Aortic pulse wave velocity improves cardiovascular event prediction: an individual participant meta-analysis of prospective observational data from 17,635 subjects

Abstract

Objectives

To determine whether aortic pulse wave velocity (aPWV) improves prediction of cardiovascular (CVD) events beyond conventional risk factors.

Background

Several studies have shown that aPWV may be a useful risk factor for predicting CVD but have been underpowered to examine whether this is true for different sub-groups.

Methods

We undertook a systematic review and obtained individual participant data from 16 studies. Study-specific associations of aPWV with cardiovascular outcomes were determined using Cox proportional hazard models and random effect models to estimate pooled effects.

Results

Of 17,635 participants, 1,785 (10%) had a cardiovascular (CVD) event. The pooled age- and sex-adjusted hazard ratio [95% CI] per SD change in log_e aPWV was 1.35 [1.22, 1.50, p<0.001] for coronary heart disease (CHD), 1.54 [1.34, 1.78, p<0.001] for stroke, and 1.45 [1.30, 1.61, p<0.001) for CVD. Associations stratified by sex, diabetes and hypertension were similar, but decreased with age (1.89, 1.77, 1.36 and 1.23 for ≤50, 51–60, 61–70 and >70 years respectively, p_interaction <0.001). After adjusting for conventional risk factors, aPWV remained a predictor: CHD 1.23, [1.11, 1.35 p<0.001]; stroke 1.28, [1.16, 1.42 p<0.001]; cardiovascular events 1.30 [1.18, 1.43, p<0.001]. Reclassification indices showed the addition of aPWV improved risk prediction (13% for 10 year CVD risk for intermediate risk) for some sub-groups.

Conclusions

Consideration of aPWV improves model fit and reclassifies risk for future cardiovascular events in models that include standard risk factors. aPWV may enable better identification of high-risk populations who may benefit from more aggressive cardiovascular risk factor management.

There is considerable interest in refining cardiovascular risk prediction in order to target better preventative therapy amongst those individuals considered by current guidelines to be at low or moderate risk. A number of additional putative cardiovascular biomarkers have been identified including C-reactive protein, carotid intima-media thickness, and a variety of genetic variants^1,2. However, these factors seem to add little to existing risk estimates, such as that derived from the Framingham Heart Study^1,3,4. Recently, aortic stiffness has emerged^5,6 as a potential additional candidate, and reference values have now been published^7,8.

Aortic stiffness can be assessed by a variety of non-invasive methods. One of the most frequently used methods is carotid-femoral (aortic) pulse wave velocity (aPWV)⁹. Data from prospective observational cohort studies indicate that aPWV relates to future cardiovascular risk even after accounting for other accepted cardiovascular risk factors. However, the extent to which aPWV improves risk prediction, whether it does so equally for cardiac and cerebral events, and if it differs by sub-groups, is unclear as most studies were under-powered to examine these issues. A recent meta-analysis using summary published data found aPWV predicted cardiovascular events but could not examine for sub-group effects at an individual level or calculate the additional prognostic value of aPWV¹⁰. We have undertaken a systematic review and have used data from both new published and unpublished cohorts with measures of aPWV and incident cardiovascular disease to conduct an individual participant meta-analysis. We aimed to address the questions of whether having information on aPWV for both unselected, population-based individuals and patients with manifest disease improved the prediction of future cardiovascular events; whether risk prediction varied by sub-groups; and whether improved risk prediction was additive to standard risk factors and how this may vary by population.

Methods

We used the PRISMA 2009 guidelines¹¹ and undertook a systematic search (details in appendix 2). We pre-specified the following inclusion criteria: (a) the study had to be a cohort design with a minimum of one year follow-up; (b) aortic stiffness had to be assessed by direct measurement of carotid-femoral PWV; (c) the study had to be able to provide relevant outcome data including all cause mortality, coronary heart disease (CHD) (myocardial infarction or revascularization or as defined by the studies) and stroke events, or CHD and stroke combined (cardiovascular events). Where available, we also tried to differentiate between fatal and non-fatal events, though not all studies collected data on non-fatal events.

Anonymized individual-level subject data were requested for each study including aPWV, a range of covariates (including age, sex, blood pressure, body mass index, smoking status, lipids, creatinine, co-morbidity) and time to the various endpoint events or censoring.

