The participants of the current study came from the longitudinal Geneva Aging Study. They all provided written informed consent, and approval was obtained from the ethics committee of the Faculty of Psychology and Educational Sciences of the University of Geneva and the Swiss Ethics Committee.

Older subjects that underwent two T1-weighted structural images, diffusion-weighted images, and fMRI scans, as well as a battery of cognitive tests were selected from a larger pool of the Geneva Aging Study, a lifespan study of 219 older adults (Fagot et al., 2018). The current analyses focused on a subset of individuals that completed longitudinal follow-up at ~2.5 years including behavioral testing and scanning. Subjects missing any one of the behavioral tests or scans were not included. At this stage, 31 subjects were selected. Two subjects were excluded due to additional missing behavioral data and one because of a lesion on the anatomical MRI. The final sample included 28 older subjects (mean age at first scan = 72 ± 6 years, mean age at second scan = 74 ± 6 years; eight males). Their demographic information is presented in Table 1. Participants were screened for health problems with a health questionnaire and their structural MRIs were inspected by a neurosurgeon with significant experience evaluating MRIs (SR) to rule out abnormalities. The battery of the behavioral tests was systematically administered before the scanning sessions (mean interval between testing and MRI = 57 days, SD = 39). Image quality was additionally ensured by examining head motion: absolute head motion at a single time point >2 mm and relative head motion >2.5 mm were criteria for participant removal. No participants were removed.

Participants underwent a battery of cognitive tests including 11 different measures which are briefly synthesized in Table 2 and reported in detail in other manuscripts (Mella et al., 2013, 2015; Fagot et al., 2018). Two of the measures, heterogeneity and amplitude of change, indicate an age-related change in cognition and the rest indicate baseline cognitive performance (i.e., measured at one timepoint). These measures are a composite score summarizing spatial and spatial-verbal working memory scores (Mat_span, MDA), crystallized abilities (MH), fluid intelligence (Raven), cognitive inhibition (Stroop), as well as mean performance and intraindividual variability in three reaction time tasks (a simple reaction time task, SRT) and a two-conditions complex processing speed task (CL6 and CL9). Intraindividual variability and mean performance have been shown to reflect different cognitive processes, intraindividual variability being linked to sustained attention abilities, while mean performance reflects the general level of processing speed. The composite scores of amplitude and heterogeneity of change in cognitive abilities at the individual level have been computed using individual analyses of variance of all tasks assessing simple reaction times, complex processing speed, and inhibition. This allowed assessing both the general amplitude of change and the heterogeneity of individual change across these tasks independently of the sample characteristics (see Mella et al., 2018 for more details). Briefly, heterogeneity of change refers to variations in the patterns of change across various cognitive measures. The homogenous decline is said to reflect a global decline in attentional resources or of mental resources, while strong heterogeneity of change may indicate a decline in one ability but stability or improvement in another (i.e., heterogenous pattern). Mean amplitude of change refers to both the direction and amplitude of the change in performance across the tasks considered. A smaller, negative change does not necessarily indicate decline but rather stability or less improvement. The higher the absolute value the stronger the change (explained in detail in Mella et al., 2018).

The 11 single measures of cognition, as well as the general amplitude and heterogeneity of change were correlated with change in SC and FC (described below) using Partial Least Square Correlation (PLSC; Mella et al., 2013, 2015; Fagot et al., 2018).

Participants were scanned in a Siemens Trio 3T magnet. The following imaging sets were analyzed in the current study: resting-state functional MRI (fMRI), structural T1-weighted MR image (T1-w), and DWI sequences. Two sequences of 30 directions DWI were acquired (TR = 8,400 ms, TE = 88 ms, b value = 1,000 s/mm², and voxel size 2.0 mm³). The two DWI acquisitions were concatenated using FSL 5.08¹ (Smith et al., 2004) to increase the signal-to-noise ratio during the postprocessing. Then, a structural T1-w MR image was acquired (TE = 2.27 ms, TR = 1,900 ms, FOV = 256 mm, voxel size 1.0 mm³). Finally, the resting-state fMRI was obtained using an echo planar imaging acquisition (echo time, TE = 30 ms, time repetition, TR = 2,100 ms, flip angle = 80°, field of view, FOV = 205 mm, voxel size = 3.2 mm³, 140 volumes).

