The effects of age on resting-state BOLD signal variability is explained by cardiovascular and cerebrovascular factors

Accurate identification of brain function is necessary to understand neurocognitive ageing, and thereby promote health and well-being. Many studies of neurocognitive aging have investigated brain function with the blood-oxygen level-dependent (BOLD) signal measured by functional magnetic resonance imaging. However, the BOLD signal is a composite of neural and vascular signals, which are differentially affected by aging. It is therefore essential to distinguish the age effects on vascular versus neural function. The BOLD signal variability at rest (known as resting state fluctuation amplitude, RSFA), is a safe, scalable and robust means to calibrate vascular responsivity, as an alternative to breath-holding and hypercapnia. However, the use of RSFA for normalization of BOLD imaging assumes that age differences in RSFA reflecting only vascular factors, rather than age-related differences in neural function (activity) or neuronal loss (atrophy). Previous studies indicate that two vascular factors, cardiovascular health and cerebrovascular function, are insufficient when used alone to fully explain age-related differences in RSFA. It remains possible that their joint consideration is required to fully capture age differences in RSFA. We tested the hypothesis that RSFA no longer varies with age after adjusting for a combination of cardiovascular and cerebrovascular measures. We also tested the hypothesis that RSFA variation with age is not associated with atrophy. We used data from the population-based, lifespan Cam-CAN cohort. After controlling for cardiovascular and cerebrovascular estimates alone, the residual variance in RSFA across individuals was significantly associated with age. However, when controlling for both cardiovascular and cerebrovascular estimates, the variance in RSFA was no longer associated with age. Grey matter volumes did not explain age-differences in RSFA, after controlling for cardiovascular health. The results were consistent between voxel-level analysis and independent component analysis. Our findings indicate that cardiovascular and cerebrovascular signals are together sufficient predictors of age differences in RSFA. We suggest that RSFA can be used to separate vascular from neuronal factors, to characterise neurocognitive aging. We discuss the implications and make recommendations for the use of RSFA in the research of aging.


