Introduction

Cardiovascular disease (CVD) is the leading cause of death worldwide [1,2,3]. The main independent risk factors for CVD are cigarette smoking, hypertension, diabetes, obesity and elevated total cholesterol or elevated low density lipoprotein (LDL) cholesterol [4,5,6]. It has been known for some considerable time that high dose radiotherapy is also associated with excess risk of CVD [7, 8]. More recently, it has become clear that there are also radiation-associated excess risks in the Life Span Study (LSS) of the Japanese atomic-bomb survivors [9, 10], and in a number of groups exposed at still lower levels of radiation dose and at lower dose rates [11]. A recent systematic review and meta-analysis of epidemiological studies highlighted evidence of association between radiation exposure and CVD at high dose, and to a lesser extent at low dose, with some indications of differences in risk between acute and chronic exposures [12]. There was inter-study heterogeneity, possibly a result of confounding or modifications of radiation effect by other factors, which complicates a causal interpretation of these findings [12]. Although a number of studies assessed in the previous review adjusted for many of the major lifestyle risk factors, relatively few studies undertook investigation of the modifying effect of these risk factors on the radiation associated CVD.

In this paper we assess effects of confounding by lifestyle, environmental, and medical risk factors, and also investigate evidence for modifying effects of these variables on radiation dose response.

Methods

The data used are those assembled in a recent systematic review [12]. In brief, the review was conducted, and reported according to PRISMA and registered in PROSPERO (https://www.crd.york.ac.uk/prospero/) (reg. no. 202036). PubMed/MEDLINE, Embase, Scopus, and Web of Science: Core Collection were used to systematically search the literature, with no limits applied (date, language), on 6th October 2022. We excluded animal studies, and any study without an abstract. The database search yielded a total of 15,098 articles.

Outcomes

CVD was defined as those causes of mortality and incidence with International Classification of Diseases 10th revision (ICD10) codes I00-I99 (or equivalently the ICD 8th revision (ICD8) codes 390–458 or ICD 9th revision (ICD9) codes 390–459).

Main exposure

Only those studies with individual organ dosimetry that enable estimation of excess relative risk per unit absorbed dose in Gy (ERR/Gy) in relation to heart or brain dose or other closely related tissue doses were used.

Potentially confounding and effect modifying variables considered

We only used those studies in which there was adjustment for any factor other than the standard demographic risk factors (age, sex, year of birth etc.), and in which ERR/Gy were reported both with and without adjustment, or alternatively in which ERR/Gy were reported of the modifying effect on radiation response of these variables. In other words, to assess potential confounding a relative risk model had been fitted of the form \(RR=\exp\lbrack\beta V\rbrack(1+\alpha D)\;\) with adjustment for a potentially modifying variable \(V\) and also without adjustment for that potentially modifying variable \(V\). A modifying variable in a particular study was any variable \(V\) for which had been assessed the interactions with radiation dose, in other words in which a model had been fitted of the form \(RR\;=\;1+\alpha D\;\exp\lbrack\beta V\rbrack\) . In some cases the only reported effect was a p-value (e.g. of significance of modification). All these studies are listed in Tables 1 and 2.

Table 1 Unadjusted or adjusted estimated excess relative risk of cardiovascular diseases in various therapeutically and diagnostically treated groups, exposed at moderate or high radiation doses and high dose rates. All analyses adjust for age
Full size table
Table 2 Unadjusted or adjusted estimated excess relative risks of cardiovascular diseases in the Japanese atomic bomb survivors and in other groups with moderate- or low-dose radiation exposure, with mean dose generally < 0.5 Gy
Full size table

Methods to evaluate the effects of confounding and effect modifying variables

The methods employed to assess the effects of potential confounding variable are comparison of the fitted ERR unadjusted for the potential confounding variable, \(ERR_{unadj}\), and the ERR \(ERR_{adj\;\lbrack V\rbrack}\) adjusted for the confounding variable \(V\). We categorized those estimates in which adjustment for potential confounders resulted in changes of the following magnitudes:

  1. a)

    more than 50% difference, i.e., with ratio of estimates outside the interval [0.667, 1.5] – labelled *

