Physical Measures (GXT, DXA, Accelerometery)

Graded Exercise Testing (GXT): The majority of participants completed a GXT using a treadmill (Quinton QStress, Cardiac Science, Chelmsford, Mass) with a substantially smaller number (<10%) who were unable to walk without holding onto the treadmill handrails using an electronically braked cycle ergometer (LODE Excalibur, Netherlands) to volitional maximal exertion. The testing protocol has been described previously [23, 54]. In brief, under the supervision of a physician, participants warmed-up for three to five minutes and then intensity was increased at a level equal to approximately 0.6 change in the metabolic equivalents of task (METS) per two-minute stage with active motivation from study staff (n=2 minimum offering ongoing and enthusiastic verbal encouragement). Exercise grew progressively harder until the physician ended the test based upon potentially dangerous changes in the ECG reading or an extreme hypertensive response (SBP>220 or DBP>110), or the participant indicated that they were unwilling/unable to continue. Individuals were excluded if they did not reach a minimum of 85% of their age predicted heart rate maximum (220-age) or testing was stopped by the physician prior to volitional fatigue. Maximal capacity was calculated in METs using the American College of Sports Medicine’s algorithm [56] designed for walking or cycling as appropriate based upon testing methodology. Based on the presumed linear relationship in oxygen consumption while coming to a metabolic steady state in response to a new workload, partial stages were scored in 30 s increments using the formula METs _{last_completed_stage}+0.25* METs _{difference_between_stages}* number of 30 s increments completed in the new stage.

Dual X-Ray Absorptiometry (DXA) to estimate body composition: A GE Lunar Prodigy at one location and an iDXA (both GE/Lunar, Madison, WI) at the other were used to estimate segmental body composition and provide absolute values (measured in grams or kilograms) for fat and lean tissue and bone mass for the arms and legs and trunk as well as secondary height-dependent regions titled android (centered on the abdomen) and gynoid (centered on the upper thighs). Visceral fat values were also derived from the android region. This method of estimating visceral fat has had good agreement compared to 3-D imaging techniques in both men and women [57, 58].

Participants were positioned in line with best practice recommendations [59] and external artifacts were removed whenever possible. For participants who were larger than the available scan field, a “hemi-scan” was acquired by scanning only the right side of the body and replicating those values for the “missing” limb.

Accelerometry: Measurement of physical activity: Participants were asked to maintain their normal behaviors during a 10 day period during which they were equipped with an Actigraph GT9X+Link (ActiGraph Inc, Pensacola, FL) deployed in line with best practice recommendations [60–63] worn on the participant’s non-dominant wrist continuously except when engaging in water-based activities like swimming or bathing. This tri-axial accelerometer has been shown to be both valid and reliable across the age span [60, 64, 65]. After 10 days of deployment, devices were recovered and data were downloaded and screened for completeness and potential device malfunction in line with established practice [60, 62, 66]. Data processing included applying a screening algorithm to detect non-wear [67, 68] and aggregating raw data into “counts per minute” in the x, y and z axes independently using Actilife (Actigraph’s proprietary software). Vector magnitude was calculated using the square root of the sum of the squares of the three axes to incorporate intensity, frequency, and duration of movement. This metric, henceforth referred to as Vector Magnitude Counts Per Minute (VM CPM) has been recommended for use in assessing physical activity during a 24-h wear period [69] particularly when the device has been deployed at the wrist.

Accelerometry: Measurement of sleep: The same device was used to calculate sleep time in total minutes, wake after sleep onset (WASO) both in terms of number of awakenings and number of minute awake, and sleep efficiency using an algorithm that has been validated for use in adults [70]. Participants were specifically asked to maintain their normal sleep rhythms during the wear period. The window of observation was derived from sleep journals in which participants indicated the time that they had begun trying to sleep, and the time that they first woke up in the morning (i.e. there was no effort to record incidental, or undesired, nighttime awakening). All records were manually entered and inspected by a specifically trained research assistant. If, based on visual inspection more than 50% of a period at the immediately following the “go to bed” or preceding the “first wakening” time appeared to have substantial amounts of movement indicative of poor subject record keeping, in-bed or first awakening time was adjusted to reflect the period when movement appeared to (mostly) cease. If a sleep journal was not maintained, previously validated methods were utilized to estimate the time of sleep onset and wake time [71]. Individuals with obstructive sleep apnea (OSA) were identified and sleep variables were compared against those without OSA.