Statistical analysis

Baseline characteristics were summarized for each study sample and reported as mean (SD) and number (%) for continuous and categorical variables respectively. For skewed continuous variables the median and inter-quartile range (IQR) are stated. aPWV varies according to the software algorithm used and the approach to transit distance measurement. Since our main aim was to examine the relative value of aPWV within a study, and then pool these estimates, we used the z-score of log_e-transformed aPWV in the analyses as aPWV values were positively skewed. Thus, effect estimates for each study reflect the change in risk of an outcome for a one standard deviation increase in log_e aPWV from average in that population.

Outcome measures were all cause mortality, cardiovascular mortality, CHD events, stroke, and cardiovascular events (CHD and stroke). For each, Cox proportional hazards models were fitted that estimated the hazard ratio (HR) of aPWV (a) adjusted for age and sex, (b) additionally adjusted for systolic blood pressure (SBP), (c) additionally adjusted for total cholesterol, HDL-cholesterol, smoking status, diabetes and current antihypertensive medication¹². We also repeated these models but replacing systolic blood pressure with pulse pressure. Continuous covariates were expressed as cohort-specific z-scores. All models were also stratified by race for one study which had a pronounced split in Caucasian and African-American populations (but see below as regards sub-group analysis for ethnicity). We checked whether the association of aPWV with outcomes was linear by visual inspection of graphs of aPWV quintiles against the corresponding hazard ratio, and formal testing for non-linearity using fractional polynomials¹³. The proportional hazards assumption was assessed using tests based on Schoenfeld residuals in models fitted separately to each study.

Models were fitted separately for each study and the fully or partially adjusted estimates pooled using random effects meta-analysis, in order to account for between-study heterogeneity. Forest plots for each model and outcome show the study-specific effects and the overall pooled estimate, with 95% confidence intervals and random-effects weightings. In sensitivity analyses we fitted all models using firstly inverse aPWV, and secondly the untransformed data (hence this is for 1 m/sec increase in aPWV) but still z-scored within studies. The sensitivity of effects to missing covariate data was examined by repeating analyses using only the 13 studies with all covariates measured (see appendix 3). The presence of small study effects and publication bias were examined using both visual inspection of funnel plots and formal Egger tests. We also considered the influence of each individual study on the pooled meta-analysis effect estimate to examine if any one study had undue influence as an outlier.

The protocol pre-specified analyses of the following potential effect modifiers: sex, age group, type of population (healthy versus disease group), smoking status, renal function measured by the MDRD¹⁴ estimated glomerular filtration rate (eGF R) (eGFR ≥90 versus eGFR<90 ml/min/1.73 m²), diabetes and antihypertensive use at baseline. For each potential effect modifier considered, we estimated the strata specific effect of PWV in each study separately. These estimates were pooled across studies, which were then tested to see if the effect of PWV differed between strata. For type of population, which is a study level variable, we used meta-regression to test for differences in effect of PWV between clinical and population-based studies. Post-hoc, we also tested for any potential differences in the results dependent on either (a) the method used to measure distance in calculating aPWV or (b) ethnic differences related to participants from the Far East versus European and North American populations (appendix 3).

To compare the discriminatory power of aPWV against simpler haemodynamic measures such as SBP, or other established risk factors, the fully adjusted models were fitted with and without log_e aPWV. We calculated study-specific measures of discrimination (Harrell’s C-index and Royston & Sauerbrei’s D-measure) and then pooled these statistics weighted by the number of events¹⁵.

We also examined reclassification of subjects to risk groups due to the addition of aPWV to conventional cardiovascular risk factors (net reclassification index – NRI)¹⁶. We used reclassification based on 5-year risk as not all studies had sufficient length of follow-up to use the standard clinical cut-points based on 10-year risk. Risk cut points were calculated in each study, based on quartiles of predicted risk from the model without aPWV, considering only those individuals with events. These cut-offs were then applied to the whole study sample. Subjects were ranked by predicted risk from the models firstly with and secondly without aPWV and assigned to low (1st quartile), medium (2^nd and 3^rd quartiles) and high-risk groups (4^th quartile). Only studies with at least two participants experiencing events within 5 years were included in this reclassification exercise. Individuals experiencing an event after 5 years were censored. The number of events available to calculate discrimination statistics is therefore less than the number available to fit the Cox proportional hazards models. Categorisations under the two models were cross-tabulated, and the number of subjects moving in the correct direction (up for those experiencing events and down for those not experiencing events) on inclusion of aPWV in the model were counted. The overall percentage of correct reclassifications was combined from those with an event, and those without, across all the studies. We calculated the net reclassification improvement based both on all participants and limited to those at intermediate risk (i.e., in risk quartiles 2 and 3)¹⁶. We also derived the integrated discrimination improvement (IDI) which measures improvement in risk prediction on a continuous scale and is independent of the choice of cut-points for risk categorisation. We also undertook a series of sensitivity analyses (see appendix 3).