  2. b)

    more than 100% difference, i.e., with ratio of estimates outside the interval [0.5, 2.0] – labelled **

  3. c)

    estimates with different signs, i.e. one positive, the other negative, labelled ***

Likewise any variable whose interaction with radiation dose resulted in the following degree of change in the ERR was labelled as follows:

  1. a)

    more than 50% difference, i.e., with ratio of estimates (with/without modification) outside the interval [0.667, 1.5] and the heterogeneity was statistically significant (p < 0.05) – labelled †

  2. b)

    more than 100% difference, i.e., with ratio of estimates (with/without modification) outside the interval [0.5, 2.0] and the heterogeneity was statistically significant (p < 0.05) – labelled ††

  3. c)

    estimates with different signs, i.e. one (with modification) positive, the other (without modification) negative or vice versa and the heterogeneity was statistically significant (p < 0.05), labelled †††

We specifically highlight the most serious discrepancies of both sorts (***, †††) in the text below. We describe these as potential confounders and potential effect modifiers, respectively. Table 3 reports those studies and results (taken from Tables 1, 2) in which one of these six categories of potentially confounding or modifying effects is observed. Table 4 separately reports effects of variation of latency.

Table 3 Unadjusted or adjusted estimated excess relative risk of cardiovascular diseases in various groups in which there is pronounced effect of adjustment for potential confounder variables, or significant variation by modifying factors. All analyses adjust for age
Full size table
Table 4 Variation with latency of estimated excess relative risks of cardiovascular diseases in occupationally and environmentally exposed groups
Full size table

Results

Of the total of 93 studies from the original systematic review and meta-analysis (detailed in Little et al. [12]), 43 studies satisfied the a priori selection criteria and were retained for this analysis.

Of the 50 studies that were omitted, the high dose radiotherapeutic studies (of the type shown in Table 1) in many cases [55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79] had information on many lifestyle and environmental variables, but only presented one type of analysis (generally fully adjusted); however, there were some high dose studies in which there was little or no information on potential confounders [80, 81]. The Danish study of Lorenzen et al. [82] was omitted as it is largely subsumed by the Nordic study of Darby et al. [19]. There was a similar situation for the lower dose studies (of the type shown in Table 2), which in some cases had rich lifestyle information but only presented a single type of analysis [83,84,85,86,87,88] although in many instances the lower dose studies that were omitted had little or no information, apart from crude markers of socioeconomic status [54, 89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111]. Of the final selected studies, 22 were of groups exposed to substantial doses of medical radiation for therapy or diagnosis (Table 1). The remaining 21 studies were of groups exposed at much lower levels of dose and/or dose rate (Tables 2 and 3). There is substantial overlap in the populations studied in some of these groups. For example three of the studies relate to various CVD endpoints in the LSS [9, 10, 37], and there are various studies of a number of CVD endpoints in the Mayak nuclear workers [40,41,42,43,44,45,46].

Effects of adjustment for potential confounding variables

Very few studies suggested substantial effects of adjustment for lifestyle, environmental, or medical risk factors. In the Rochester study of Adams et al. [34] the unadjusted ERR/Gy for coronary heart disease was 0.08 (95% confidence interval (CI) -0.01 to 0.20), and the ERR/Gy adjusted for sex, diabetes, dyslipidemia, smoking and hypertension was -0.03 (95% CI -0.07 to 0.10) (Tables 1 and 3). In the LSS the unadjusted ERR/Gy for hypertensive heart disease incidence was 0.01 (95% CI -0.08 to 0.10), and the ERR/Gy adjusted for smoking and drinking was -0.01 (95% CI -0.09 to 0.09) [9] (Tables 2 and 3). In the French uranium miners case–control study of Drubay et al. [48] the unadjusted ERR/Gy for all CVD was 0.4 (95% CI -1.8 to 3.0), and adjusted for smoking, body mass index (BMI), diabetes, hypertension, hypercholesterolemia, hypertriglyceridemia, resting heart rate, chronic kidney disease, hyperuricemia, gamma-glutamyl transpeptidase was -0.7 (95% CI -3.2 to 2.9) (Tables 2 and 3). The unadjusted ERR/Gy for hypertension, or CVD excluding cerebrovascular disease (CeVD) and some other endpoints in the Korean medical diagnostic workers were 0.3 (95% CI -6.9 to 9.6) and 1.3 (95% CI -5.0 to 9.0), whereas for the same endpoints adjusted for sex, smoking, alcohol intake, BMI, blood pressure, cholesterol and blood glucose the ERR/Gy were -1.8 (95% CI -8.3 to 6.8) and -0.5 (95% CI -6.3 to 6.7), respectively [51] (Tables 2 and 3).