The diffusion data were corrected for eddy currents and movements using FSL (Andersson and Sotiropoulos, 2016). Structural T1-w MR images were preprocessed using Freesurfer version 6.0² (Fischl, 2012) with a standard automated preprocessing pipeline (i.e., “recon-all” with the default set of parameters). Tissue-segmented images were obtained (i.e., white matter, gray matter, and cerebrospinal fluid), and used at the diffusion data analysis stage. For each participant, the T1-w image was registered to the diffusion data with cross-modal registration (i.e., rigid with six degrees of freedom) using a boundary-based cost function (Greve and Fischl, 2009). The diffusion data were further processed using MRtrix3³ (Tournier et al., 2019) following the following standard structural connectome construction steps. Briefly, to estimate white matter fiber orientation distributions the white matter response function was estimated and used to perform single-shell, single-tissue constrained spherical deconvolution (MRtrix command “dwi2 response Tournier”; Tournier et al., 2004, 2007, et al., 2013). Whole-brain tractography was generated using anatomical-constrained tractography (MRtrix command “tckgen”, 50 million streamlines, maximum tract length = 250, fractional anisotropy cutoff = 0.06; Smith et al., 2012). Spherical-deconvolution informed filtering of tractograms (SIFT2, command “tcksift2”) algorithm was used to estimate structural connection density (Smith et al., 2015; Tournier et al., 2015). Finally, the structural connectome was produced by mapping a multi-modal parcellation atlas with 360 regions (180 per hemisphere) described in detail in Glasser et al. (2017), to the streamlines obtained from SIFT2. A 360 × 360 connectivity matrix was created, representing the number of white matter reconstructed pathways for each pair of regions, normalized by the sum of the volumes of the two regions.

Resting-state fMRI was preprocessed using SPM8⁴ and functions of the data processing assistant for resting- state fMRI (Yan, 2010) and individual brain atlases using statistical parametric mapping (Alemán-Gómez, 2006) toolboxes. Standard preprocessing steps were followed: realignment of the functional scans, spatial smoothing with an isotropic Gaussian kernel of 5-mm full width at half maximum, co-registration of the T1-w image to the functional mean, and segmentation of the structural images (white matter, gray matter, cerebrospinal fluid). The average signal from cerebrospinal fluid and white matter then was regressed out from functional time courses, together with six motion parameters (translation and rotation along the three dimensions) and linear/quadratic trends. Next, Glasser’s multimodal parcellation atlas (the same used for the structural connectome) was resliced to functional resolution and applied to the fMRI data to estimate regional average time courses. For each participant, functional connectivity was calculated as Pearson correlation between each pair of time series. These were then Fischer z-transformed and stored in a 360 × 360 FC matrix.

For each subject, the change in SC and FC was calculated as the difference in connectivity values between time 2 and time 1, after taking the absolute values of the matrices for FC. The absolute change in connectivity at the individual level, reflecting therefore a change in connectivity strength (regardless of the sign, for FC), was then stored as final connectivity difference matrices.

Age and gender were regressed from the SC and Fischer’s z-transformed FC, and residuals were used for the PLSC analysis.

The behavioral measures (n = 13) were correlated (Pearson’s) with each other to better understand the relationships between them, resulting in a 13 × 13 cross-correlation matrix of correlation coefficients r. The p-values were corrected for multiple comparisons using a false discovery rate (FDR; Benjamini et al., 2006). MATLAB version 2019a⁵ was used for all statistical analyses.

