified linked joint ICs, certain highly weighted cognitive control constructs (i.e., one common and two specific residuals) or brain features may contribute more or less. They can also have negative or positive values, thus a brain feature may be negatively associated with (or subtracting from) the overall cognitive control constructs. Likewise, given cognitive control constructs may make a positive or negative contribution to the overall joint ICs. By using this concept (see Pearlson, Liu, & Calhoun, 2015), our negative value of IC brain image indicates negative association among FA‐GM‐ALFF in this region, whereas a positive value indicates positive associations among FA‐GM‐ALFF in the region.

3. RESULTS

3.1. Behavioral performance

3.1.1. Stop‐signal task performance: SSRT

The mean stop inhibition rate (stop success rate) was 53.38% (mean) ± 0.15% (SE), and the mean SSRT was 256.36 ± 14.58 ms. The SSRT was significantly correlated with age (r = 0.35; p < 0.00001). It was not correlated with mean RT of the correct go‐trials (r = −0.09; Go trials' RT: 658.61 ± 12.13 ms; mean accuracy rate: 90.39 ± 3.3%), which is consistent with the “horse‐race” model that assumes the independence of the process between go‐trials and stop‐trials.

3.1.2. Task‐switching performance: Switch‐cost

The mean switch‐cost was 79.22 ± 9.84 ms—this is not significantly correlated with age (r = −0.11; p = 0.17). The mean accuracy rate for all repeat trials in the single blocks was 98.01 ± 1.52%, repeat trials in the mixed blocks was 95.82 ± 2.32%, and switch trials in the mixed blocks were 92.75 ± 3.21%.

3.1.3. 2‐back task performance: d’

The overall mean accuracy for the 2‐back task was 80.71 ± 12.73%. The results of the 2‐back tasks are reported based on the signal detection theory. For the 2‐back task, the mean sensitivity (d’) was (−)1.93 ± 0.88. There was a significant correlation between age and the 2‐back d’ (d’ in a negative value; r = 0.48; p < 0.01) suggesting a worse 2‐back task performance with increasing age.

3.1.4. Summary of the behavioral results

The behavioral results showed that inhibitory (reflected by SSRT) and working memory updating (reflected by 2‐back d’ in a negative value) decayed with increasing age; shifting ability (reflected by switch cost) did not decay with age.

3.2. Cognitive control constructs based on Miyake et al.'s model: Common (inhibition) and specific (shifting and updating)

Here, following Miyake et al.'s (2000) model, we calculated a cognitive control common and two specific constructs: shifting and updating. The common EF and updating‐specific residuals significantly correlated with age (common EF: r = 0.41, 95% CI: 0.4397 ~ 0.8645; updating‐specific residual: r = 0.43, 95% CI: 0.2846 ~ 0.5852) but not the shifting‐specific residual in correlation with age (r = −0.09, 95% CI: −0.2362 ~ 0.1440).

3.3. Independent components (ICs) and cognitive control constructs

Of the eight ICs, we selected only five ICs for further processing because the other three ICs' spatial maps containing obvious artifacts had sharp edges at the brain boundary or within the cerebrospinal fluid (CSF) region (Griffanti et al., 2017). The five sets of IC's mixing coefficients—each of which represents the relative degree to which an individual participant contributes to the joint component—were robustly regressed with participant's ages and one common and two specific cognitive control function residuals: shifting and updating‐specific residuals. Additional analysis was also conducted on inhibition‐specific residuals, results showed highly similar patterns between common and inhibition‐specific residuals (see Supporting Information Table S7). All robust regression standardized coefficient r values between each of the ICs and variables are listed in Table 2.

Table 2.

Robust regression r values for the mixing coefficients of five joint independent component analysis (jICA) components (IC#) with age and cognitive control common and specific (shifting and updating) residuals

Component	Age	Common	Shifting‐specific	Updating‐specific
IC #3	−0.51*	−0.55*	−0.22*	−0.32*
IC #4	−0.42*	−0.05	0.19	−0.37*
IC #5	0.10	0.13	0.17*	−0.07
IC #6	0.54*	0.35*	0.03	0.18
IC #8	0.26*	0.07	0.00	0.15

^* bootstrap upper and lower bound do not pass 0.

3.4. ICs, age, and cognitive control constructs

3.4.1. ICs and age factor

The mixing coefficient of the ICs (i.e., #3, #4, #6, #8) except #5 showed a significant correlation with age. In addition, the mixing coefficients of the joint IC #3 not only exhibited an age effect but also correlated with all other cognitive control indexes including one common and two specific residuals (see the next section for details). It can be interpreted as IC #3 had greater contribution to these age‐associated cognitive control constructs among these joint ICs. Therefore, we reported the results of the IC #3 in relation to the age factor in details. The spatial maps were transformed into Z values visualized at |Z| > 2.5 (Figure 2).

Figure 2.