Imaging Measures

Imaging Acquisition: Three magnetic resonance imaging (MRI) scanners at two sites were used to acquire resting state functional MRI (rs-fMRI) data (UCSD: GE MR750 3 T scanner (GE, Milwaukee, WI) with an 8-channel head coil; WUSTL: Siemens 3 T Trio and 3 T Prisma-FIT (Erlangen, Germany) with a 20-channel head coil. T1-weighted (T1w) and T2-weighted (T2w) structural scans were performed for purposes of image registration and radiological screening of the participants (UCSD: T1 MPRAGE, TE=3.036 ms, TI=1.000 ms, 1.0 mm³ voxels; T2 CUBE, TR=3.300 ms, TE=73.37 ms, 1.0 mm³ voxels; WUSTL: T1 MPRAGE, TR=2.400 ms, TE=3.16 ms, TI=1.000 ms, 1 mm³ voxels; T2 SPACE, TR=3.200 ms, TE=458 ms, 1 mm³ voxels). Four rs-fMRI scans (140 frames per run) were acquired per subject using a multi-echo sequence (UCSD: TR=2.740 ms, TE=14.8, 28.4, 42, 55.6 ms; 4.0 mm³ voxels; WUSTL: TR=2.960 ms, TE=15, 31.3, 47.6, 63.9 ms; 4.0 mm³ voxels) with a total run time of 6.4 min (UCSD) or 6.9 min (WUSTL) per run or 25.6 and 27.6 total scan minutes respectively. Gradient echo field maps were acquired for later use in correction of susceptibility-related image distortion. During fMRI, participants were shown a neutral video without audio (e.g., nature scenes) synchronized to the start of each run and were asked to stay awake without engaging in any sort of meditation.

Image Preprocessing: fMRI data processing largely followed previously described methods [72] which makes use of several FSL modules [73]. Briefly, rigid body motion correction both within- and across-runs was computed on data summed over all echos. Slice timing correction was applied to each echo. Bias field inhomogeneities were corrected using the FAST module in FSL [74]. Atlas transformation was computed by composition of transforms (individual frame→frame average→T2w→T1w→atlas representative target). T1w→atlas registration was computed with FSL FNIRT. Three atlas representative targets, all representing the 711-2B version of Talairach space, had been previously prepared for each of the scanners to accommodate scanner-specific differences in T1w contrast. Final resampling of the fMRI data in 3 mm³ voxel 711-2B space was accomplished in one step incorporating distortion correction, previously computed bias field correction, and the composition of all spatial transforms.

The multi-echo data were then modeled according to standard theory [75] fitting the four echoes to a monoexponential model S_t=S_0texp(−R_2t^TE_k:, where indexes frame, indexes echo, and is reconstructed intensity extrapolated to echo time 0. Frame-to-frame variation in was suppressed by averaging over the whole run and the fMRI data were modeled at a TE of 30 ms according to S_t=(S_0) ®exp(−R_2t^30 ms). The modeled data in each run then were intensity normalized (one multiplicative scalar applied to all voxels and frames) to achieve an intensity mode value of 1000.

Denoising was effected on the fMRI data virtually concatenated across the 4 runs. Frame censoring was computed on the basis of DVARS [76], with the criterion adjusted to compensate for baseline variability using a previously described method based on fitting the distribution DVARS values to a gamma function [77]. Subsequent steps ignored all censored frames. The data were denoised using a CompCor-like scheme with regressors derived from motion correction temporally filtered to suppress respiration-related factitious head motion [78], white matter, ventricles, extra-axial cerebral spinal fluid (CSF), and the whole brain global signal [79]. Image derived regressors were based on tissue class segmentations computed by FreeSurfer 6.0.0 [80]. Additional denoising included bandpass temporal filtering retaining frequencies in the range 0.01–0.1 Hz and spatial filtering (Gaussian blur of 6 mm in each cardinal direction). Finally, the (scanner-specific) response evoked by the movie was averaged over all participants and subtracted from each individual’s data.