PLSC for neuroimaging (McIntosh and Lobaugh, 2004; Krishnan et al., 2011) was used to assess the multivariate patterns of correlations between behavioral variables (i.e., cognition) and SC and FC measures. The performed analysis relates to that described in Zimmermann et al. (2018). Briefly, the upper triangle of the symmetric connectivity difference matrices (360 × 360 regions of interest) was vectorized for each of the subjects. Structural connections that were 0 at both timepoints, for 66% of subjects or more, were excluded from the analysis to avoid issues with resampling statistics (Zimmermann et al., 2018). Vectorized brain connectomes were then stacked across subjects. The behavioral variables were stored in a matrix of subjects^*13 behavioral measures. Next, to start the PLSC, we computed a correlation matrix between the brain and behavioral matrices (i.e., a cross-correlation matrix between two data matrices). The cross-correlation matrix representing the brain-behavior correlation was submitted to singular value decomposition (SVD), which yields orthogonal latent variables (LVs) that capture the covariance between the variables from the behavioral and brain datasets (McIntosh and Lobaugh, 2004; Ziegler et al., 2013). The significance of the LVs was determined via permutation tests (1,000 iterations) of the singular values from the SVD, and the stability of the brain and behavior weights (also called saliences) was assessed using bootstrapping (500 bootstrap samples). For each significant LV, bootstrap ratios (BSR) for brain and behavior weights were calculated as the ratio of the weight over its estimated standard error. The stability of each connection in brain weights and behavioral weights was assessed based on its BSR: a positive high BSR >2.5 contributed positively and reliably to the brain-cognition correlations while a negative high BSR <−2.5 contributed negatively and reliably to the brain-cognition correlations. The positive and negative dimensions reflect distributions of connections that covary in a similar pattern to one another. Therefore, a BSR with a larger magnitude indicates that the connection with which it is associated has a large singular vector weight (i.e., contributes to the LV) and a small standard error (i.e., stable across participants).

To explore the contributions of each imaging modality to behavior, a matrix of the pairwise linear correlation coefficients (Pearson correlation coefficient) between each pair of regions was calculated. Next, each matrix was vectorized and correlated using Pearson correlation.

The PLSC analysis was conducted with the 13 cognitive measures (see Table 2) as behavioral variables and with SC and FC change as brain variables.

The cognitive measures were mostly weakly correlated with each other. Only measures CL9 MRIM1 and CL6MRI1M, both measures of processing speed involving mean performance on letters comparison task with nine letters and six letters, respectively, were positively and significantly correlated after FDR correction (0.83, p < 0.00001; Figure 1).

The relationship between the change in FC and SC and cognitive function (resulting from the PLS analysis) was captured by one LV (17% of total covariance; singular value = 2,406.87; p = 0.0070), showing a significant contribution to the covariance. The LV revealed functional and structural correlations expressed across an array of behavioral measures, that optimally covaried with each other, and contributed negatively or positively to the FC and SC—cognition relationship. The contribution of the connections was expressed as BSRs, which indicates how robustly each connection contributed to the weighted pattern of the SC and FC—cognition matrix.

The two imaging modalities were overall weakly correlated (−0.072, p < 0.0001). The strength of correlations varied based on region as depicted in Figure 2. For example, regions in the posterior cortex (e.g., temporo-parieto-occipital junction, superior parietal, inferior parietal, posterior cingulate) were positively correlated with medial temporal and lateral temporal cortices. The anterior cortex and sensorimotor cortices largely showed negative correlations with the rest of the brain. Positive correlations indicate that both FC and SC change contribute in the same direction towards behavior, while negative correlations indicate that FC and SC change have opposite contributions.

The contributions to behavior can be better appreciated in Figure 3, which depicts the association of each region (i.e., node) to the brain-behavior relationship.

Specifically, several sets of functional and structural connections were found to have stable weights by bootstrapping (Figure 4). The behavioral weights corresponding to the significant LV are shown in Figure 5. In brief, Figure 5 indicates the behavioral pattern determined by the LV. PLS selects patterns of connections whose signal change covaries with the behavioral pattern across subjects identified by the LV. Those behaviors with higher BSR are said to modulate or contribute to a greater extent to the functional or structural connections than behaviors with lower BSR. The behavioral interpretation is presented in the following subsections (i.e., “Structural connections”, “Functional connections”) in conjunction with the brain weights.

Several sets of overlapping and non-overlapping functional and structural connections, specified in detail in the following subsections, were found to uniquely and independently contribute both negatively and positively to the association between change in brain connectivity and cognition in aging.

To facilitate interpretation of the data, in each hemisphere the 180 regions were separated into 22 larger partitions or “cortices” based on their topological proximity, common properties, based on architecture, task-fMRI profiles, and/or FC. Therefore, each of the 360 regions occupies one of the 22 cortices. The assignment is described in detail in Supplementary Neuroanatomical Results in Glasser et al. (2017).