(a) The spatial maps of IC #3 for amplitude of low‐frequency fluctuation (ALFF), gray matter (GM), and fractional anisotropy (FA) tensor‐based white matter (WM) tract, respectively. For display, the spatial maps were transformed into Z values, visualized at |Z| > 2.5. (b) The scatterplots and robust regression fitting lines for IC #3 in relation to age, common and two specific residuals (shifting and updating), respectively [Color figure can be viewed at http://wileyonlinelibrary.com]

Independent component #3

The specific identified regions in IC #3 are summarized in Supporting Information Table S1 for rfMRI (Talairach labels for ALFF), Supporting Information Table S2 for sMRI (Talairach labels for GMV), and Supporting Information Table S3 for dMRI (Johns Hopkins white‐matter tractography labels for FA tensor‐based map).

The positive contributing ALFF regions in IC #3 included postcentral gyrus, precuneus, paracentral lobule, inferior parietal lobule, superior parietal lobule, posterior cingulate, inferior frontal gyrus, superior temporal gyrus, cuneus, and fusiform gyrus, whereas the negative contributing regions included superior frontal gyrus, medial and middle frontal gyrus, precentral, lingual gyrus, superior temporal gyrus, and cuneus. (Figure 2b; Supporting Information Table S1).

The positive contributing GM regions in the IC #3 included precentral gyrus, cuneus, inferior frontal gyrus, inferior parietal lobule, superior temporal gyrus, fusiform gyrus, middle frontal gyrus, inferior temporal gyrus, lingual gyrus, superior parietal lobule, postcentral gyrus, precuneus, parahippocampal gyrus, middle temporal gyrus, cingulate gyrus, posterior cingulate, middle occipital gyrus, superior frontal gyrus, insula, angular gyrus, anterior cingulate, medial frontal gyrus, transverse temporal gyrus, and thalamus. (Figure 2b; Supporting Information Table S2). The negative contributing GM regions in the IC #3 included lentiform nucleus, superior temporal gyrus, middle temporal gyrus, and inferior parietal lobule.

The positive contributing WM tracts in IC #3 included ATR L/R, CG L/R, CST L/R, Fmaj, Fmin, IFF L/R, ILF L/R, SLF L/R, and SLFT L/R (Figure 2b; Supporting Information Table S3).

Interaction of modalities among rfMRI, sMRI, and dMRI in IC #3

Figures 3a,b summarized the main results across different modalities. Our goal is to evaluate the intersection of the results with (a) known tracts for dMRI and (b) known regions for fMRI and sMRI, as well as to identify which known tracts both intersect with the regions of FA changes and touch the regional changes. Thus, we present the significant ALFF and GM regions that are adjacent to significant WM tracts.

Figure 3.

(a) Interaction conceptual illustration between IC #3's WM tracts (colored line) and ALFF region (blue square); (b) Interaction conceptual illustration between IC #3's WM tracts (colored line) and GM regions (yellow square). Regions with colored square indicate that a major portion of it is significant; (c) Overlapping (red) between IC #3's ALFF (blue) and GM regions (yellow); x, y, z = −3, −68, 10. Different colored line indicates different WM tracts. MFG: middle frontal gyrus; PRG: precentral gyrus; POG: postcentral gyrus; CC: cingulate cortex; pCC: posterior cingulate cortex; PH: parahippocampal gyrus; STG: superior temporal gyrus [Color figure can be viewed at http://wileyonlinelibrary.com]

We found an ALFF cluster in the posterior cingulate cortex (PCC) adjacent to the WM tract of CG) (Figure 3a). There were a few GM clusters at the cingulate cortex surrounded by a cingulum bundle of CG (Figure 3b). The WM tract of the CST was projected from the precentral and postcentral gyrus, while the ATR consists of fibers between the thalamus and the frontal cortex and was adjacent to the parahippocampal gyrus (Figure 3b). There were some clusters at the cuneus surrounded by the WM tract of Fmaj in both GM and ALFF modalities (Figure 3a,b). We further overlapped the different modalities of IC #3 to find connections between functional regions (fMRI) and structural regions (sMRI) of the brain (Figure 3c). Most overlapping regions between ALFF and GM were located in the PCC region, which is closer to the CG tract (Figure 3C).

Comparing IC #3 with the visual network

The brain regions involved in IC #3 appear to comprise some critical regions of the visual network. We thus overlapped this component (GM and ALFF feature spatial maps) to the visual network template provided by Yeo et al. (2011). The results showed that the ALFF feature map of IC #3 overlapped with the visual network mostly at clusters of cuneus and fusiform gyrus (Figure 4a). The GM feature map of IC # 3 also overlapped with the visual network at the clusters of the cuneus and fusiform gyrus (Figure 4b).

Figure 4.

(a) Overlapping (red) between IC #3's ALFF (blue) regions and the visual network (cyan); x y z = −5 −74 10; (b) Overlapping (red) between IC #3's GM (yellow) regions and the visual network (cyan); x y z = −36 −76 9 [Color figure can be viewed at http://wileyonlinelibrary.com]

Independent component #8

The IC #8 was found to be only significantly associated with age, but not cognitive control constructs (Table 2). The positive contributing ALFF regions in IC #8 included superior temporal gyrus, inferior frontal gyrus, insula, anterior cingulate, transverse temporal gyrus, middle temporal gyrus, lingual gyrus, precentral gyrus, precuneus, cingulate, and parahippocampal gyrus (Figure 5). The negative contributing areas included precuneus, and inferior parietal lobule.