Introduction 1
The worldwide population is rapidly aging with an increasing number and proportion of older 2 adults across the globe (Beard et al., 2016). Considering the cognitive decline and increasing burden of 3 dementia in aging societies, there is a pressing need to understand the neurobiology of cognitive aging. 4 This will inform efforts to maintain mental wellbeing into late life, allowing people to work and live 5 independently for longer. Research in cognitive neuroscience of aging has used blood-oxygen level-6 dependent (BOLD) signal measured by functional magnetic resonance imaging (fMRI) as one of the 7 standard ways to examine the neural mechanisms of cognition. However, the BOLD signal measures 8 the activity of neurons indirectly through changes in regional blood flow, volume and oxygenation. This 9 makes BOLD a complex convolution of neural and vascular signals, which are differentially affected by 10 aging (Logothetis, 2008). Without careful correction for age differences in vascular health, differences 11 in fMRI signals can be erroneously attributed to neuronal differences (Liu et  2010; Restom et al., 2007). However, such methods have not been widely used, in part to 20 impracticalities in large-scale studies, and poor tolerance by older adults (for a review see Tsvetanov et 21 al., 2020). Additionally, a hypercapnic challenge may not be neuronally neutral, given participants' 22 awareness of the aversive challenge, which may differ with age (Hall et al., 2011). Breath-hold 23 compliance may also decrease with age (Jahanian et al., 2017). Such biases affect data quality and 24 reliability measures (Magon et al., 2009), highlighting the advantage of non-invasive and "task-free" 25 estimates of vascular components in the BOLD time series. 26 The BOLD signal variability in a resting state ("task-free") is one such estimate and is also known 27 as resting state fluctuation amplitudes (RSFA) (for a review see Tsvetanov et al., 2020). It has been 28 proposed as a safe, scalable and robust cerebrovascular reactivity mapping technique (Golestani et al., Evidence in support of cerebrovascular factors comes from Garrett et al. (2017) who found that "gold-41 standard" measures of cerebrovascular function (arterial spin labelling, ASL, and CO2 inhalation-induced 42 hypercapnia) are associated with RSFA. Importantly, both studies reported age-related differences in 43 RSFA that remain after adjusting for individual differences in either cardiovascular or cerebrovascular 44 factors. However, neither study considered jointly cardiovascular and cerebrovascular factors, and it 45 remains unclear whether the unexplained age-related differences in RSFA reflect joint contributions 46 from cardiovascular and cerebrovascular factors, as in the case of BOLD signal fluctuations (Chang et 47 al., 2013(Chang et 47 al., , 2009). Alternatively, the unexplained age differences in RSFA may reflect neuronal factors, 48 such as atrophy (Grady and Garrett, 2013), even though variation in neuronal activity does not explain 49 the effect of age on RSFA (Tsvetanov et al., 2015). 50 Cardiovascular, cerebrovascular and other physiological signals, but not neuronal signals, 51 contribute to the age-related differences in RSFA, yet none of these non-neuronal measures on their 52 own could fully account for the effects of age on RSFA. It is possible that various vascular signals 53 contribute to different components of the age effects on RSFA (Tsvetanov et al., 2020). However, no 54 study to date has tested whether the cardiovascular and cerebrovascular signals together fully capture 55 the effects of age on RSFA -an assumption underlying the use of RSFA as a scaling method. In this study 56 we sought to investigate the effects of age on RSFA by the simultaneous assessment of the independent 57 and shared effects of cardiovascular, cerebrovascular and neuronal effects on age-related differences 58 in RSFA. To this end, we used a set of cardiovascular, cerebrovascular and volumetric measures in a 59 population-based study of healthy ageing (age 18-88, N > 250, www.cam-can.org). We hypothesized 60 that age-related variation in RSFA are predicted by cardiovascular and cerebrovascular factors, but not 61 grey matter volume, and therefore that the residuals in RSFA -after adjusting for these vascular factors 62  Folstein et al., 1975) and Edinburgh Handedness Inventory (Oldfield, 74 1971 time (TI) 1, TI2 = 1800 ms, 1600 ms saturation stop time). The imaging volume was positioned to 126 maintain maximal brain coverage with a 20.9 mm gap between the imaging volume and a labelling slab 127 with 100mm thickness. There were 90 repetitions giving 45 control-tag pairs (duration 3'52"). In 128 addition, a single-shot EPI (M0) equilibrium magnetization scan was acquired. Pulsed arterial spin 129 labelling time series were converted to cerebral blood flow (CBF) maps using ExploreASL toolbox 130 (Mutsaerts et al., 2018). Following rigid-body alignment, the spatial normalised images were smoothed 131 with a 12 mm FWHM Gaussian kernel. In brief, the fastICA algorithm was applied after the optimal number of sources explaining the variance 174 in the data was identified using PCA with Minimum Description Length (MDL) criterion (Hui et al., 2011;175 Li et al., 2007;Rissanen, 1978). By combining the PCA and ICA, one can decompose an n-by-m matrix 176 of participants-by-voxels into a source matrix that maps independent components (ICs) to voxels (here 177 referred to as "IC maps"), and a mixing matrix that maps ICs to participants. The mixing matrix indicates Additional inclusion of total intracranial volume (TIV) did not change the principal results. Non-normally 205 distributed variables were logarithmically or exponentially transformed to conform normality (Fink, 206 2009