Several distinct sets of functional and structural connections were identified and assigned to 22 larger partitions (see “Materials and Methods” Section). This can be visualized in Figure 2. Areas that contribute to behavior in the same way (either both negatively or both positively) are depicted in yellow (i.e., positive correlations), while areas that have opposing contributions are in dark blue. Visual, auditory, and temporal cortices mostly present positive correlations between FC and SC contributions indicating that in these areas FC and SC change contribute in the same direction to behavior. Anterior cortices had largely a negative correlation with other cortices. For example, FC change (Figure 7) indicates negative loading onto behavior while SC change (Figure 6) indicates positive loading. This indicates that there is preservation of structural connections but decline in functional ones.

The majority of the structural connections loaded positively onto the latent variables identified by PLSC, indicating stability or increased connectivity. Given the relatively short time of 2.5 years between the scans, and that our study sample is of healthy older adults we did not expect an accelerated global decrease in structural connectivity. Although the majority of structural connections contributed strongly and positively to the brain-behavior relationship, there were some exceptions with some areas contributing strongly and negatively (see Figure 4). Notably, interhemispheric structural connections between the right dorsolateral prefrontal cortex and several other regions (frontal, sensory, motor) in the left hemisphere: anterior cingulate and medial prefrontal cortex, inferior frontal, premotor, dorsal visual stream, and dorsolateral prefrontal cortex. The dorsal visual stream contains higher visual areas implicated in perceiving where visual stimuli are located and in planning visually guided actions (Ungerleider and Mishkin, 1982; Goodale and Milner, 1992). Other negatively contributing regions were left intra-hemispheric fronto-parietal and temporo-parieto-occipital structural connections. Decreased SC in these regions was associated with the decline in cognitive functioning (i.e., measured by heterogeneity and amplitude of change), and inferior baseline performance on crystallized intelligence, verbal-spatial memory, and processing speed. Our results are congruent with other studies that indicate that normal aging involves decreasing myelination and reduced microstructural integrity of the white matter in the dorsolateral prefrontal cortex as well as in the corpus callosum (i.e., reduced interhemispheric connections), leading to reduced conduction of nerve fibers and cognitive decline (Peters and Sethares, 2002; Bowley et al., 2010; Kaller et al., 2015). Such changes are commonly seen in other prefrontal areas with aging and correlated with reduced cognitive ability especially involving processing speed, planning, and distractibility/inhibition (Eckert et al., 2010; Luebke et al., 2010; Kaller et al., 2015). Our measures of cognitive functioning, heterogeneity, and amplitude of change, were derived from tests assessing simple reaction times, complex processing speed, and inhibition. The findings of a decline in these measures correlating with decreased SC in prefrontal, motor, and sensory regions are in line with the existing literature. Studies indicate the decline in white matter fractional anisotropy is associated with lower processing speed, reflecting a regionally specific structural decline in attention-related regions in the prefrontal cortex, and a structural decline in the motor and sensory cortex, as these last are essential for performing processing speed tasks (Salthouse, 2000; Coffey et al., 2001; Charlton et al., 2006; Kennedy and Raz, 2009; Eckert et al., 2010).

There was variation in the contribution of functional connections to the brain-behavior relationship, which is in line with the existing literature, which indicates that age-related functional reorganization does not follow a global, whole-brain pattern rather is region-dependent (Goh et al., 2013; Cao et al., 2014; Zimmermann et al., 2016). Some connections showed a strong, positive contribution including somatosensory, motor, and temporo-occipital connections. Thus, preserved or increased FC in these regions was correlated with higher scores in baseline spatial working memory and fluid intelligence, decline in cognitive functioning, inferior baseline performance on crystallized intelligence, verbal-spatial memory, and processing speed. These observations are largely consistent with other studies reporting age-related increases in FC within motor cortices (Meier et al., 2012; Tomasi and Volkow, 2012), somato-motor connections (Betzel et al., 2014), and visual networks (Betzel et al., 2014; Geerligs et al., 2015).