Figure 5.

(a) The spatial maps of IC #8 for amplitude of low‐frequency fluctuation (ALFF), gray matter (GM), and fractional anisotropy (FA) tensor‐based white matter (WM) tract, respectively. For display, the spatial maps were transformed into Z values, visualized at |Z| > 2.5. (b) The scatterplots and robust regression fitting lines for IC #8 in relation to age, common and two specific residuals (shifting and updating), respectively [Color figure can be viewed at http://wileyonlinelibrary.com]

The positive contributing GM regions in IC #8 included inferior temporal gyrus, cuneus, middle temporal gyrus, cuneus, lentiform nucleus, superior temporal gyrus, fusiform, inferior parietal lobule, supramarginal gyrus, superior occipital gyrus, middle occipital gyrus, posterior cingulate, precuneus, thalamus, middle frontal gyrus, superior parietal lobule, parahippocampal gyrus, insula, caudate, cingulate gyrus, postcentral gyrus, precentral gyrus, medial frontal gyrus, paracentral lobule, lingual gyrus, and superior frontal gyrus. The negative contributing regions included inferior temporal gyrus, middle frontal gyrus, superior frontal gyrus, superior parietal lobule, middle temporal gyrus, cingulate gyrus, superior temporal gyrus, postcentral gyrus, inferior parietal lobule, precuneus, medial frontal gyrus, middle occipital gyrus, inferior frontal gyrus, precentral gyrus, paracentral lobule, lingual gyrus, cuneus, and angular gyrus.

The positive contributing FA tensor‐based WM tracts in IC #8 included CH L, and UF R.

Interaction of modalities among rfMRI, sMRI, and dMRI in IC #8

Our goal is to examine the relationship between multimodal brain features and their correlations with cognitive control functions. However, IC #8 did not show significant associations with any of the cognitive constructs. Therefore, we do not discuss these results further among rfMRI, sMRI, and dMRI in the following sections.

3.4.2. ICs and cognitive control constructs

Independent component #3

We next to determine which IC might be significantly related to the common and specific constructs (see Table 2). The mixing coefficients of IC #3 showed significant correlations with common and the two specific residuals. Figure 2a demonstrates the scatter plots of cognitive control constructs versus the mixing coefficient of IC #3 as well as linear trends with robust regression. Figure 2b displays IC #3's spatial maps of ALFF (see also Supporting Information Table S1), GM (Supporting Information Table S2), and FA tensor‐based WM tracts (Supporting Information Table S3). The details of IC #3 spatial maps and their overlapping regions can be found in the previous section regarding the effect of age.

Independent component #4

In contrast to IC #3, IC #4 showed significant correlation only with age and updating‐specific residuals. Figure 6b demonstrates the scatter plots of age and updating‐specific residuals versus IC 4's weights; there is a linear trend. Figure 6a displays IC 4's spatial maps for ALFF (Supporting Information Table S4), GM (Supporting Information Table S5), and FA tensor‐based WM tracts (Supporting Information Table S6).

Figure 6.

(a) The spatial maps of IC #4 for amplitude of low‐frequency fluctuation (ALFF), gray matter (GM), and fractional anisotropy (FA) tensor‐based white matter (WM) tract, respectively. For display, the spatial maps were transformed into Z values, visualized at |Z| > 2.5. (b) The scatterplots and robust regression fitting lines for IC #4 in relation to age, common and two specific residuals (shifting and updating), respectively [Color figure can be viewed at http://wileyonlinelibrary.com]

The positive contributing ALFF regions in IC #4 included superior temporal gyrus, medial frontal gyrus, postcentral gyrus, and inferior semi‐lunar lobule (see Figure 7b; Supporting Information Table S4).

Figure 7.

(a) Interaction conceptual illustration between IC #4's WM tracts (colored line) and ALFF regions (blue square); (b) Interaction conceptual illustration between IC #4's WM tracts (colored line) and GM regions (yellow square). Regions with colored square indicate that a major portion of it is significant; (c) Overlapping (red) between IC #4's ALFF (blue) and GM regions (yellow); x, y, z = 35, 9, 10. Different colored line indicates different WM tracts. MFG: middle frontal gyrus; PRG: precentral gyrus; POG: postcentral gyrus; CC: cingulate cortex; pCC: posterior cingulate cortex; PH: parahippocampal gyrus; STG: superior temporal gyrus; MeFG: medial frontal gyrus; MTG: middle temporal gyrus [Color figure can be viewed at http://wileyonlinelibrary.com]

The positive contributing GM regions in IC #4 included middle temporal gyrus, middle frontal gyrus, superior temporal gyrus, fusiform gyrus, angular gyrus, inferior frontal gyrus, superior frontal gyrus, parahippocampal gyrus, precuneus, precentral gyrus, superior parietal lobule, cingulate gyrus, middle occipital gyrus, medial frontal gyrus, insula, posterior cingulate, insula, and inferior parietal lobule (see Figure 7b; Supporting Information Table S5). The negative contributing regions included thalamus, precuneus, lingual gyrus, cuneus, inferior frontal gyrus, posterior cingulate, middle temporal gyrus, precentral gyrus, postcentral gyrus, fusiform gyrus, insula, superior temporal gyrus, inferior parietal lobule, middle occipital gyrus, and anterior cingulate.