ROI/Network Creation: Relevant voxel locations were initially identified using the Big Brain 300 parcellation described by Seitzman and colleagues [81] excluding subcortical and cerebellar RIO’s. In an effort to confirm/replicate earlier findings we projected the regions described by Voss and colleagues [7] onto the Seitzman ROI’s. Thus, three networks were defined; the salience (SAL), motor control (MOT), and visual (VIS) networks. Two additional networks described by Voss et al. [7] the default mode network (DMN) and the dorsal attention network (DAN) were substantially different when compared to Seitzman and colleagues parcellation [81]. Accordingly, we utilized both with labels DMN and DAN for Voss defined regions and BSDMN and BSDAN for Seitzman. Finally, because Seitzman and colleagues did not identify an executive control network (ECN) we utilized the Voss visual representations and included voxels identified as frontoparietal and DMN within the Sietzman designations. The voxel locations and network designates are included in the supplementary materials. A visual representation of Seitzman-Voss mapping for SAL is shown in Fig. 1 below, and mapping for all five examined networks are shown in supplementary Figs. 1 through 5. The average correlation value, defined as the average correlation across the ROI x ROI pairs within each network was calculated and assessed as the within network connectivity value for subsequent analyses.

Open in a separate window

Fig. 1

Seitzman and Voss coordinates and ROI’s—salience network. Seitzman ROI’s as designated by different colors on the top left (A) and Voss ROI’s are designated on the bottom right (B). The overlay areas of the two are shown on the top right (A’). Note: Visual representation of overlay between other networks of interest are available in supplementary materials

Locational Differences

There was a significant difference between the locations for participants’ maximal cardiovascular fitness and connectivity with WUSTL having a population with higher fitness and less connectivity. Means, standard deviations and p values for difference by location for all demographic, physiological, and connectivity values are shown in Table 1. Mean differences and confidence intervals of the difference for variables with significant differences for both location and sex are shown in supplementary materials Table 1.

Correlations

Significant correlations between age and the SAL, VIS and BSDMN networks were observed (p=<0.001, 0.026 and 0.016 respectively). As such, age was included as a covariate in other correlational analyses. After controlling for age, sex, and location all resting state functional networks had significant positive correlation (p=<0.001–0.005) except for between VIS and BSDAN (p=0.763).

Similarly, many physiological variables were associated. Specifically, cardiorespiratory fitness was associated with total physical activity (p<0.001 r=0.374) and all metrics of body composition (p range <0.001 and =0.024). Additionally, all metrics of body composition were associated with each other (p<0.001 for all). All sleep-based variables were also significantly associated (p<0.001 for all). In addition to the relationship to CRF mentioned above, total daily physical activity was negatively associated with metrics of body fatness (i.e. more movement=less fatness). Further, PA was negatively associated with sleep efficiency and time, and positively associated with wake after sleep time (i.e. more movement=less sleep time and lower sleep quality). A scatterplot showing the significant associations between fitness and DMN is shown in Fig. 2 below. All other correlation matrix tables and scatterplots for both network to network associations and associations between physiological variables are provided in the supplementary materials.

Open in a separate window

Fig. 2

Cardiorespiratory fitness and DMN connectivity shown both continuously and split into best and worst quartile. Left-hand graph=Continuous comparison of unstandardized residuals of the CRF and DMN connectivity (p=0.040). Right-hand graph=T-test based comparison of the best and worst CRF residual groups with DMN connectivity (p=0.049). CRF=Cardiorespiratory Fitness; DMN=Default Mode Network

Cardiorespiratory fitness was correlated with DMN (p=0.008 r=0.142), SAL (p=0.005, r=0.152) and BSDMN (p=0.008 r=0.143) but not with ECN, DAN, MOT, VIS, or BSDAN (p range=0.248–0.982). Total physical activity was not significantly associated with greater MOT connectivity (p=0.060; r=0.124), or any other network (p range=0.156–0.862).

