Distinct components of cardiovascular health are linked with age-related differences in cognitive abilities

Cardiovascular ageing contributes to cognitive impairment. However, the unique and synergistic contributions of multiple cardiovascular factors to cognitive function remain unclear because they are often condensed into a single composite score or examined in isolation. We hypothesized that vascular risk factors, electrocardiographic features and blood pressure indices reveal multiple latent vascular factors, with independent contributions to cognition. In a population-based deep-phenotyping study (n = 708, age 18–88), path analysis revealed three latent vascular factors dissociating the autonomic nervous system response from two components of blood pressure. These three factors made unique and additive contributions to the variability in crystallized and fluid intelligence. The discrepancy in fluid relative to crystallized intelligence, indicative of cognitive decline, was associated with a latent vascular factor predominantly expressing pulse pressure. This suggests that higher pulse pressure is associated with cognitive decline from expected performance. The effect was stronger in older adults. Controlling pulse pressure may help to preserve cognition, particularly in older adults. Our findings highlight the need to better understand the multifactorial nature of vascular aging.


Methods
Participants. Figure 1 illustrates the analytical strategy and the study design with the Cam-CAN cohort, n = 708 54,71 . The methods were carried out in accordance with guidelines approved by Cambridgeshire 2 (now East of England-Cambridge Central) Research Ethics Committee, who approved all experimental protocols. All participants gave full, informed, written consent. Participants were recruited from Cambridge City GPs, randomly selected from this complete population sampling frame. The detailed recruitment pathway is outlined in Supplementary Fig. 1. For a full list of exclusion criteria, see Supplementary Table 1. In brief, those participating represented the healthier and more advantaged spectrum within the population at all ages 54 . Self-reported general health was reported across four categories of: excellent, good, fair or poor. The diastolic and systolic blood pressure observations were excluded for one participant due to data entry errors. Education was reported across four categories of: none, GCSE or O-Level, A-Level, Degree (College or University). Medication status (binary on/off) was reported across four categories of drugs with cardiovascular relevance: [1] anti-hypertensives 69 , including angiotensin receptor blockers, angiotensin converting enzymes inhibitors, calcium channel blocking agents and thiazide diuretics; [2] beta blockers, including beta selective and non-selective beta blockers; [3] other diuretics, including loop and potassium sparing diuretics; [4] dyslipidemics drugs, including statins [67][68][69] . www.nature.com/scientificreports/ Vascular factors. Systolic and diastolic blood pressure, and heart rate, were measured using the A&D Medical Digital Blood Pressure Monitor (UA-774). Measurements were taken after at least 10 min of a participant being seated and repeated 3 times in succession. BMI was calculated using portable scales as weight (kg)/ height (m) 2 . Heart rate variability was based on the frequency-domain information of normal-to-normal beats and extracted from resting state electrocardiogram recordings while seated during a separate MEG scan. We separated low-and high-frequency components: high frequency heart rate variability (0.15-0.4 Hz) principally indexes parasympathetic vagal influences, while low frequency heart rate variability (0.05-0.15 Hz) indexes non-vagal and sympathetic nervous system influences 20,72 . These two branches of the autonomic nervous system exhibit different non-linear trajectories with age, and might relate differently to cognition 73 . Heart rate variability data was processed using the PhysioNet Cardiovascular Signal Toolbox 74,75 in MATLAB (Mathworks, MA). Following Tsvetanov et al. 36 , segments classified as atrial fibrillation were excluded and data for any participant with > 50% atrial fibrillation (n = 1) were excluded. The heart rate variability at low and high frequency, and BMI, were log-transformed to render them more Gaussian.