The positive contributing FA tensor‐based WM tracts in IC #4 included ATR L/R, whereas the negative contributing tracts included CH L/R, Fmaj, IFF L/R, ILF L, and CST L/R (see Figure 7b; Supporting Information Table S6).

Interaction of modalities among rfMRI, sMRI, and dMRI in IC #4

Figures 7a,b summarized the main results across different modalities. We overlapped the rfMRI and sMRI results of IC#4 to find the relations between brain functional activity (ALFF) and brain (GM) regions (Figure 7c). The relations between the WM tracts and the GM of IC #4 show that the CST is projected from the precentral and passes through the parahippocampal gyrus (Figure 7b). There is also a marginal overlapping region between the ALFF and GM maps in the superior temporal gyrus (STG), as shown in Figure 7c.

Independent component #5

IC #5 was only significantly associated with the shifting‐specific residual. The scatterplots and brain maps of this component are shown in Figure 8b. The positive contributing ALFF brain regions in IC #5 included cuneus, posterior cingulate, lingual gyrus, precuneus, fusiform, precentral gyrus, parahippocampal gyrus, middle, and inferior occipital gyrus (Figure 8a), whereas negative contributing areas included medial frontal gyrus, middle frontal gyrus, and precuneus.

Figure 8.

(a) The spatial maps of IC #5 for amplitude of low‐frequency fluctuation (ALFF), gray matter (GM), and fractional anisotropy (FA) tensor‐based white matter (WM) tract, respectively. For display, the spatial maps were transformed into Z values, visualized at |Z| > 2.5. (b) The scatterplots and robust regression fitting lines for IC #5 in relation to age, common and two specific residuals (shifting and updating), respectively [Color figure can be viewed at http://wileyonlinelibrary.com]

The positive contributing GM regions in IC #5 included fusiform, precuneus, middle occipital gyrus, middle temporal gyrus, postcentral gyrus, lentiform nucleus, superior temporal gyrus, medial frontal gyrus, lingual gyrus, inferior parietal lobule, supramarginal gyrus, superior frontal gyrus, middle frontal gyrus, paracentral lobule, caudate, cuneus, paracentral lobule, parahippocampal gyrus, insula, inferior temporal gyrus, inferior frontal gyrus, inferior occipital gyrus, superior parietal lobule, anterior cingulate, and angular gyrus (Figure 8a). The negative contributing regions included middle temporal gyrus, inferior parietal lobule, middle occipital gyrus, cuneus, inferior temporal gyrus, middle frontal gyrus, fusiform gyrus, lingual gyrus, superior temporal gyrus, precuneus, superior frontal gyrus, inferior frontal gyrus, medial frontal gyrus, precentral gyrus, superior parietal lobule, postcentral gyrus, cingulate gyrus, and posterior cingulate (Figure 8a, the middle panel).

The positive contributing FA tensor‐based WM tracts in IC #5 included ATR L/R, and Fmaj (Figure 8a; bottom panel).

Interaction of modalities among rfMRI, sMRI, and dMRI in IC #5

Figures 9a,b summarized the main results across different modalities. There are GM and ALFF clusters near the cuneus region that are connected with the WM tract of Fmaj (Figure 9a,b). There is a GM caudate connected with the WM tract of the ATR (Figure 9b). There are also overlapping regions located in the precuneus and cuneus regions between ALFF and GM (Figure 9c).

Figure 9.

(a) Interaction conceptual illustration between IC #5's WM tracts (colored line) and ALFF regions (blue square); (b) Interaction conceptual illustration between IC #5's WM tracts (colored line) and GM regions (yellow square). Regions with colored square indicate that a major portion of it is significant; (c) Overlapping (red) between IC #5's ALFF (blue) and GM regions (yellow); x, y, z = −11, −73, −3. Different colored line indicates different WM tracts. MFG: middle frontal gyrus; PH: parahippocampal gyrus [Color figure can be viewed at http://wileyonlinelibrary.com]

Independent component #6

IC #6 is significantly related to age and common cognitive control (see Table 2). The positive contribution ALFF areas in IC #6 included middle frontal gyrus, superior frontal gyrus, medial frontal gyrus, inferior occipital gyrus, lingual gyrus, fusiform gyrus, and superior temporal gyrus (Figure 10a), whereas negative contributing areas included superior temporal gyrus, and parahippocampal gyrus (Figure 10a, the top panel).