Modifying effects of radiation

There are only two studies reporting the most serious level of modifying effect, those of the International Nuclear Workers Study (INWORKS) workers by Gillies et al. [39] and the Korean medical diagnostic worker study of Cha et al. [51] (Tables 2 and 3). Gillies et al. [39] reported markedly higher risks for females, with ERR/Gy of 4.22 (90% CI 1.72 to 7.21) for all CVD, 6.17 (90% CI 2.44 to 10.92) for ischemic heart disease (IHD), and 2.67 (90% CI < 0 to 9.79) for CeVD, compared with ERR/Gy for males for the same endpoints of 0.20 (90% CI 0.07 to 0.36), 0.16 (90% CI -0.01 to 0.34), and 0.48 (90% CI 0.10 to 0.91), respectively; these differences were highly significant for all CVD (p = 0.005) and IHD (p = 0.004), but not for CeVD (p > 0.50) (Tables 2 and 3). Gillies et al. [39] also reported modifications by attained age and duration of employment, and although some of these were substantial for certain groups (Tables 2 and 3) none were statistically significant (p > 0.10). The Korean worker study of Cha et al. [51] also reported markedly higher ERR/Gy for females, with CVD ERR/Gy of 42.1 (95% CI 3.0 to 99.2) compared with ERR/Gy for males of -0.7 (95% CI -7.6 to 7.6). This study also reported significant modifications of ERR/Gy for CVD in relation to age, birth year, year started work and years worked (Tables 2 and 3). There were similar indications of heterogeneity for all these variables for hypertension and CVD excluding CeVD and other CVD (ICD10 I70-I83, I85-99). The French Childhood Cancer Study of Mansouri et al. [16] reported marked discrepancies with treatment status by anthracyclines (cardiotoxic anticancer drugs), with an ERR/Gy of 0.44 (95% CI 0.18 to 1.12) for those not treated compared with an ERR/Gy of 0.09 (95% CI 0.02 to 0.22) for those treated with anthracyclines (Table 1). They also reported that “HF [heart failure] risk increased with each category of age, with ERR/Gy values for the MHD [mean heart dose] of 0.06, 0.33, 0.38 and 0.48 in those aged < 15, 15–25, 25–35, and ≥ 35 years, respectively” [16]. The Netherlands-NKI-Rotterdam breast cancer case control study of Jacobse et al. [22] reported borderline significant modifications in risk by age at exposure (ERR/Gy age < 45 = 0.242 (95% CI 0.044 to 0.823), age 45–49 = 0.111 (95% CI 0.012 to 0.401), age 50–70 = 0.025 (95% CI -0.014 to 0.119), heterogeneity p = 0.07) (Table 1). In no other high radiation dose studies (reported in Table 1) were any modifying effects reported [17,18,19, 22, 23]. In the LSS, there are significant modifying effects on CeVD risk of attained age and age at exposure [10, 11], with ERR decreasing in each case with increases in each variable, but no such modifications in ERR by any of these variables when the investigators used only CVD as outcome; Little et al. [11] reported a change in ERR/Gy per year of age at exposure by -0.050 (95% CI -0.099, -0.015) for stroke and by -0.012 (95% CI -0.041, 0.018) for heart disease. There is a borderline significant effect (p = 0.022) of sex on heart disease but not for stroke (p > 0.9) [11].

Although not reported in Table 2, because the analysis was not generally aligned with that of the main analysis (using a lag of 5 years, and wherever possible dose < 4 Gy), there is analysis of the Mayak worker data by sex, attained age and duration of employment group; there was no statistically significant heterogeneity (p > 0.1) in effect suggested by these analyses [41, 42, 44,45,46].