Behavioural tasks. Crystallized  Statistical analyses. Analyses were performed in R (version 4.0.2) and R-Studio 80 . For initial checks, the three observations of diastolic and systolic pressure were correlated with age, using the Pearson's product moment correlation coefficient. They were then used to calculate three sets of total blood pressure (systolic + diastolic) and pulse pressure (systolic−diastolic). The resulting scores, and the three observations of heart rate, were log transformed to conform more closely to Gaussian distributions. These three sets of blood pressure, pulse pressure and heart rate measures were then standardised (mean = 0, standard deviation = 1) and condensed into a single latent variable per domain, using confirmatory factor analysis (CFA) in the lavaan package 81 . Latent Figure 1. Schematic representation of the data processing and analysis pipeline to investigate shared and unique relationships between vascular and cognitive factors in the Cam-CAN dataset (n = 708). BMI body mass index, BP total blood pressure (systolic + diastolic), dia diastolic, Cattell 1-4 sub-scores across the four Cattell tasks, CFA confirmatory factor analysis, Discrepancy, the ability discrepancy, defined as LCF2 (crystallized) minus LCF1 (fluid), ECG electrocardiogram, EFA exploratory factor analysis, HR heart rate, HRV HF heart rate variability at high frequency, HRV LF heart rate variability at low frequency, LCF latent cognitive factor, LVF latent vascular factor, PP pulse pressure (systolic − diastolic), sys systolic, STW spot the word. www.nature.com/scientificreports/ variables reduce error and estimation bias, while increasing precision 82,83 . Note that the confirmatory models were saturated, where three indicators loading onto one latent variable gave zero degrees of freedom. Missing data were imputed using Full Information Maximum Likelihood in cases where data were recorded for at least one of the three domain-specific observations; where data were missing entirely, participants were omitted automatically (leaving n = 579 for blood pressure and pulse pressure; n = 580 for heart rate). Next, we sought to identify the optimal number of factors among all vascular variables using Exploratory Factor Analyses (EFA). EFA is a common multivariate statistical method used to uncover the structure of a large set of variables 84 . Thus, EFA was used to identify the smallest number of latent vascular factors (latent vascular factors) that can parsimoniously explain the covariance observed among all vascular variables (blood pressure, n = 579; pulse pressure, n = 579; heart rate, n = 580; heart rate variability, at low and high frequencies, n = 604; BMI, n = 587). EFA was performed with the Psych package and imputing missing data (n = 668) 85 . Allowing two variables per latent factor is the upper boundary limit for model identification, meaning that one-, two-and three-factor solutions can be explored for six vascular variables in the current study. Note that a three-factor model with 6 variables will be fully saturated, not allowing estimation of the absolute fit indices. Therefore, model validity was based on model comparisons using the chi-squared statistic (p < 0.05), i.e. using comparative fit indices to determine the optimal number of latent vascular factors. Factor score estimates for each latent vascular factor were then extracted from the winning model for further regression analyses, below.
To explore the robustness of the winning EFA vascular model (and its loadings), we performed an additional structural equation model that included cognitive variables too (see Supplementary Section A). To additionally investigate whether the EFA vascular model structure is influenced by age, we repeated the EFA and model comparisons on sub-groups of young (n = 158, 18-37 years), middle (n = 311, 38-67 years) and old (n = 199, 68-88 years) participants (Table 1; Supplementary Section B).
Observed cognitive variables were standardised and condensed into two latent variables, using confirmatory factor analysis. The two-factor structure was based on the established dissociation between crystallized and fluid intelligence 86 . Scores on the Proverbs (n = 655) and Spot the Word tests (n = 705) loaded onto one latent cognitive factor (LCF1), representing crystallized intelligence. Scores on the Cattell tests (n = 660) loaded onto LCF2, representing fluid intelligence. Missing data were imputed using Full Information Maximum Likelihood in cases where data were recorded for at least one observed variable, producing LCFs for n = 678. The difference between LCF1 and LCF2 was calculated to give the ability discrepancy 47 . The calculation of the ability discrepancy was based on three assumptions, namely that (1) measurement of fluid and crystallized intelligence is invariant to age, (2) the two are highly correlated in youth, and (3) crystallized measures do not change with age. We investigated these assumptions using moderated non-linear factor analysis, correlations and visualisations, in Supplementary Section C.
The latent vascular factors and ability discrepancy were standardised to allow interpretation in terms of standard deviations from the mean. Linear and quadratic age predictor terms were also standardised. The relationships between the latent vascular factors, ability discrepancy and age were examined in multiple linear regression, using complete case analysis (n = 655). The presence of outliers with undue influence, as identified with Cook's criteria 87 , motivated the use of robust linear regression, implemented in the MASS package 88 . We performed a series of regression models from simple to complex, all including general health, sex and education as covariates of no interest, but dropping effects that did not improve overall model fit. Fit was investigated with the Akaike Information Criterion, Bayesian Information Criterion and Sum of Squares. Results were reported at p < 0.05. To guide the interpretation of significance of parameters in the larger models, model specific p-values after Bonferroni corrections are also reported.