Figure 10.

(a) The spatial maps of IC #6 for amplitude of low‐frequency fluctuation (ALFF), gray matter (GM), and fractional anisotropy (FA) tensor‐based white matter (WM) tract, respectively. For display, the spatial maps were transformed into Z values, visualized at |Z| > 2.5. (b) The scatterplots and robust regression fitting lines for IC #6 in relation to age, common and two specific residuals (shifting and updating), respectively [Color figure can be viewed at http://wileyonlinelibrary.com]

The positive contributing GM regions in IC #6 included caudate, thalamus, middle frontal gyrus, inferior occipital gyrus, inferior temporal gyrus, inferior frontal gyrus, superior temporal gyrus, middle temporal gyrus, fusiform gyrus, superior frontal gyrus, precuneus, angular gyrus, precentral gyrus, inferior parietal lobule, medial frontal gyrus, middle occipital gyrus, lingual gyrus, superior parietal lobule, rectal gyrus, supramarginal gyrus, cingulate gyrus, postcentral gyrus. The negative contributing regions included inferior frontal gyrus, transverse temporal gyrus, middle frontal gyrus, lentiform nucleus, lingual gyrus, superior temporal gyrus, middle temporal gyrus, anterior cingulate, precuneus, parahippocampal gyrus, cuneus, precentral gyrus, inferior temporal gyrus, postcentral gyrus, insula, middle occipital gyrus, and posterior cingulate (Figure 10a, the central panel).

The positive contributing FA tensor‐based WM tracts in IC #6 included CST L/R, whereas negative contributing tracts included ATR L/R, CG L, Fmin, and UF L (Figure 10a, the bottom panel).

Interaction of modalities among rfMRI, sMRI, and dMRI in IC #6

Figures 11a,b summarized the main results across different modalities.. We overlapped the rfMRI and sMRI results of IC #6 to display the connections between functional activity and regions of the brain (Figure 11c). The WM tracts and ALFF for IC #6 show that the ALFF clusters are mainly located in the frontal lobe, which is connected with the ATR tract (Figure 11a). The relations between the WM tract and the GM show that there are many GM clusters located in the subcortical region, such as the thalamus and caudate. These are surrounded by the WM tract of the CST and the medial ATR tracts (Figure 11b). The overlap between ALFF and GM is located in the bilateral middle frontal gyrus (Figure 11c).

Figure 11.

(a) Interaction conceptual illustration between IC #6's WM tracts (colored line) and ALFF regions (blue square); (b) Interaction conceptual illustration between IC #6's WM tracts (colored line) and GM regions (yellow square). Regions with colored square indicate that a major portion of it is significant; (c) Overlapping (red) between IC #6's ALFF (blue) and GM regions (yellow); x, y, z = −14, 2, −16. Different colored line indicates different WM tracts. MFG: middle frontal gyrus; PRG: precentral gyrus; POG: postcentral gyrus; CC: cingulate cortex; pCC: posterior cingulate cortex; PH: parahippocampal gyrus; STG: superior temporal gyrus; MeFG: medial frontal gyrus; MTG: middle temporal gyrus; SFG: superior frontal gyrus; ITG: inferior temporal gyrus [Color figure can be viewed at http://wileyonlinelibrary.com]

4. DISCUSSION

The aims of this study were to empirically and systematically examine which cognitive control functions are most sensitive to age differences and to identify the neural features that correlate to cognitive control functions across the adult lifespan via different neuroimaging modalities. Prior research has indicated that cognitive control is not entirely a unitary function but is rather a collection of processes (Duncan, Humphreys, & Ward, 1997; Godefroy, Cabaret, Petit‐Chenal, Pruvo, & Rousseaux, 1999; Shallice & Burgess, 1996; Stuss & Alexander, 2007; Teuber, 1972). We examined the model of unity versus diversity of cognitive control functions proposed by Miyake et al. (2000) to guide the investigation of the relationship among multimodal neuroimaging data that arise from independent but partially correlated subcomponents (diversity) (see also Friedman & Miyake, 2017; Miyake & Friedman, 2012).

For the first aim, we observed that not all cognitive control functions deteriorated as a function of age as predicted by the unitary theory of cognitive control deficit in aging. Examples include inhibition account (Hasher & Zacks, 1988; see also Braver & Barch, 2002), working memory capacity (Craik & Byrd, 1982; see also Gordon, Tse, Gratton, & Fabiani, 2014), and speed of processing (Salthouse, 2000). Of the three types of cognitive control functions examined in this study, only stopping efficiency (as indexed by SSRT) and working memory (as indexed by d’ of the 2‐back task) were sensitive to age, but not shifting ability (as indexed by the RT switch cost). The results remained the same after transforming the original task performance into three specific residuals (unity versus diversity of cognitive control based on the executive function framework developed by Miyake et al. (2000) and Miyake and Friedman (2012)).