Results

The flow chart of the selection of papers in the systematic review is shown in Figure 1. From a potential of 29 papers assessed for eligibility, 9 were dropped that were either duplicates or did not fulfil the eligibility criteria on further examination. A further 7 studies were unable to supply individual participant data^5,17–22. The resulting 13 eligible studies (see web table 1 for details) for which the original investigators were willing to provide data access were supplemented by 3 additional studies that were not formally published and were identified through other methods^{23,24,25,26,27,28,29,30,31,32,33,34,35,36,37}. The study by Cruickshank and colleagues recruited two cohorts, a population-based sample and a sample of diabetic patients; therefore, the study was considered as two study cohorts in the analyses, resulting in 17 cohorts in the main analysis.

Figure 1. Flow diagram illustrating the process of study identification.

There was a mix of cohorts, with 8 of the 17 cohorts based on patients with known diseases, and the rest from population-based studies. Baseline characteristics of the various cohorts are shown in Web table 1. Most studies included approximately equal numbers of men and women, except for the Caerphilly Prospective Study, which included only men. All except four studies had information on all adjustment variables, and all except five studies had event rates and follow-up times for all outcome measures. The distribution of raw aPWV measures across the studies is shown in Web figure 1.

In Cox proportional hazards models, log_e aPWV was linearly associated with risk for each of the outcomes and proportional hazards assumptions were valid. Table 1 shows the HRs for the pooled associations of aPWV with our outcome measures for each of the three models. For all outcomes, log_e aPWV was strongly associated with increased risk, though additional adjustment resulted in some attenuation. The study-specific HRs of combined cardiovascular events for aPWV together with the pooled estimate are shown in Figure 2 (a) adjusted for age and sex, and in Figure 2 (b) fully adjusted for all risk factors. Funnel plots and formal Egger tests fitted to estimates from the simple age and sex-adjusted models indicated limited problems of small study effects, with any differences likely to be due to some studies having limited numbers of events. There were no overly influential studies.

Table 1.

Pooled adjusted hazard ratios (95% CIs) of a 1 SD increase in log_e-transformed aPWV for all cause mortality, cardiovascular mortality, CHD events, stroke events and cardiovascular events.

Outcome	Model 1* (Hazard ratio, 95% CI)	Model 2* (Hazard ratio, 95% CI)	Model 3* (Hazard ratio, 95% CI)
CHD events (n=1195)	1.35 (1.22, 1.50)	1.32 (1.18, 1.48)	1.23 (1.11, 1.35)
CVD events (n=1785)	1.45 (1.30, 1.61)	1.37 (1.23, 1.52)	1.30 (1.18, 1.43)
Stroke events (n=641)	1.54 (1.34, 1.78)	1.37 (1.21, 1.54)	1.28 (1.16, 1.42)
CVD mortality (n=395)	1.41 (1.27, 1.56)	1.35 (1.20, 1.53)	1.28 (1.15, 1.43)
All cause mortality (n=2041)	1.22 (1.16, 1.27)	1.20 (1.15, 1.26)	1.17 (1.11, 1.22)

^*Model 1 adjusts for sex and age group; Model 2 adjusts for sex, age group and systolic blood pressure; Model 3 additionally adjusts for other risk factors (cholesterol, HDL-cholesterol, smoking status, presence of diabetes, anti-hypertensive medication), stratified by race in the Sutton-Tyrell study. Not all studies had data on every risk factor.

Figure 2. Forest plot for aPWV and combined cardiovascular events adjusting for various risk factors.

(a) Adjustment for age and sex

Log_e aPWV is shown. Size of box represents the study specific weight for the meta-analysis. BTC – Belgian Transplant Cohort, CORD – Calcification Outcome in Renal Disease, CaPS – Caerphilly Prospective Study. ES=effect size

(b) Adjustment for age, sex and other cardiovascular risk factors.