Five models were used to test different hypotheses. For model syntax, see Supplementary Section D. The first model examined the relationship between the ability discrepancy and the three latent vascular factors, ignoring any shared dependence on age, to reveal which latent vascular factor(s) make unique contributions to the ability discrepancy. The Second model investigated whether any relationships between the ability discrepancy and latent vascular factors remained over and above a second-order polynomial expansion of age, and/or whether any effects of latent vascular factors depended on age. Note that, since the latent vascular factors were highly correlated with age, if effects of latent vascular factors from Model 1 are no longer significant in Model 2, then this could simply be because age shares variability with the latent vascular factors. General health was covaried in Models 1 and 2, and taken forwards into further models if it significantly predicted the ability discrepancy. Models 1 and 2 were compared and if Model 2 was shown to better fit the data, then the age terms were taken forwards into further models. Model 3 investigated whether our findings could be explained by medication status. Model 4 accounted for the interacting effects of sex on Vascular factors with age 9,89-91 . Since model comparisons showed that medications did not improve overall fit in Model 3, medications were not specified in Model 4. Model 5 investigated whether latent vascular factors interact with each other in order to determine the ability discrepancy. Since model comparisons showed that the inclusion of Sex interaction terms did not improve overall fit in Model 4, these interactions were not perpetuated to Model 5.
We also investigated whether the relationships between latent vascular factors and the ability discrepancy score, as explored in regression models, were robust to the effects of age on observed vascular measures. The EFA on vascular health was repeated over sub-groups of young, middle and old participants, and the resulting latent vascular factors were input to regression Models 1-5 (Supplementary Section B).

Results
Participants. Characteristics of the 708 participants in the Cam-CAN Phase 2 are outlined in Table 1. Rate of missing data varied between 0 and 18% (see Table 1). When the cohort (n = 708) was split into three age groups, the proportion of each on medications was: 0% of the younger (18-37 years); 12.6% of the middle ( Vascular factors and age. Diastolic and systolic blood pressure increased with age ( Supplementary Fig. 6).
Within each confirmatory factor analysis on the repeated measures of blood pressure, heart rate and pulse pressure, there were positive associations between all observed and latent variables, with p < 0.001 for all factor loadings. The resulting latent variables for blood pressure, heart rate and pulse pressure showed significant positive associations with age ( Supplementary Fig. 7). BMI also correlated significantly positively with age, while heart rate variability at both frequencies showed a significantly negative association ( Supplementary Fig. 7).

Vascular factor analysis.
We used EFA to estimate models with one, two and three factors. The two-factor model did not converge well. The three-factor model was fully saturated (no residual degrees of freedom). Model comparisons showed two factors to fit better than one (∆ X 2 = 236.57, p < 0.001), and three factors to fit better  www.nature.com/scientificreports/ than two factors (∆ X 2 = 95.04, p < 0.001). We confirmed the validity of the three-factor model in combination with cognitive measurements (see Supplementary Section A). The estimates of factor loadings and covariances for the EFA vascular model are visualised in Fig. 2A and detailed fully in Supplementary Table 5. Blood pressure loaded strongly onto the first latent factor, with a small contribution from BMI. Pulse pressure loaded strongly onto the second latent factor, with a small negative contribution from heart rate. Both frequencies of heart rate variability loaded similarly onto the third latent factor. All latent vascular factors correlated significantly with age (Fig. 3).
To explore whether the model structure remained consistent with age, we repeated EFA on three sub-groups of young, middle and old aged participants (Supplementary Section B). In all age groups, the three-factor model consistently fit best (Supplementary Table 2). Latent vascular factors produced in the age group specific and whole sample EFA correlated highly (r > 0.71, p < 0.001) (Supplementary Fig. 4).