-Model 5: Covariates and grey matter volume measures 217
Note that the independent variables in Models 2, 4 and 5 included measures with voxel-specific 219 information, i.e. RSFA values across subjects in a given voxel were predicted by the CBF/GM values for 220 the corresponding voxel. 221 The residuals, ɛ, from each model were then used in a second-stage linear regression (i.e. 222 correlational analysis) to estimate their association with age. Voxels where the residuals correlate with 223 age (p<.05, FDR-corrected) indicate that the independent variables in first-stage model could not 224 explain sufficiently the age-dependent variability in RSFA. Conversely, residuals not associated with age 225 would suggest that the independent variables considered in the model are sufficient to explain age-226 dependent variability in RSFA. 227 This two-stage procedure was performed for each voxel of RSFA maps resulting in a statistics 228 map for each model indicating the association between residuals and age. Statistical maps were 229 corrected for multiple comparisons at p <0.05 (FDR-corrected). To further address multiple 230 comparisons and voxel-voxel mapping between modalities, we performed complementary analysis 231 where voxel-wise estimates of brain measures were substituted with subject-wise IC loadings, see 232 Section 2.6. 233 We also tested whether the distribution of age-RSFA residuals correlations across all voxels 234 formed differed from the predicted distribution under pure randomness. We constructed 5000 235 distributions of age-RSFA residual correlations across all brain voxels (DVoxels), where RSFA residuals were 236 based either on a model with obseved RSFA values (DVoxels1) or permuted RSFA values (DVoxels2-5000). 237 Distribution medians and distribution shapes were compared using Wilcoxon rank sum test and 238 Kolmogorov-Smirnov test respectively. We performed a pair-wise comparison across all 5000 239 distribution shapes using Kolmogorov-Smirnov test, resulting in a distribution of 4999 similarity scores 240 (D Similarity ) between each D Voxels with the remaining 4999 D Voxels . Next, we estimated the number of times 241 (Np) the distribution of similarity for observed RSFA values (DSimilarity1) is statistically different than the 242 permuted distributions of similarities (DSimilarity 2-5000) using Wilcoxon rank sum test. The ratio Np/5000 243 provided a level of significance, e.g. a value < 0.05 suggested that the distribution of age-RSFA residual 244 values is not as predicted by a model with pure randomness (at significance level p<0.05) and suggests 245 an association between age and RSFA residuals. The procedure was applied separately for each of the 246 five models across all brain voxels, as well as for different tissue types (cerebrospinal fluid, grey matter 247 and white matter voxels with values above 0.4 in SPM's tissue probability maps). indicating that variability in blood pressure was not associated uniquely with aging over and above their 282 contribution to other factors in the analysis. Factor 2 was mainly expressed by heart rate and HRV 283 measures, where individuals with high subject scores had low resting pulse and high HRV metrics. 284 Subject scores were correlated negatively with increasing age (r = -.417, p<.001), consistent with 285 findings of age-related decrease in HRV ( Figure 3). Finally, Factor 3 was expressed negatively by HRV 286 and positively by WMH and systolic blood pressure, indicating that individuals with high subjects scores 287 were more likely to have high burden of WMH, high systolic blood pressure and low HRV (Figure 3). 288 Subject scores were associated positively with age (r = +.713, p<.001), suggesting that a portion of the 289 age-related decrease in HRV is coupled with increase in WMH and systolic blood pressure. The whole-group voxel-wise analysis of RSFA maps revealed brain regions with high vascular 294 reactivity including frontal orbital, inferior frontal gyrus, inferior frontal gyrus, dorsolateral prefrontal 295 cortex, superior frontal cortex, anterior and posterior cingulate, and lateral parietal cortex. We 296 observed age-related decrease in RSFA in the bilateral inferior frontal gyrus, bilateral dorsolateral 297 prefrontal cortex, bilateral superior frontal gyrus, primary visual cortex, cuneus, precuneus, posterior 298 and anterior cingulate, superior temporal gyrus, medial parietal cortex, and lateral parietal cortex. In 299 addition, we observed age-related decrease in RSFA in the proximity of ventricles and large vascular 300 vessels. 301

Controlling for Cerebrovascular Factors (Model II) 302
We observed significant correlations between age and the RSFA residuals after controlling for

Controlling for Cardiovascular Factors (Model III) 308
We observed no significant correlations between age and the RSFA residuals after controlling 309 for variability in CVH and covariates of no interest at an FDR-adjusted p-value of 0.05 (Figure 4, model 310 III), suggesting that CVH can explain sufficiently age-dependent variability in RSFA, at least at the level 311 of statistically-corrected voxels. 312

Controlling for Cardiovascular and Cerebrovascular Factors (Model IV) 313
We observed no significant correlations between age and the RSFA residuals after controlling 314 for variability in CVH, CBF and covariates of no interest at an FDR-adjusted p-value of 0.05 (Figure 4, 315 model IV), suggesting that CVH and CBF together explain sufficiently age-dependent variability in RSFA. 316

Controlling for Grey Matter Volume (Model V) 317
We observed significant correlations between age and the RSFA residuals after controlling for 318 grey matter volume (GMV) and covariates of no interest at an FDR-adjusted p-value of 0.05 (Figure 4, 319 model V), suggesting that GMV does not adequately explain variability in RSFA, at the voxel-wise level. 320