Modifying effects of latency

Table 4 shows the modifying effects of latency. There is very little variation of ERR/Gy with lag, although there is a slight tendency for ERR/Gy to increase with increasing lag period.

Modifying effects in animal data

Table 5 illustrates what little is known about potential modifying factors from radiobiological animal data. There is some evidence of the modifying effects of age at exposure and chemotherapy in certain systems.

Table 5 Effect of modifying variables on absolute risk in radiobiological animal data
Full size table

Discussion

We have assessed effects of confounding by lifestyle, environmental or medical risk factors on radiation-associated CVD, using data assembled for a recent systematic review [12]. We found only limited evidence that adjustment for potential confounding made substantial difference to risk estimates. Only in four studies, in a group treated in childhood for hemangioma [34], in the LSS [9] and in two groups of nuclear workers [48, 51], were the adjusted and unadjusted ERR substantially different (Tables 1, 2 and 3). However, it was hard to assess whether these variables were true confounders of the association between radiation and CVD; there were also substantial uncertainties in all of these studies, so that not much weight can be attached.

We also investigated evidence for modifying effects of these variables on CVD radiation dose response, again using data assembled for the systematic review [12]. There are fewer suggestions of the most serious level of modifying effect, with age at exposure modifications in the same direction reported in two studies [11, 16], although for different disease endpoints. In the study of Mansouri et al. [16] there were substantial modifying effects of anthracycline exposure (Tables 1, 2 and 3). In the LSS and in two groups of nuclear workers there are significant modifying effects of sex [11, 39, 51], although for discrepant endpoints. However, in most of the studies reported here no analysis has been reported of modifying effects of these or any other variable (Tables 1, 2 and 3). There is little variation of ERR with lagging period, although there is a slight tendency for ERR/Gy to increase with increasing lag period. (Table 4).

The radiobiological animal data has rather less information (Table 5). The metrics used are heterogeneous, and in general the internal dose trends (ERR/Gy) used in the epidemiological data given in Tables 1, 2 and 3 are not given. Indeed, in most studies there is only a single irradiated group, and the relative effects of the extra covariate on the radiation-associated relative risk (irradiated vs control) difficult to determine. Given the heterogeneity in endpoints used and in the animal systems employed one should probably not attach much weight to these findings.

Assessment of outcomes is a complication, as mortality outcomes could be less accurate than studies of incidence. In incidence studies, medical and lifestyle factors are more likely to be collected, as reported in our systematic review [12]. Pooling mortality and incidence data could explain part of the heterogeneity of the summarized results observed in the meta-analysis. Indeed, summarized risks were significantly higher for mortality endpoints compared with those of incidence in the meta-analysis [12]. An additional complication in many of the studies presented, particularly of the LSS [9,10,11, 37], groups of nuclear workers (IARC 15-country and INWORKS nuclear worker studies [38, 39], Mayak nuclear workers [40,41,42, 44,45,46]) is the overlap in subjects included, a feature also of the systematic review from which they were drawn [12]. It is very likely that there is inter-study heterogeneity of effect in the present study, reflecting the heterogeneity that was observed in the systematic review from which it is drawn [12]. Some part of the heterogeneity in the previous review is clearly driven by differences in endpoint sensitivity, by age at exposure, by dose and dose rate, and as noted above mortality vs incidence, but even after accounting for the effect of these heterogeneity remained [12]. Such heterogeneity complicates any causal interpretation of the results presented.

Confounding is likely to be specific to each study and the effects of adjustment could rarely be generalized to other studies. Residual confounding could be an issue, if the potential confounding variable is measured with substantial error. Effect modification is likely to be much more easily compared between studies, although the evidence assembled here does not suggest that even for such easily and reliably measured variables as sex and age at exposure there are consistent effects within studies (Tables 1, 2 and 3). These differences in modifying effect between cohort may reflect the play of chance, but it is also possible that there are underlying differences between the cohorts. Medical and lifestyle factors are differently available in the studies considered, with more detailed information among studies of radiotherapeutic exposure, where radiation doses are high to moderate. However, the number of patients included in these medical studies is generally rather small, limiting study of specific endpoints. In contrast, in studies on workers or in general population, generally few potential confounding variables can be collected as the large number of people recruited and the way of collecting the information does not usually allow such information to be obtained. In our systematic review [12], among the lower dose studies with detailed information on lifestyle factors and a large number of included people there were only a few occupational cohorts, principally the Mayak worker cohort [40,41,42, 44,45,46], the Semipalatinsk cohort [53], and, with a rather smaller number of people included, the French nuclear fuel cycle workers [47, 48] and the Korean radiation worker cohorts [51, 52]. For the purposes of maximizing statistical power, specific CVD outcomes are analyzed together, but potential confounders could act differently on the different outcomes and their specific effect could be unseen in a pooled analysis of heterogeneous outcomes.