Ability discrepancy. The cognitive variables (Supplementary Fig. 8) were entered into a two-factor CFA and produced latent cognitive factor 1, representing crystallized intelligence, and latent cognitive factor 2, representing fluid intelligence (Fig. 2B). The ability discrepancy was calculated by subtracting the participant loadings on the fluid factor from those on the crystallized factor 47 . The calculation of the ability discrepancy was based on three assumptions, firstly that the measurement of fluid and crystallized intelligence is invariant to age. Moderated non-linear factor analysis suggested that the measurement model did not differ substantially across the continuous covariate of age, warranting us to use the factor scores across the lifespan (Supplementary Section C). On the second and third additional assumptions, crystallized and fluid intelligence correlated highly in young adults, and crystallized intelligence remained stable with age (Supplementary Section C). The ability discrepancy showed a strong positive association with age (Fig. 3). It also correlated significantly with the three latent vascular factors, with substantial effect sizes (Fig. 3).
Compared to Model 1, Model 2 (n = 655, DoF = 634, residual standard error = 0.65) fit the data better (Supplementary Table 7). Model 2 revealed that the main effects of latent vascular factors 2 and 3 did not remain significant when accounting for age. The linear effect of age was significant, std β = 0.711, SE = 0.047, p < 0.001. More interesting was a significant interaction between latent vascular factor 2 and the quadratic effects of age  Table 5. B The two-factor Confirmatory Factor Analysis model of cognition. BMI body mass index, BP total blood pressure, HR heart rate, HRV HF heart rate variability at high frequency, HRV LF heart rate variability at low frequency, LCF latent cognitive factor, LVF latent vascular factor, PP pulse pressure, STW spot the word.  Fig. 4, by splitting the data into three age groups. It can be seen that the positive relationship between latent vascular factor 2 and the ability discrepancy is over 7 times stronger in the older group compared with the two younger groups.  Table 8). Given the lack of evidence supporting these more complex models, any significant parameters in Models 3-5 (Supplementary Tables 9-11) should be considered as suggestive only, and may not survive correction for multiple comparisons.

Discussion
In this study, we show that changes in vascular systems across the lifespan have multifactorial effects on cognitive function. There are three key observations. First, we identify three latent vascular factors that broadly dissociate the autonomic nervous system from distinct components of blood pressure. Second, these factors make distinct contributions to age-related cognitive decline, as indexed by the "ability discrepancy" score 47 . Third, the pulse pressure factor was particularly associated with the cognitive ability discrepancy, increasingly so for older adults. This remained even after controlling for the use of hypertensive medications and the covariates of sex, education and general health. Importantly, the effect of pulse pressure was independent of other latent vascular factors. We propose that steps to maintain lower pulse pressure may help to preserve cognitive function into old age.
Three components of vascular health. A single composite measure was insufficient to capture vascular health in our exploratory analyses, and the model fit was better with three factors. The evidence for these multiple latent vascular factors is consistent with previous model-based and data-driven approaches 35,36,[92][93][94][95] . There was no evidence that the number of factors changed with age, at least in the sense of fewer factors being needed for optimal fit in young, middle or older age sub-groups. The three latent vascular factors were composed predominantly of two major blood pressure variables and an autonomic nervous system variable; a decomposition that agrees with previous studies 36,95 . This decomposition also mimics established models of cardiovascular health (discussed below). The unsupervised construction of these latent vascular factors highlights their distinct contributions, which may involve different pathways and require different interventions.
The first factor, latent vascular factor 1, expressed total blood pressure, which is the steady state component of blood pressure (Fig. 2). This component is proposed to be mainly influenced by cardiac output and peripheral vascular resistance 64 . The additional contribution of BMI to latent vascular factor 1 fits well with early work showing a strong correlation between BMI and the steady component of blood pressure 59 . Latent vascular factor 1 was positively associated with age and the ability discrepancy (Fig. 3), however it did not significantly predict the ability discrepancy, over and above other latent vascular factors, in Model 1 (Supplementary Table 4). Consistent www.nature.com/scientificreports/ with these observations, Lefferts et al. 62 showed that steady blood pressure no longer predicts cognition over and above the effects of pulsatile blood pressure and covariates. We previously found that the steady component of blood pressure is associated with age-related cerebrovascular dysfunction of the sensorimotor regions, independently of the pulsatile component 36 . This suggests a unique contribution of steady blood pressure to brain health, with regional specificity. Future work should establish the possibility of a specific contribution of steady state blood pressure to brain functioning, and whether this varies across the lifespan 96 . The second component, latent vascular factor 2, expressed pulse pressure, with some contribution from heart rate (Fig. 2). Latent vascular factor 2 was positively associated with age (Fig. 3). Given the predominant loading by pulse pressure, latent vascular factor 2 likely represents a cerebrovascular element 97,98 . This is also consistent with previous findings that pulse pressure and white matter lesion burden expressed a common latent cerebrovascular factor 36 .