322
The medians of observed and permuted data did not differ significantly (p>.1 for all five models). In 323 terms of the distributions, the level of statistical significance decreased after controlling for 324 cardiovascular, cerebrovascular and GMV signals (p <.001, p<.001, p=.015, and p<.001 for models 1, 2, 325 3 and 5 respectively), see Table 2. The model considering jointly cardiovascular and cerebrovascular 326 signals (model 4) indicated a difference in the distribution of observed and permuted data (p = 0.016), 327 reflecting a small level of correlation between age and RSFA residuals in some voxels. It is unclear 328 whether the signal originated in a particular tissue type, so we repeated the permutation approach for 329 each tissue type separately (Table 2). For models 1, 2 and 5 the RSFA residuals were associated with 330 age across all three tissue types, suggesting that variability in cerebrovascular and grey matter cannot 331 account fully for the effects of age on RSFA in all tissue types. However, the models controlling for 332 cardiovascular health (Models 3 and 4) were not significant for grey matter and white matter tissue. 333 The analysis on CSF voxels was highly significant suggesting that any potential age-related effects on 334 RSFA not captured by cardiovascular and cerebrovascular signals on voxel-level are focal to CSF areas, 335 rather than grey matter or white matter. 336 Next, we turn to the correlations between age and residuals of the RSFA ICs. We focused on 360

Component-based analysis
ICs that showed age-related differences in the subject loading values (10 out of 18), after controlling 361 for CBF IC loading values, GMV IC loading values or CVH factor loadings ( Figure 6). 362

Controlling for Cerebrovascular Factors (Model II) 363
The associations between age and RSFA residuals after controlling for CBF loading values were 364 weaker in vascular ICs and abolished in GM ICs compared to the analysis with covariates only (Figure 6, 365 Model I vs Model 2). Unlike in the voxel-based analysis, this ICA approach suggests that CBF does explain 366 some age-related variability in RSFA across many networks, especially those in GM areas, which may be 367 due to reduced number of comparisons and improved characterisation of sources of signals in RSFA 368 and CBF data using ICA. 369

Controlling for Cardiovascular Factors (Model III) 370
After controlling for differences in CVH, RSFA residuals in two ICs (IC3 and IC7) were correlated 371 with age (uncorrected p-value at 0.05 significance level), although to a lesser extent compared to the 372 analysis with covariates only (Model III vs Model I), indicating that CVH can explain age-dependent 373 variability in most, but not all, RSFA ICs. 374

Controlling for Cerebrovascular and Cardiovascular Factors (Model IV) 375
We observed no significant correlations between age and the RSFA residuals after controlling 376 for variability in CVH and CBF (even at an uncorrected p-value of 0.05, see Figure 6), suggesting that 377 together, CVH and CBF can explain age-dependent variability in RSFA. 378

Controlling for Grey Matter Volume (Model V) 379
RSFA ICs adjusted for GMV ICs demonstrated reduced correlations between RSFA and age 380 (particularly RSFA ICs of grey matter territories), indicating that age-related differences in RSFA ICs can 381 be partly explained by grey matter atrophy. 382