Log_e aPWV is shown. Adjusted for age, sex, systolic blood pressure, total cholesterol, HDL-cholesterol, diabetes, and antihypertensive use. Data from BLSA were excluded as there were too few events. Size of box represents the study specific weight for the meta-analysis. BTC – Belgian Transplant Cohort, CORD – Calcification Outcome in Renal Disease, CaPS – Caerphilly Prospective Study, CVD – cardiovascular disease. ES=effect size

There was no evidence that the increased risk associated with aPWV was modified by sex, population type, smoking status, renal function, baseline diabetes or anti-hypertensive use. However, aPWV was more strongly related to the risk of CHD (p_interaction=0.001) and cardiovascular events (p_interaction=0.004) in younger participants. This age effect remained in the fully adjusted models for both CHD and cardiovascular events (p=0.006 and p=0.03 respectively). The results of sub-group analyses for combined cardiovascular events are shown in Figure 3, and in the supplementary figures for other outcomes (Web Figure 2a–d).

Figure 3. Forest plot for a PWV with cardiovascular events according to pre-specified subgroups.

Log_e aPWV is shown. Data are adjusted for age and sex where applicable. Data from BLSA were excluded as there were too few events. ES=effect size

Results from the sensitivity analyses that used inverse aPWV and the untransformed aPWV did not materially differ from those using log_e aPWV (data not shown). We found that the models that used pulse pressure rather than systolic blood pressure were essentially the same though the hazard ratios were attenuated for stroke (data not shown), but with clearly overlapping 95% confidence intervals. A change in aPWV of 1 m/s (weighted mean and SD were 10.1 and 3.3 m/s) was associated with a hazard ratio for cardiovascular events of 1.07 (95% CI 1.02, 1.12) for a male aged 60 years old who was a non-smoker, not diabetic, not on any blood pressure medication and with systolic blood pressure of 120 mm Hg, total cholesterol of 5.5 mmol/L and HDL cholesterol of 1.3 mmol/L. We found that the fully adjusted HRs were slightly attenuated in the models that were restricted to the studies with a full set of covariate data (results not shown).

The discrimination and reclassification statistics calculated to assess improvement in 5-year risk prediction associated with the inclusion of log_e aPWV in models are shown in Table 2. Small differences in C and D statistics and IDI indicated modest improvement in risk prediction when log_e aPWV was added to conventional Framingham risk factors. The integrated discrimination improvement presented evidence of improvements in discrimination for all outcomes when including log_e aPWV in the models. However, calculation of the net reclassification improvement for each outcome indicated improvements in reclassification that have some clinical relevance especially for those at intermediate risk (Table 3).

Table 2.

Improvement in discrimination^* and reclassification of 5-year risk prediction^α associated with including log aPWV as a risk factor.

Outcome	C-Index*			D-Statistic*			IDIα
	Without log aPWV**	With log aPWV	Change (95% CI)	Without log aPWV**	With log aPWV	Change (95% CI)	Event Numbers
All Cause Mortality	0.7012 (0.6897, 0.7127)	0.7057 (0.6942, 0.7172)	0.0046 (0.0010, 0.0082) p=0.013	1.2176 (1.1436, 1.2915)	1.2552 (1.1809, 1.3292)	0.0375 (−0.0672, 0.1422) p=0.483	n=1,023	0.00589 (0.0033,0.0085)
CVD Mortality	0.7476 (0.7252, 0.7699)	0.7566 (0.7342, 0.7790)	0.0098 (−0.0010, 0.0207) p=0.076	1.5031 (1.3341, 1.6722)	1.7043 (1.4708, 1.9379)	0.2012 (−0.0871, 0.4895) P=0.171	n=219	0.00632 (0.0024, 0.0103)
CHD events	0.6780 (0.6630, 0.6929)	0.6832 (0.6685, 0.6980)	0.0053 (0.0013, 0.0093) p=0.010	1.0734 (0.9777, 1.1691)	1.1147 (1.0183, 1.2111)	0.0413 (−0.0945, 0.1771) p=0.551	n=730	0.00366 (0.0010, 0.0063)
CVD events	0.6797 (0.6669, 0.6923)	0.6848 (0.6721, 0.6976)	0.0052 (0.0011, 0.0092) p=0.013	1.1436 (1.0649, 1.2222)	1.1965 (1.1174, 1.2756)	0.0529 (−0.0586, 0.1645) p=0.352	n=1,060	0.00797 (0.0047, 0.0112)
Stroke events	0.7243 (0.7054, 0.7433)	0.7315 (0.7123, 0.7508)	0.0072 (−0.0003, 0.0147) p=0.059	1.3649 (1.2314, 1.4985)	1.4171 (1.2848, 1.5494)	0.0522 (−0.1358, 0.2401) p=0.587	n=315	0.00954 (0.0042, 0.0149)