Only the common/inhibition factor and the updating‐specific residuals were sensitive to age, but not to the shifting‐specific residual. Thus, it is not clear why these two cognitive constructs of common/inhibition and updating were more sensitive to the age factor. One possible explanation is that the stopping test assesses inhibitory control, especially motor response inhibition, which might be more sensitive to individual differences in the intrinsic functional architecture of the brain, including age differences and specifically those related to inhibition. On the other hand, the 2‐back task is a measure of working memory, which has been repeatedly shown to decline with age, as in our results (Babcock & Salthouse, 1990; Hasher & Zacks, 1988; Li, Lindenberger, & Sikström, 2001). It should be noted that our interpretation of the effect of age is strictly based on differences between cohorts and individuals but does not refer to biological aging given the absence of longitudinal data (Goh, An, & Resnick, 2012).

Regarding the second aim, we observed three main findings. First, there are widespread structural and functional differences in the brain (i.e., ICs #3, #4, #6, #8) across the adult lifespan, which is by no means restricted to the frontal lobe as proposed by the frontal lobe aging theory (e.g., Craik, Morris, Morris, & Loewen, 1990; Dempster, 1992; West, 1996). Second, there is one component (IC #3) that is sensitive to age and all cognitive control constructs, including one common and two specific constructs (i.e., shifting‐ and updating‐specific). Lastly, there are other components that are sensitive to age but only certain cognitive control constructs. Specifically, IC #4 is sensitive to age and only to the updating‐specific construct, IC #5 is only sensitive to the shifting‐specific construct but not to the other constructs, and IC #6 is sensitive to both age and the common cognitive control construct. IC #8 is only sensitive to age (see Table 2), which suggests neural correlates of common and specific cognitive control constructs.

4.1. ICs and age differences

The multimodal neuroimaging results suggest that the effect of age difference is not limited to only the frontal lobe region, as previously suggested (Miller & Cohen, 2001; Miyake et al., 2000; West, 1996). Rather, it is more widespread and includes non‐frontal brain regions, such as the parietal, temporal, occipital, cuneus, insula, and parahippocampal gyrus regions. In addition, age differences are also associated with decreased FA in several WM tracts, such as the CG, CST, and forceps major/minor.

The results did not support a strong version of the frontal aging theory (Craik et al., 1990; Dempster, 1992; Fabiani & Friedman, 1997; Hartley, 1993; Shimamura & Jurica, 1994; West, 1996) but rather a more widespread degeneration in aging3 (Greenwood, 2000). However, it is worth noting that despite the widespread neural differences with age, it is not necessarily associated with impaired cognitive control function as a whole. It only implies impaired common/inhibition and updating‐specific cognitive control functions.

4.2. IC #3 and cognitive control constructs

The increasing coefficients of IC #3 are the most sensitive multimodal signature that is associated with all cognitive control constructs (a lower score indicates better performance). Specifically, a reduction in the WM integrity of CST was observed in this component and revealed a strong association with all cognitive control constructs. This tract has been evident as a mediator for controlling voluntary movements, and the neurons of the CST mainly arise from the sensory‐motor system and parietal region. Moreover, several parietal and frontal regions are consistently reported in both GM and ALFF modalities, particularly the precuneus and middle frontal gyrus. Previous studies suggest that age‐related alterations of these parietal–frontal regions are linked to executive function and dysfunction (Kievit et al., 2014), which is consistent with our findings of an age‐related association with cognitive control constructs.

It is worthwhile to determine why this component exhibits the greatest association with age and all constructs of the cognitive control function. We speculate that the dynamic interactions across different modalities have a crucial role in unraveling the underlying mechanism. For example, other GM regions and WM tracts of the brain in IC #3 show that the GM cuneus and some parts of the cingulate cortex are connected with WM tracts of the CG and forceps major. In addition, the WM tract of the CST was found to be adjacent to the GM parahippocampal gyrus (see Supporting Information Table S1). The functional ALFF in the PCC is connected with the WM tract of the CG (see Figure 3a).

The GM and WM tracts of IC #3 appear to form similar connectivity to the visual resting‐state network that is derived from resting fMRI (Yeo et al., 2011). The major brain region in the visual network is the cuneus and is found in the occipital lobe of the human brain beneath the parieto‐occipital fissure. This is where the parietal and occipital lobes meet above and within the calcarine fissure located in the lower section of the occipital lobe. The cuneus (synonymous with Brodmann area 17) is involved in processing visual information and is a part of the dorsal and ventral visual streams. Therefore, we reason that visual function is critical for a wide range of cognitive control tasks and is also the most sensitive to aging compared to other neural systems.