Other functional connections showed a strong negative contribution to the brain-behavior relationship. Largely, these were composed of interhemispheric cingulo-opercular, and insular and opercular connections with parietal, paracentral lobule, medial prefrontal cortex, somatosensory, motor, anterior, mid, and posterior cingulate functional connections. To provide behavioral context to the above results, decreased FC in these regions was associated with the decline or less improvement in cognitive functioning (i.e., measured by heterogeneity and amplitude of change), inferior baseline performance on crystallized intelligence, verbal-spatial memory, processing speed, and higher scores in baseline spatial working memory, fluid intelligence.

Interestingly, decreased cingulo-opercular network (including the insula) FC has been reported to correlate with age-related reductions in visual processing speed (Ruiz-Rizzo et al., 2019), and so in our study, this may explain the lower scores on heterogeneity and amplitude of change since these were derived using processing speed tasks with visual stimuli. Some of the other functional connections mentioned (posterior cingulate, prefrontal, parietal cortex, etc.) are part of the default mode network (DMN; Alves et al., 2019), and lower DMN FC has been reported to be decreased in age-related decline in processed speed (Damoiseaux et al., 2008). We found that decreased FC between right insular and frontal opercular cortex and several other regions: left paracentral lobular and midcingulate cortex, left anterior cingulate and medial prefrontal cortex, left inferior parietal cortex, and posterior cingulate was associated with decreased cognitive functioning (i.e., measured by heterogeneity and amplitude of change), inferior baseline performance on crystallized intelligence, verbal-spatial memory, processing speed, and higher scores in baseline spatial working memory, fluid intelligence. Disruption of coordinated functional activity of key network regions such as the ones mentioned (posterior cingulate, prefrontal, parietal cortex), has been strongly linked to cognitive decline in advancing age without an overt disease (Andrews-Hanna et al., 2007). Some studies found that the relationship between decreased DMN and reduced processing speed is attenuated by controlling for whole-brain FC and white matter hyperintensities (Staffaroni et al., 2018) and reduced functional anisotropy between DMN areas (Andrews-Hanna et al., 2007). Therefore, there is a proposed vulnerability of these connections with cognitive aging. Similarly, reduced FC was reported for the insular and cingulated cortex with aging and cognitive decline (Onoda et al., 2012). Our study did not focus on within and between network analysis, however, that would be an interesting direction to explore. For example, the anticorrelation between the DMN and dorsal attention network is well-studied metric for cognitive functioning in aging. However, multi-modal longitudinal studies are limited in this area (Zhu et al., 2016).

Moreover, right intra-hemispheric functional connections involving cingulo-opercular, fronto-parietal, and prefrontal-opercular (including insula) connections, were also found to be negatively contributing to the brain-behavior relationship. In a comprehensive review by Robertson (2014), cognitive reserve (known to be protective in cognitive aging) is linked to the functional integrity of the frontoparietal connections in the right hemisphere, meaning that less FC in these regions may indicate a decline in cognitive functioning involving arousal, sustained attention, response to novelty, and awareness (Robertson, 2014; Haupt et al., 2019). Thereby, corroborating our results of decreased FC in these regions associated with a decline or less improvement in cognitive functioning.

A significant strength of our study is our longitudinal design, as only longitudinal studies can measure change. It allowed the investigation of individual brain-behavior developmental trajectories in aging, rather than just general tendencies in age differences, an important approach considering that aging is a heterogenous process. The within-subject approach minimizes cohort and period effects typical of cross-sectional designs and may help disentangle to what degree change in FC and SC ultimately drives aging-related changes in cognition, as well as investigate the behavioral relevance of FC and SC metrics.

Further longitudinal studies investigating age-related change in FC and SC while considering cognition are necessary to elucidate the underlying brain mechanisms behind cognitive decline in normal aging. It would be valuable to investigate the change in FC within and between brain networks (i.e., DMN, salience network, etc) and correlate it with white matter integrity and SC, with various brain parcellations, and across large longitudinal samples of healthy older adults and in different pathological states (i.e., Alzheimer’s).

Some weaknesses of our study are our relatively small sample size (n = 28) and the small number of cognitive measures that represent change (i.e., heterogeneity and amplitude of change). We also had only two timepoints at a relatively short interval (2.5 years), which may limit the generalizability of our study to the lifespan scale. Additional time points and a larger time frame would allow a more robust assessment of where on the trajectory of age-related change an individual’s brain is.