^*Harrell’s C-Index and Royston & Sauerbrei’s D-Measure

^**Clinical model includes systolic blood pressure, total and HDL-cholesterol, smoking status, diabetes and antihypertensive medication status.

^αIntegrated Discrimination Improvement. Event numbers for the IDI are reduced due to individuals with no event within 5 years being censored at this point.

Table 3.

Net reclassification statistics showing percentage change in 5-year risk prediction (and 5 & 10 year overall reclassification) associated with including log_e aPWV as a risk factor in the fully adjusted model; results shown for the whole sample and those at intermediate risk.

Whole Sample
Controls (Event=0)					Cases (Event=1)				5-year overall reclassification	10-year overall reclassification
Outcome	Whole Sample (n=)	Clinical Population (n=)	General Population (n=)	Aged ≤61 years (n=)	Whole Sample (n=)	Clinical Population (n=)	General Population (n=)	Aged ≤61 years (n=)	Whole Sample NRI (95% CI) (n=)	Whole Sample NRI (95% CI) (n=)
All Cause Mortality	0.66 (n=14,125)	1.11 (n=4,703)	0.44 (n=9,422)	−0.17 (n=7,011)	4.30 (n=1,023)	3.37 (n=356)	4.80 (n=667)	4.08 (n=147)	4.96 (4.11, 5.81) (n=15,148)	1.73 (0.87, 2.59) (n=12,837)
CVD Mortality	3.95 (n=9,275)	0.99 (n=2,618)	5.11 (n=6,657)	1.43 (n=3,975)	8.22 (n=219)	10.45 (n=67)	7.24 (n=152)	16.00 (n=25)	12.17 (10.68, 13.66) (n=9,494)	8.34 (7.17, 9.51) (n=10,271)
CHD events	0.28 (n=14,158)	3.05 (n=3,212)	−0.54 (n=10,946)	−0.84 (n=7,158)	4.66 (n=730)	7.55 (n=212)	3.47 (n=518)	8.77 (n=114)	4.94 (4.00, 5.88) (n=14,888)	3.03 (2.24, 3.82) (n=12,503)
CVD events	0.28 (n=13,828)	3.06 (n=3,104)	−0.52 (n=10,724)	−0.72 (n=7,092)	5.09 (n=1,060)	5.31 (n=320)	5.00 (n=740)	10.56 (n=180)	5.37 (4.38, 6.36) (n=14,888)	4.43 (3.53, 5.33) (n=12,503)
Stroke events	−0.04 (n=13,397)	0.33 (n=2,142)	−0.12 (n=11,255)	−0.73 (n=6,407)	9.52 (n=315)	8.49 (n=106)	10.05 (n=209)	13.79 (n=58)	9.48 (8.36, 10.60) (n=13,712)	5.60 (4.39, 6.81) (n=10,465)