Latent vascular factor 3 expressed resting heat rate variability (Fig. 2), which indexes a specific component of the autonomic nervous system 22,99,100 . The decomposition of autonomic nervous system signals separately from cerebrovascular health signals (latent vascular factors 1 and 2) confirms that these are distinct, yet partly correlated, constructs of vascular ageing 101 . The convergence of low and high frequencies of resting heart rate variability onto latent vascular factor 3 does not necessarily rule out frequency-specific effects on cognition, or frequency-specific effects of task-based heart rate variability modulation/reactivity 102,103 . Latent vascular factor 3 associated negatively with age, consistent with previous studies on heart rate variability.
Pulse pressure and age-related cognitive function. A novel aspect of our work was to simultaneously relate the three latent vascular factors to age-related differences in cognition, specifically the cognitive discrepancy score. In the absence of longitudinal data, this discrepancy is arguably (see below) a better estimate agerelated change than raw individual differences in fluid intelligence. When relating directly to cognitive ability, only latent vascular factors 2 and 3 made unique contributions. Note that this does not mean latent vascular factor 1 has no relationship with cognitive ability; only that we cannot distinguish any such contribution from those of factors 2 and 3. The negative relationship for latent vascular factor 3 shows that higher heart-rate variability is associated with lower ability discrepancy, i.e., more variable heart rate is associated with less discrepancy, i.e., fluid intelligence that is closer to what would be expected from crystallized intelligence.
However, the relationship between latent vascular factor 3 and cognitive discrepancy was no longer significant when adjusting for age. This is consistent with cross-sectional studies where the association between resting heart rate variability and executive functions is accounted for by age and systemic vascular health [104][105][106] . Future studies should investigate how the shared variance between heart rate variability, systemic vascular health, age and cognition is linked to changes in cerebral blood flow, tissue integrity and neural function 9,20,58 . Heart rate variability is www.nature.com/scientificreports/ also theorised to link to domain specific measures of cognition, including emotional regulation [20][21][22]91,107 . Future research should also explore whether heart rate variability affects emotional regulation independently of other vascular factors, and whether this relationship changes with age. Latent vascular factor 2 made a unique, positive contribution to the ability discrepancy, consistent with higher pulse pressure being detrimental to cognitive ability. Like latent vascular factor 3, this contribution was no longer significant when adding age to the model. This could be because age is the true driver of ability discrepancy, or that age and latent vascular factor 2 are so highly correlated that we can no longer detect a unique effect of the latter. More importantly, we did find a significant interaction between age and latent vascular factor 2. This was a quadratic effect, consistent with pulse pressure being especially important for cognition in old age, rather than changing linearly with age. These findings are consistent with a growing body of literature suggesting that pulsatile, rather than steady, blood pressure is an important factor for brain health and higher cognitive functions 15,16,62,[108][109][110] . Future work needs to evaluate whether maintaining normal pulse pressure across the lifespan is the mediating factor of cognitive function and plays a role in the increasing relationship between brain function and cognition in old age [111][112][113][114][115] .
The mechanism by which pulse pressure relates to the ability discrepancy has yet to be identified. Pulse pressure has been proposed to be associated with a trigger point of a positive feedback loop of rising arterial stiffness and pressure that penetrates increasingly into deep brain tissue 11 . This causes a cascade of molecular and cellular damage to cerebral vessels, which ultimately injures the blood brain barrier, promoting the aggregation of beta-amyloid [116][117][118] , phosphorylated tau 119 and white matter hyperintensities 11,97,98,120 . Pulse pressure induced hippocampal damage is theorised to result in impaired episodic memory 98 , while it also uniquely contributes to cerebrovascular dysfunction in frontoparietal regions 36 . The recruitment of frontoparietal regions is of particular interest here, given that their involvement in fluid ability processing is partly explained by age-related hypoperfusion in these areas 121 . Separately, pulse pressure is associated with amyloid-dependent hypometabolism in frontoparietal regions 122 , which has in turn been associated with the ability discrepancy score 47 . Our research adds to this evidence by showing, in a population-based lifespan cohort, that pulsatile, rather than steady state blood pressure, is a predictor of cognitive decline (ability discrepancy), and this is particularly so for older adults. We propose that pulse pressure links to cognitive ageing through a distinct mechanism of cellular and molecular changes in cerebral vessels 36,98,121 . We further highlight pulse pressure as an emerging therapeutic target to prevent cognitive decline in ageing 97 .