Controlling for Grey Matter Volume independent of Cardiovascular Factors 383
Some degree of association between age differences in RSFA and grey matter atrophy is 384 expected given cardiovascular health has been linked to brain-wide atrophy ( . Therefore, to test whether the effects of brain atrophy on RSFA were 387 independent of the effects of CVH on brain atrophy, we controlled for the effects of CVH in GMV ICs. 388 Then we used the GMV residuals after fitting CVH to GMV IC loadings (i.e. GMV orthogonalised with 389 respect to CVH) to estimate RSFA residuals and subsequently their correlation with age ( Figure 6 The principle result of this study is to confirm the suitability of resting state fluctuation amplitude (RSFA) 396 to quantify vascular influences in BOLD-based fMRI signals, and to demonstrate that the age effects on 397 RSFA reflect variability in vascular factors rather than neuronal factors. We demonstrate that the effects 398 of age on RSFA can be sufficiently captured by the joint consideration of cardiovascular (based on ECG, 399 BP, WMH and BMI measures) and cerebrovascular factors (CBF from ASL). Variance in brain atrophy 400 (GM volume Figure 6 two factors associated with blood pressure and heart rate variability (factors 1 and 2, respectively). A 411 third factor expressed white matter hyperintensities, blood pressure, heart rate variability and body-412 mass index, suggesting a cerebrovascular origin. 413 These three factor indices of cardiovascular health explained most of the age-related variability 414 in RSFA, leaving little to no associations between age and RSFA residuals in grey matter regions (after 415 controlling for these cardiovascular signals). This suggests that differences in cardiovascular health 416 mediate most of the age effects on RSFA (Tsvetanov et al., 2015). Interestingly, each CVH factor was 417 associated with a distinct spatial RSFA pattern (Supplementary Figure 2) and collectively provided 418 additional explanatory value for the overall age-differences in brain-wide RSFA. Next, we turn to neural 419 and cerebrovascular contributions to BOLD. 420 Cerebrovascular signals and age-differences in RSFA  Figure 3). But, we also observed negative associations between RSFA and CBF in 446 inferior brain areas, mainly close to vascular territories i.e. the higher the RSFA the lower the CBF values 447 were in these regions (Supplementary Figure 3) the additional contribution of short-term but stable cardiovascular health signals (e.g. heart condition 464 or white matter hyperintensities), which are independent of cerebrovascular factors. RSFA reflects both 465 cardiovascular and cerebrovascular signals, which are associated with distinct spatial patterns (see 466 section Spatial distribution and age effects on RSFA). RSFA can help dissociate age-related differences 467 in cardiovascular, cerebrovascular and neural function in task-based BOLD signal, which is important 468 for using fMRI to understand the mechanisms of cognitive aging. 469 Grey matter volume and age-differences on RSFA There are limitations to the current study. In terms of cardiovascular health, there may be more 519 important measures that were not present in the CamCAN sample. Moreover, the analysis of heart rate 520 variability estimates was based on normal-to-normal beats (Vest et al., 2018). The difference between 521 NN-and RR-beat analysis is that the former considers the detection and exclusion of segments and 522 participants with atrial fibrillation and other abnormal beats. While NN-beat analysis optimises the 523 detection of unbiased estimates of cardiovascular health, it also precludes sensitivity to potential 524 effects of arrhythmia and abnormal heart beats on RSFA in our analysis, which might be relevant to 525 regions susceptible to pulsatility effects (Webb and Rothwell, 2014). 526 In terms of cerebrovascular signals, the use of ASL-based CBF measurements could be complemented 527 with individual-based arterial transit time measurement in order to improve the accuracy of ASL 528 imaging in older populations (Dai et al., 2017). There are also other means to assess cerebrovascular 529 function, including cerebrovascular reactivity, including CO2-inhalation-induced hypercapnia (Liu et  In the current study, RSFA was estimated from resting state fMRI BOLD data prior to collection of other 551 task-based fMRI scanning as in previous validation studies of RSFA (Kannurpatti and Biswal, 2008;552 Tsvetanov et al., 2015). Other means of RSFA-like estimates have been proposed for scaling BOLD 553 activation data using fMRI BOLD data at different non-resting cognitive states, e.g. during task periods 554 (Kazan et al., 2016) or fixation-/resting-periods succeeding task periods (Garrett et al., 2017). Given that 555 short periods of cognitive engagement can modulate the BOLD signal in a subsequent resting state scan 556 (Sami et al., 2014;Sami and Miall, 2013), future studies are required to generalise our findings to RSFA-557 like estimates derived from other types of fMRI BOLD acquisition. 558 Finally, this study has focussed on the effects of aging, but other studies aiming to understand individual 559 differences or drug effects in fMRI BOLD might be affected in a similar manner. Therefore, future 560 studies should consider the origins of the signal contributing to RSFA (cerebrovascular vs 561 cerebrovascular) and more broadly their influence in fMRI BOLD imaging studies. In the light of 562 increasing evidence of the role of cardiovascular and cerebrovascular factors in maintaining cognitive 563 function, future studies might even consider RSFA as a predictor, rather than just as a covariate of no 564 interest, when modelling the effects of interest (e.g. age or performance). Furthermore, while the 565 proposed approach is based on plausible neurophysiology that can be used to evaluate its contribution 566 to cognitive function, future studies could improve absolute quantification of neural function together 567 with its integration with deoxyhaemoglobin-dilution-based modelling (Davis et al., 1998;Hoge et al., 568 1999aHoge et al., 568 , 1999b Concluding remarks 573 Cardiovascular and cerebrovascular signals together predict the age differences in RSFA, establishing 574 RSFA as an important marker that can be used to accurately separate vascular signals from neuronal 575 signals in the context of BOLD fMRI. We propose that RSFA is suitable to normalize BOLD, and control 576 for differences in cardiovascular signals. This is particularly relevant to the research in neurocognitive 577 aging, and may reduce selection bias, for example by permitting the inclusion of individuals with a wider 578 range of hypertension, cardiovascular conditions or comorbidity. The use of RSFA as a mechanism to 579 adjust for confounds in BOLD-fMRI, or as a predictor, will allow the development of better models of 580