When exploring associations between FC and body composition, the only significant associations were between percent body fat and the VIS (p=0.03, r=0.117), DMN (p=0.05; r=0.105), and DAN (p=0.047; r=0.107) networks.

Greater sleep efficiency was associated with greater connectivity in the SAL (p=0.007; r=0.137) and BSDMN (p=0.016; r=0.122) networks. Total sleep time was inversely associated with ECN, MOT and VIS connectivity (p=0.016, 0.025 and 0.027 and r=−0.121, −0.114 and −0.112 respectively). Total amount of time awake during (attempted) sleep periods was negatively associated with multiple networks including SAL (p=<0.001, r=−0.18), MOT (p=0.007; r=−0.136), BSDMN (p=0.038; r=−0.105) and BSDAN (p=0.043; r=−0.102). The number of times awakening during a sleep period (regardless of total length of time awake) was negatively associated with ECN (p=0.001; r=−0.162) and MOT (p=0.013; r=−0.126).

Cardiorespiratory Fitness

We found relations between greater aerobic fitness and stronger FC in the DMN and SAL networks. Interestingly, we did not find a predictive association when CRF and the SAL network are considered continuously, but did find significant differences in this network when comparing “best” and “worst”. This suggests that, at least for some networks, there is a threshold above which fitness is protective. These data strengthen, through replication, previous results [7, 84, 85] indicating that the most fit individuals had stronger connectivity in the DMN and SAL networks than the least fit, and that greater connectivity in these regions is associated with “younger” brains [7]. The significance of these associations is further strengthened by the fact that the mean difference in connectivity was greater than 7% in these networks relative to the FC mean over the total population.

These observed associations are consistent with the cardiovascular fitness hypothesis, which postulates that to meaningfully impact health, physical activity must be at an intensity sufficient to improve fitness [7, 84, 85], and thus light physical activity is unlikely to be beneficial. There are several mechanisms that could possibly contribute to the observed associations. Potentially the most widely studied is the increased/improved cerebral blood flow related to greater microvasculature associated with better fitness. [86]. Others have suggested that individuals with higher CRF may have actually have lower blood flow needs within key regions associated with the DMN and SAL [87]. Alternatively, given that higher levels of fitness are often associated with less systemic inflammation ([88], it may be that greater CRF is also protective against neuroinflammation and its negative downstream effects [89]. Regardless of mechanism, given the known associations between the ability to reflect/learn and FC in the DMN [90, 91] and to regulate emotions and FC in the SAL network [92, 93], aerobic capacity may be important to successful aging of the brain in addition to its role in cardiovascular and metabolic health.

Physical Activity

In contrast, we did not find that daily quantity of physical activity was significantly associated with FC in cognitively important networks. Although wrist worn devices limited our ability to differentiate activity intensity, Bassett and colleagues [69] found that this metric was sufficient to identify meaningful differences in health across the NHANES population and recommended it as a metric to accurately measure total daily physical activity. Our results contrast with previous results which have found increased FC in the DMN, SAL, and ECN and with greater amounts of self-reported physical activity [94] and in the DMN with greater accelerometer measured activity [95, 96]. It is worth noting that none of these papers included/controlled for measures of fitness or sleep in their analyses, which may help to explain the discordant findings. Additionally, and in line with the cardiovascular fitness hypothesis introduced earlier it is likely that not enough of our (exclusively sedentary) population had a sufficiently large amount of activity (particularly moderate to vigorous activity) to detect meaningful differences in FC that might occur with more intense activity. However, the present results do further validate prior [7] findings that differences in FC attributable to cardiovascular fitness are independent of regular physical activity.