Limitations.
Our study has limitations. The results were based on a population-based cross-sectional cohort and cannot directly speak to longitudinal ageing, i.e. individuals' progression over time. For example, crosssectional data may also be confounded by generational effects, such as the general increase in education and a decrease in blood pressure seen across recent decades. While we tried to approximate cognitive change via the discrepancy score (see below), our conclusions are restricted to the effects of age and its correlates, as assessed across individuals, and we cannot rule out the possibility that differences in vascular health are the consequence rather than the cause of differences in cognitive ability. Nonetheless, though the cross-sectional nature of our study cannot speak directly to the expansive body of literature on effects of mid-life blood pressure on late-life cognition 109,123,124 , our findings can generate hypotheses to test in longitudinal datasets. It should also be noted that the population-based adult-lifespan sample (18-88 years) used here is likely to be healthier, with lower variability in cardiovascular function, than samples used in other reports 54 .
The discrepancy between an individual's score on fluid versus crystallized intelligence was used to approximate cognitive decline, on the assumptions that 1) measurement of fluid and crystallized intelligence is invariant to age, 2) the two are highly correlated in youth, and 3) crystallized measures do not change with age. Each of these assumptions was tested and satisfied (see Supplementary Section C). In brief, we showed that the measurement model for fluid and crystallized intelligence was invariant to age. Secondly, though crystallized intelligence showed some increase from youth to mid-life (which could be generational), and though it has previously been shown to decline in very old age 70 , any such effects of ageing were much smaller than those on fluid intelligence. Finally, the two measures were highly correlated in young adulthood. It is possible that this high correlation means that some aspects of general cognitive ability are lost when subtracting them, though the assumption here is that this shared variance largely reflects age-invariant individual differences (such as genetics), and so is less relevant to our present question about vascular effects on cognitive ageing. Indeed, it has been argued that accounting for crystallized abilities helps adjust scores on tests of fluid ability that may be artificially lower than expected, e.g. owing to verbal materials and/or complex instructions in such tests 125 .
We estimated a subset of potential vascular factors, based on a limited set of measures on a single visit. Our measures were relatively easy to acquire in practice on a large scale. The reactivity of the autonomic system to an event or stressor, which may be cognitive, emotional or physical in nature, e.g. phasic heart rate variability 102 , could prove more sensitive to the resting state heart rate variability estimates used here. It is also possible that there are more than three latent vascular factors, but we cannot test that here with six vascular measures 84 , and to do so would require a greater number of vascular measures to be collected. For example, future work could test whether cholesterol levels 126 comprise an independent factor, or instead load on one or more of the three latent vascular factors identified here, or examine lifestyle factors such as physical activity, smoking and socioeconomic status.
It may seem surprising that we found no evidence that medications related to vascular health contribute to ability discrepancy, or at least modulate the effects of our latent vascular factors on ability discrepancy. This may be because the indirect effects of medication on ability discrepancy are mediated fully through their direct effects on latent vascular factors. Alternatively, it could be that, while medications affect current latent vascular factors, they may be given too late to prevent pre-medication levels of vascular factors like pulse pressure from already www.nature.com/scientificreports/ causing irreversible effects on cognitive ability 127 , or that the effects of chronic stable medication are mitigated by homeostasis. Either way, future research could investigate the mechanisms through which latent vascular factors mediate cognitive change, for example through damage to cerebral vessels, changes in brain perfusion or neuroinflammation, perhaps by direct manipulation of medications. In summary, we show that vascular ageing has multifactorial relationships with cognitive ageing. Of the three latent vascular factors, an increase in the factor expressing pulse pressure was uniquely associated with the cognitive discrepancy score, and this relationship was stronger for older adults. We suggest that maintaining low pulse pressure may help to preserve cognitive function into old age.