Intermediate Risk Only (Quartiles 2 & 3)
Controls (Event=0)					Cases (Event=1)				5-year overall reclassification	10-year overall reclassification
Outcome	Whole Sample (n=)	Clinical Population (n=)	General Population (n=)	Aged ≤61 years (n=)	Whole Sample (n=)	Clinical Population (n=)	General Population (n=)	Aged ≤61 years (n=)	Whole Sample NRI (95% CI) (n=)	Whole Sample NRI (95% CI) (n=)
All Cause Mortality	5.49 (n=3,933)	7.69 (n=1,053)	4.69 (n=2,880)	9.92 (n=393)	9.18 (n=512)	10.67 (n=178)	8.38 (n=334)	0 (n=31)	14.67 (12.63, 16.71) (n=4,445)	6.14 (4.05, 8.23) (n=3,970)
CVD Mortality	11.68 (n=1.970)	15.72 (n=439)	10.52 (n=1,631)	17.86 (n=84)	11.43 (n=105)	13.33 (n=30)	10.67 (n=75)	37.50 (n=8)	27.17 (29.61, 37.73) (n=2,075)	24.27 (20.65, 27.89) (n=2,164)
CHD events	6.85 (n=3,929)	13.80 (n=1,196)	3.81 (n=2,733)	9.26 (n=994)	7.92 (n=366)	7.48 (n=107)	8.11 (n=259)	0 (n=50)	14.77 (12.41, 17.13) (n=4,295)	9.69 (7.61, 11.77) (n=3,346)
CVD events	5.99 (n=3,774)	15.65 (n=1,061)	2.21 (n=2,713)	7.72 (n=868)	7.97 (n=527)	4.43 (n=158)	9.49 (n=369)	7.25 (n=69)	13.96 (11.41, 16.51) (n=4,301)	13.05 (10.69, 15.41) (n=3,473)
Stroke events	5.80 (n=3,740)	8.66 (n=693)	5.15 (n=3,047)	8.66 (n=658)	13.38 (n=157)	11.32 (n=53)	14.42 (n=104)	20.83 (n=24)	19.18 (16.38, 21.98) (n=3,897)	10.89 (7.97, 13.81) (n=3,141)

The NRI for 10-year predicted risk was slightly lower than that for 5-year risk which may be due to the attenuation of the accuracy of predictions with increasing extrapolation beyond the actual period of observation. We did not find any evidence that any of the results for our various outcomes differed either by the method used to define the distance over which aPWV was calculated or whether the study populations came from the Far East as compared to Europe or North America (see appendix 3 for results from sensitivity analysis).

Discussion

The main finding of the current study is that aortic stiffness, assessed by measurement of aPWV, predicts future cardiovascular events and mortality, even after accounting for other established cardiovascular risk factors. The predictive value of aPWV was stronger in younger versus older subjects but was not modified by hypertension, smoking, sex, diabetes or kidney disease. Addition of aPWV into risk prediction models also increased the number of participants correctly classified, particularly amongst younger individuals at intermediate risk and improved the overall 10 year classification by 13%.

The optimal approach to cardiovascular disease screening and risk stratification remains controversial, with some favouring a strategy based on targeting high risk individuals³⁸ and others arguing for a population-based approach³⁹. The former strategy focuses on measuring traditional risk factors and the relative cost effectiveness of such an approach has not been assessed in clinical practice³⁸. Novel biomarkers may improve risk stratification. However, when these potential biomarkers have been entered into risk prediction models, such as Framingham, they do not appear to improve risk prediction very much beyond that already provided by established risk factors such as blood pressure, blood glucose and cholesterol. Interest has also focused on markers of tissue or end-organ damage such as carotid intima media thickness, which has been included in European Society of Hypertension (ESH) and European Society of Cardiology (ESC) guidelines⁴⁰. However, despite the recommendation in published guidelines, carotid intima media thickness is rarely measured in routine clinical practice and its utility remains controversial^3,41,42.

During the last 10 years, a large amount of evidence has accumulated demonstrating that arterial stiffness is an important risk factor for cardiovascular disease. Aortic stiffness can be assessed in a number of ways, but aPWV is regarded as the current gold-standard⁹ and has the most evidence available linking it to cardiovascular risk. aPWV can be assessed in a routine clinical setting using a number of commercially available devices, making it a potentially attractive cardiovascular biomarker. Indeed, assessment of arterial stiffness is included in the latest ESH/ESC guidelines⁴⁰ but the American College of Cardiology Foundation and American Heart Association felt that there was insufficient evidence to recommend measures of arterial stiffness for asymptomatic individuals⁴³.