As summarized in Tables 1, 2 and 3, in general many epidemiological studies now have quite rich lifestyle, environmental and medical information. As highlighted in the Results there are over 30 other studies that clearly have such cofactor information, although in all cases best use has not been made of this in the publications for the purposes of assessing effects of potential confounding factors or effect modifying factors.

A limitation of our analysis is that statistical significance cannot be attached to the difference made by potentially confounding factors, since the reported coefficients would necessarily be highly correlated, and from the published data this correlation is impossible to determine. We therefore judge that this has to remain as we describe it in the Methods, based simply on the size and sign of the coefficients. However, as we outline in the Methods, one of the criteria for risk modifying factors is based on statistical significance. Clearly the particular levels of magnitude we chose to determine the seriousness of potentially confounding variables, likewise the levels of difference made by risk modifying factors are both somewhat arbitrary. Another limitation of the analysis is the degree of overlap in two particular studies, specifically two of the most important and informative ones, the LSS [9, 10, 37] and the Mayak workers [40,41,42,43,44,45,46], although not in any of the other studies listed in Tables 1, 2, 3 and 4. However, given the form of the analysis, bias would not result from this. At most there would be a tendency for findings (of potential confounding or risk modification) to be inflated by these correlated findings. As may be inferred from the results presented in Tables 3 and 4, there is little evidence for this, although the Mayak worker data do show a similar direction of effect made by adjustment to two overlapping mortality endpoints, IHD and all CVD [41].

In many studies in which adjustment is made for certain covariates, these are assayed at a number of time points and this information is then used to adjust for health endpoints after that point. Some of these covariates may also affect competing risks, for example cancer, and it is possible that they may affect both baseline CVD risk as well as radiation-associated excess risk; however, the evidence we have presented (Tables 1, 2 and 3) does not suggest that this is likely. Competing risks may well not be independent of CVD, so that the censoring they introduce will be informative. In this case consideration may have to be given to non-standard ways of analyzing the data. There are a number of statistical methods to assess effects of two or more competing risks [126]. One of the most popular is the so-called subdistribution hazard of Fine and Gray [127].

In summary, because of the multifactorial etiology of CVD, medical and lifestyle factors are clearly crucial variables to take into account in analysis of the dose response of these endpoints. The heterogeneity of the studied populations and of the type of exposure and dosimetry complicates drawing conclusions on the impact of medical and lifestyle factors on the dose response relationship between exposure to radiation and CVD. Nevertheless, we found a large number of studies in which there is information on effect of adjusting for certain lifestyle/environmental/medical variables, although in the larger number of studies previously assessed this information was not available, even if the relevant variables had clearly been assessed. We found limited evidence of potential confounding of radiation effects on CVD (Tables 1, 2 and 3); substantial differences were made by adjustment in four studies, but the uncertainties in all cases were substantial, so that little weight can be attached. There is much less information on potential modifying variables of radiation effect; nevertheless there is some evidence of the effects of age at exposure and sex, although not always in the same endpoints in different studies. However, in most of the studies reported here no analysis was reported of modifying effects of these or any other variable (Tables 1, 2 and 3). There is little evidence of modification resulting from variation in lagging period (Table 4). Efforts should be made to include in future studies as much as possible precise information on these variables and if available specific analysis on their impact on the dose response relationship should be assessed. It is important that analyses of radiation-associated CVD clearly demonstrate the effect of adjustment for the available lifestyle/environmental/medical variables, and also assess the potential modifying effect of these variables on the radiation dose response.