Our primary findings again challenge the strong theory of frontal aging (Craik et al., 1990; Dempster, 1992; Fabiani & Friedman, 1997; Hartley, 1993; Shimamura & Jurica, 1994; West, 1996). Furthermore, IC #3 involves the CG tract, and the literature has shown that the cingulum bundle, notably its dorsal/anterior portions, mediates performance in “frontal” tests of cognitive control and executive function (Bettcher et al., 2016). In addition, IC #3 also involves the forceps major, which is also known as the posterior forceps, a fiber bundle that connects the occipital lobes and crosses the midline via the splenium of the corpus callosum. Previous studies have shown that this callosal visual tract is associated with age‐related decline (Fjell, Sneve, Grydeland, Storsve, & Walhovd, 2017). We extend the current understanding of age‐related brain alternations by linking these multimodal visual functions network to cognitive control constructs. Thus, IC #3 is suggested to be the most sensitive component and is not only susceptible to age‐related decline but also associated with all cognitive control constructs.

4.3. IC #6 and common construct

In contrast to IC #3, IC #6 appears to occur in the ventral pathway (i.e., the frontal‐temporal regions) in both ALFF and GM modalities. Specifically, regions in the frontal lobe are particularly highlighted in ALFF, including the medial, middle, and superior parts of the frontal gyrus, as well as the temporal lobe, such as the superior temporal gyrus. The results in GM regions were consistent with ALFF in the SFG, middle temporal gyrus (MTG), and STG. Previous studies have reported that these frontotemporal regions are related to cognitive decline. Miyake and colleagues argue that basic processing speed is primarily related to the common EF, reflecting that speed seems to be related to all three of these cognitive control functions because it is genetically related to the common EF that they all share (Friedman et al., 2008). A previous study showed that processing speed training for the elderly alters the GM volume in the STG and around the occipitotemporal regions (Takeuchi et al., 2011), which is in line with our results in both GM and ALFF modalities and in concert with Miyake's model.

Another major component of Miyake's common EF is inhibitory control. Coincidently, extensive overlapped frontal regions in both ALFF and GM modalities are known to be related to inhibitory control processing (Ridderinkhof, Ullsperger, Crone, & Nieuwenhuis, 2004), which reflects that these motor inhibition regions might suffer the most from age‐related decline. In the WM tracts, the ATR appears to connect the frontal lobe and several subcortical regions, including the thalamus and caudate. The CST appears to project from the motor cortex to the spinal cord and passes through subcortical regions, suggesting that this component involves frontal and subcortical regions. Such a connection (the frontal‐subcortical circuit) has shown to be highly sensitive to common/inhibition cognitive control (Aron, Herz, Brown, Forstmann, & Zaghloul, 2016).

A series of parallel segregated frontal‐subcortical circuits are now known to link specific regions of the frontal cortex to the striatum, globus pallidus (GP), substantia nigra (SN), and thalamus. This is an important effector mechanism that allows an organism to interact adaptively with its environment (Alexander, 1994). Impaired executive functions, apathy, and impulsivity are hallmarks of frontal‐subcortical circuit dysfunction. Our results showed that frontotemporal regions are not only consistent with previous studies related to basic sensorimotor speed but are also linked to motor inhibition circuitry. Remarkably, the results show that the frontal‐temporal networks and the frontal‐subcortical circuit are critical for basic processing speed and motor inhibition, which is primarily consistent with Miyake's model regarding the common EF.

Intriguingly, these results somewhat contrast with those of Reineberg et al. (2015), who showed that increased common EF was associated with connectivity in the frontal pole with an attentional resting‐state network, as well as Crus I and II of the cerebellum with the right frontoparietal resting‐state network. We also found that the common control function was associated with the frontal pole, but we did not find a relationship with the cerebellum. This study is the first multimodal analysis of cognitive control, and future studies are needed to clarify whether the cognitive control tasks adopted by the two different studies resulted in discrepancies or if the analysis methods affected the results (unimodal vs. multimodal and ICA vs. joint ICA).

4.4. IC #5 and shifting construct

The structural and functional regions of IC #5 were associated with the shifting construct, which mainly consists of the cuneus and precuneus (overlapped regions of ALFF and GM). Consistently, functional MRI studies on task switching have shown activity evident in these regions (Dove, Pollmann, Schubert, Wiggins, & Yves Von Cramon, 2000), particularly in the cuneus and precuneus for shifting attention (Le et al., 1998). Moreover, the WM tracts mainly involve the ATR L/R and are connected with the caudate, while the forceps major is connected with the cuneus. Overall, this component clearly shows a relation to the shifting constructs of cognitive control functions.

Again, the current findings appear to be different from the results of Reineberg et al. (2015), who showed that somatomotor/attentional resting‐state networks were associated with higher shifting‐specific construct. Our data appear to involve some major clusters of the visual network such as the cuneus. However, our data show that the ATR involves a shifting construct, which is consistent with the literature. We also noted that the ATR consists of fibers connecting the thalamus with the prefrontal cortex and involves higher‐order cognition (Biesbroek et al., 2013; Duering et al., 2011; Mamah et al., 2010; Wright, Vann, Aggleton, & Nelson, 2015).