Our results confirm those of a previous summary meta-analysis¹⁰ that aPWV predicts future fatal and non-fatal cardiovascular events. We have greatly extended this with the addition of new data, the ability to examine important sub-groups and by specifically calculating the prognostic value of aPWV over and above established risk factors. Following full adjustment, a 1 m/s increase in aPWV was associated with a 7% increased risk of a cardiovascular event for a 60 year old man (non-smoker, not diabetic, not on any blood pressure medication and with systolic blood pressure of 120mmHg, total cholesterol of 5.5mmol/L and HDL cholesterol of 1.3mmol/L). We have shown that aPWV was a stronger risk factor amongst younger individuals although it was still predictive in older individuals. This may be because individuals with stiff aortae who are susceptible to cardiovascular disease die younger (“healthy survivor effect”), other risk factors attenuate the effects of aPWV at older ages and/or systolic pressure is a better surrogate of aortic stiffness in older people than in younger people; therefore including aPWV in models already containing systolic pressure would be expected to add less predictive value. Indeed, the age-related rise in systolic pressure, and development of isolated systolic hypertension, closely mirrors the age-related rise in aPWV^7,44. Conversely, systolic hypertension in younger individuals seems to be driven predominantly by an elevated cardiac output and stroke volume, and as such systolic or pulse pressure is a poor surrogate for stiffness in the young⁴⁵.

Addition of aPWV to the adjusted cardiovascular prediction models only increased the C and D statistics to a modest degree, suggesting that aPWV may not add much to standard risk equations when all participants are analysed together. However, they are relatively insensitive methods for assessing the potential value of new biomarkers and do not specifically focus on people in whom better risk prediction is likely to make an important clinical difference i.e. those who are at moderate or intermediate risk⁴⁶. Indeed, many current guidelines advocate 10-year cardiovascular risk estimation and the targeting of therapy toward individuals whose estimated risk exceeds a particular threshold. However, refining estimation in those at high or low risk is unlikely to alter management or risk prevention in a substantial way. The performance of aPWV on the net reclassification improvement appeared more clinically informative in terms of risk stratification for those at intermediate cardiovascular risk and in younger subjects, though we have presented data on reclassifying subjects at both low (general population sample) and high (e.g. clinical sample) absolute risk for completeness.

Our results also suggest that aPWV may be a suitable target for novel risk reduction strategies. Although we did not investigate pathophysiological mechanisms underlying cardiovascular disease, previous studies suggest that aPWV attenuation is associated with improved survival⁴⁷. The majority of existing drugs do not appear to lower aPWV in a blood pressure-independent manner, but long-term blockade of the renin-angiotensin system⁴⁸ and novel agents targeting elastic fibre cross-linking⁴⁹ or calcification may afford some benefit. However, these strategies remain to be tested directly and remain speculative⁵⁰.

Our analysis has a number of limitations. Almost all the studies were from Caucasian or participants from the Far East, limiting the generalizability of these findings to other ethnic populations. A variety of different methods and devices were used to assess aPWV which are known to influence absolute values. However, we tried to minimize methodological influence by calculating study-specific effects and our analyses did not reveal significant heterogeneity between studies or devices. A sensitivity analysis examining the method used to calculate the distance for the carotid-femoral path failed to find any evidence of heterogeneity. We extrapolated the results from some short term studies to predict 10 year risk so these results should be treated with some caution given the limited long-term data. Our CVD outcome measure was primarily based on myocardial infarction and stroke so the predictive value of aPWV on heart failure has not been explicitly examined. We tried to include all published studies, but two large studies were not included. However, our observations appear consistent with those reported in excluded studies, and benefit from significantly larger sample size than any of the individually published studies. Recently, the Rotterdam study has published its own data on risk prediction⁵¹ and showed a similar 9% reclassification of intermediate risk group subjects. We were able to include data from several new studies including three unpublished studies, two of which have published previous data on their aPWV measures^23,24. All our estimates come from observational studies and a previous meta-epidemiological study found that the effects of cardiovascular biomarkers were stronger in such studies compared to randomised controlled trials⁵².

In summary, aPWV predicts future cardiovascular risk and improves risk classification adjusting for established risk factors. As aPWV can now be reliably and easily measured⁵³, it may serve as a useful biomarker to improve cardiovascular risk prediction for patients at intermediate risk. However, before its adoption can be recommended, randomized-controlled trials using aPWV to guide risk stratification and/or treatment are required to provide convincing evidence that this method has clinical value.

Supplementary Material

1

Web Figure 1 Box and whisker plot of raw PWV measures across all studies

Data from all 17 studies. Note that the Cruickshank study was considered as 2 separate samples for analysis (participants with and without diabetes).

Web figure 2: Forest plot for aPWV with other outcomes according to pre-specified subgroups

(a) All cause, (b) CHD events, (c) CVD mortality, (d) stroke