The observed association between all day physical activity and increased FC in the MOT network is intuitively reasonable. Although there is ample literature concerning changes in motor performance with age [97, 98] and exercise-based interventions [98, 99], there is little scholarship on the links between exercise/physical activity and MOT network functional connectivity in healthy human populations. In one of the few extant studies, sedentary youth had significantly higher MOT connectivity at rest compared to age-matched endurance runners [100]. In clinical populations, greater MOT connectivity has been observed in stroke survivors compared to healthy controls [100, 101]. Studies in both rats [102] and humans [103] following stroke suggest that these effects may reflect new neural pathways in the MOT network to compensate for lesions of the primary motor tracts. This, along with findings suggesting that older age is not associated with differences in the MOT network [7] suggest that increased resting state functional connectivity may not always represent a positive biomarker, depending on brain region/network and clinical status of the participant.

Body Composition

The present sample indicates minimal associations between body composition and FC. When viewed continuously, there does appear to be some association with higher body fat predicting reduced connectivity. However, these contributions are not sufficiently robust to establish a significant difference between the most and least fat quartiles in this relatively large sample. This conflicts with previous results by Kullman et al. [104] who found significantly increased connectivity with increasing BMI in a continuous fashion and with other group based work that showed increased functional connectivity in the SAL network in an obese compared with non-obese population [38]. In contrast, although not significantly associated when viewed continuously (p=0.061, see supplementary materials) when top vs. bottom quartiles were examined, we found that the highest fat quartile had significantly greater connectivity in the DAN, a network associated with attentional focus [98, 99]. Although minimal declines in DAN FC have been observed with advancing age [7], significantly weaker FC has been observed both in children [105] and adults with obesity [104] Thus, the current findings do not replicate certain prior results. This discrepancy could reflect population differences, as the current study sample consisted entirely of older adults with approximately one third of women, and just over half of men having body fat percentages qualifying them as obese using gender and age specific cut points derived from DXA that correspond with the traditional BMI values designating obesity. Further, it may be that the presence of additional (non-visceral) body fat is protective in older age in a way that it is not for younger adults. This has been shown in a systematic review looking at all-cause mortality and morbidity where older adults that were overweight or mildly obese had the lowest mortality [106]. The advantages of (slightly) higher non-visceral body fat may translate to brain health as well, as the associations between higher body fat and leptin production, and the associations between increasing age and decreased leptin production are both well documented [107, 108]. Although far from conclusive, there is some evidence suggesting that maintaining normal levels of leptin into older age may help to preserve memory and attentional focus [107].

Sleep

In agreement with our a priori hypothesis, better sleep, defined as more minutes of total sleep, greater sleep efficiency, and less nightly wakeful periods, was positively associated with FC across multiple networks. This is true of sleep quality, defined as greater efficiency and less time awake during the attempted sleep period, when the data are viewed both continuously and in “best” vs “worst” quartiles. In contrast, sleep quantity (i.e. total sleep time) was only associated with ECN connectivity when viewed continuously. Although some literature suggests that more minutes of sleep minutes of sleep are associated with greater FC, and improved cognitive outcomes, particularly in working age adults [44, 109, 110] in this instance, the contribution of sleep minutes to ECN connectivity was not sufficient to establish a significant difference between the best and worst sleepers.

Greater sleep efficiency was associated with greater FC in both the SAL and MOT network, while longer amount of time awake was associated with weaker FC in the SAL and MOT networks. These findings agree with prior findings in children [42] and adults [111] in which poor sleep both acutely and over extended periods of time contributed to emotional dysregulation, decreased coordination, and reduced connectivity in the SAL and MOT networks.

This combination in which sleep quality has a larger impact than sleep quantity in terms of mean difference and number of affected networks could be explained by the importance of different stages of sleep which cannot be measured through accelerometery alone. Indeed, it is well known that deep and REM sleep plays an important role in synaptic plasticity [112]. Further, it may be that continuous sleep is important for the clearing of metabolic waste. Indeed, clearance is mediated by the glymphatic system [113] which appears to be more active during extended sleep periods [113, 114]. Considered in aggregate, the available evidence indicates that good sleep is associated with strong FC across multiple regions although it is very possible that this association bi-directional in nature.

Not applicable.