4.5. IC #4 and updating construct

In contrast to Reineberg et al. (2015), we observed that a higher updating ability was associated with structural and functional overlapped regions in the temporal lobe and with the WM tract of the CST. Prior research has indicated that the function of working memory load updating (load‐dependent activity for updating) is associated with the temporal, occipital, and subcortical regions (Leung, Oh, Ferri, & Yi, 2007). Charlton, Barrick, Lawes, Markus, and Morris (2010) discovered that the integrity of white‐matter pathways connecting the prefrontal cortex (PFC), parietal cortex, and temporal cortex correlates with working memory performance. Thus, the current findings appear to agree with the literature and show a strong relationship between working memory and the temporal lobe.

4.6. Summary of associations with cognitive control constructs across all brain imaging modalities

We revealed common and discrete cognitive control constructs across different brain imaging modalities. Particularly, in the WM modality, we found that the ATR tract is the most consistent structure shown among all components associated with cognitive control constructs, whereas CST was observed among the components associated with cognitive control constructs except shifting ability. These WM alterations suggest a potential role of aging in executive functions. In the results of ALFF, IC #3 showed that the parietal–frontal regions are the most associated with all cognitive control constructs. However, the STG of IC #4; the cuneous, PCC, and lingual gyrus of IC #5; and the middle frontal gyrus (MFG), superior frontal gyrus (SFG), medial frontal gyrus (MeFG), and STG of IC #6 are associated with the updating, shifting, and common abilities in Miyake's model, respectively.

For the GM modality, structural features are more spatially consistent across cognitive control constructs. For example, the MTG is one of the susceptible structures across all cognitive control constructs (i.e., updating of IC #4, shifting of IC #5, and common control of IC #6). This indicates that this region is not only correlated with age‐related decline but also with higher‐order cognitive control constructs. Notably, multimodal brain features including the CST and ATR tracts, middle frontal gyrus, precentral/postcentral gyrus, cuneus/precuneous cortex, and superior temporal gyrus all play important roles in age differences of common cognitive control. Furthermore, despite shifting ability not being sensitive to age, it is associated with the middle frontal gyrus, parahippocampus gyrus, cuneus, precuneus, ATR, and forceps major tract. The updating construct is sensitive to aging and is mainly associated with the precentral/postcentral gyrus, precuneous/cuneous, middle frontal gyrus, and the WM tract of the CST. Ultimately, the dynamic interaction of these brain features (i.e., increasing or decreasing) among these modalities in each component shapes the common and distinct cognitive control constructs.

4.7. Limitations and supplemental analyses

The results generally agree with the literature in showing multimodal independent components associated with common and distinct cognitive control constructs. Nevertheless, there are multiple approaches to fusion multimodal neuroimaging data, such as parallel ICA and mCCA + jICA. Each approach has advantages and limitations. To generalize the current findings, we also analyzed the data from these two approaches and report them in Supporting Information Files SS1 and SS2.

For consistency, we applied resting‐state ALFF data in both para‐ICA and mCCA + jICA fusion approaches to see whether these fusion models also support the current findings. We first regressed out the gender factor in para‐ICA and mCCA + jICA since brain data might be sensitive to it. The results obtained from these two methods were generally similar to those reported with the jICA approach regarding the components associated with the cognitive control constructs. Particularly, several of the same key brain features as in the Summary section are constantly reported across different modalities, such as the STG, MTG, precuneus, and cuneus. Regarding the association with cognitive control constructs, this study showed robust results in terms of the joint components associated with the common cognitive control score across different approaches to multimodal fusion analysis.

5. CONCLUSIONS

This study revealed that age differences are associated with widespread brain degeneration that is not limited to the frontal region. Particularly, we extended the current knowledge of the frontal–parietal network in relation to not only all cognitive control constructs but also those that are susceptible to the aging process. Distinct brain features related to processing speed are correlated with common EF, whereas the visual functions network is critical in shifting ability. Updating ability mainly consists of the dorsal stream, suggesting vulnerability to aging effects in cognitive control. Crucially, these findings suggest there are some common and distinct joined multimodal components that correlate with the psychological constructs of common and discrete cognitive control functions, respectively. Dynamic interactions among these brain features across different brain modalities could serve as possible developmental mechanisms associated with these age effects.

Supporting information

Supplementary Figure S1 Supplementary Information.

Supplementary Table S1 Regression r values for the mixing coefficients of 5 joint independent component analysis (jICA) components (IC #) with age and cognitive control common and specific (shifting, updating, and inhibition) residuals.

Supplementary Table S2. Talairach map of rfMRI (ALFF) in the component #3

Supplementary Table S3. Talairach map of sMRI (gray matter) in the component #3

Supplementary Table S4. The white matter tract labels in the component #3

Supplementary Table S5. Talairach map of rfMRI (ALFF) in the component #4

Supplementary Table S6. Talairach map of sMRI (gray matter) in the component #4

Supplementary Table S7. The white matter tract labels in the component #4

Supplementary File S1: Parallel ICA approach

Supplementary File S2: mCCA + Joint ICA approach

Yang M‐H, Yao Z‐F, Hsieh S. Multimodal neuroimaging analysis reveals age‐associated common and discrete cognitive control constructs. Hum Brain Mapp. 2019;40:2639–2661. 10.1002/hbm.24550