Study identification

A two-step search process was employed to identify relevant studies. First, relevant SRs were identified from MEDLINE (via PubMed) and Scopus databases through 7 November 2023. We included SRs with or without meta-analyses that evaluated the effect of e-cigarettes on the risk of CVD. The original studies included in these eligible SRs, along with those identified from their reference lists, were further assessed for eligibility.

Second, an updated search was conducted in the same databases, covering studies published since the last search date of the most recent SR^14,15, i.e. 1 January 2020 through 7 February 2024. Details of the search terms and strategies used for each database are provided in Supplementary file Tables 1 and 2.

Study selection

Titles and abstracts of identified studies (694 records for step one and 561 records for step two) were screened by two independent reviewers (ATa and TA). The full texts of potentially eligible records (80 studies for step one and 14 studies for step two) were reviewed. Studies that met the pre-defined eligibility criteria were included. Disagreement between reviewers was resolved by third party consensus.

Eligibility criteria

Observational studies (i.e. cross-sectional, case-control, or cohort studies) and randomized controlled trials (RCTs) were eligible for inclusion if they met the following criteria: 1) conducted in the general population; 2) compared the effect of e-cigarette use with any comparator including cigarette, dual use of e-cigarettes and cigarettes, or non-use; and 3) assessed CVD outcomes, including MI, stroke, or composite CVD endpoints. There were no language restrictions; however, non-English studies were excluded if their key information could not be adequately translated using available translation tools. For evidence synthesis, the included studies were categorized based on their reported outcomes.

Exposures and outcomes of interest

The primary exposures of interest were e-cigarette, cigarette, and dual users, which were categorized into current and former users based on the definitions provided in each original study. The comparator was any never user of either product. Ultimately, exposures were classified into 7 groups including never user, former e-cigarette, former cigarette, former dual use, e-cigarette, cigarette, and dual users. The outcomes of interest were CVDs, including MI, stroke, and composite CVD endpoints defined according to the original studies.

Data extraction

Two independent reviewers extracted the following data: study characteristics (i.e. first author’s names, publication year, study design, database used, and the number of participants), population characteristics (i.e. mean age, sex, mean body mass index (BMI), and comorbidity status), and type of exposures (i.e. e-cigarette, cigarette, dual use, former e-cigarette, former cigarette, former dual use, and never use).

Data for pooling (i.e. number of participants between exposure groups and outcomes) were extracted as a contingency table if available. Otherwise, a crude or adjusted summary statistic and standard error (SE) or upper and lower bound of 95% confidence interval (CI) were extracted. These included odds ratios (ORs) and hazard ratios (HRs). Covariates adjusted in the model were extracted. For HRs, they were converted to risk ratios (RRs) and subsequently ORs using the following equations¹⁸:

and

where r₀ is the incident rate of the outcome in the reference group.

Risk of bias assessment

Risk of bias for the studies included was assessed by two independent reviewers using the Joanna Briggs Institute (JBI) appraisal tool¹⁹ for cross-sectional studies, which consists of eight questions. A study was considered to have a high risk of bias if it received at least one negative answer, while a study was classified as low risk of bias only if all questions received positive answers. The Newcastle-Ottawa Scale (NOS)²⁰ was used to assess cohort studies. The NOS tool has three domains including cohort selection (4 items), comparability (1 item), and outcome assessment (3 items). For each domain, the number of rated stars indicates quality of that domain. A maximum of one star per item was rated except for the comparability domain which was allowed up to two stars. Studies were considered poor quality if they received ≤1 star for cohort selection or outcome assessment domains or zero stars for the comparability domain. Again, disagreement between both reviewers was resolved by third party consensus.

Statistical analysis

To avoid potential confounding effects, only adjusted effect sizes were considered in the NMAs. Data pooling was performed separately for each outcome (i.e. composite CVD, MI, and stroke). A multivariate random-effects model with a consistency assumption was used to estimate the adjusted relative effects (i.e. AORs) of all pairwise comparisons among exposures. Surface Under the Cumulative Ranking (SUCRA) curves and P-scores ranked exposure effects for each outcome. Transitivity was assessed by comparing study characteristics between comparisons/studies. Global heterogeneity was assessed using τ² and global chi-squared test. A design-by-treatment interaction model and node-splitting (back-calculation method) were used to investigate violations of the consistency assumption. Publication bias was assessed using a comparison-adjusted funnel plot and the Egger’s test. Asymmetry of the comparison-adjusted funnel plot was considered indicative of publication bias. P-values less than 0.05 were considered significant. All statistical analyses were performed using R version 4.4.1 (netmeta package)²¹. Confidence in evidence synthesis from NMA was assessed using the Confidence In Network Meta-Analysis (CINeMA) framework²².

Study selection

The initial search (step one) identified 986 records; 292 and 614 records were excluded by deduplication and title and abstract screening, respectively. Among 80 records entering full-text screening, only 13 SRs were eligible. Eight individual studies were identified from these SRs (Figure 1A). In the second search phase (step two), 707 records were retrieved. After removing duplicates (146 records), 561 records remained for title and abstract screening. Of these, 547 records were excluded and 14 studies plus eight individual studies identified from the previous SRs, were assessed by full-text screening. Duplications and studies that used overlapping data sources were further excluded at this point. Therefore, a total of 11 primary studies^23-33, comprising 6022304 participants, were included in this review (Figure 1B).

Figure. 1

PRISMA diagram of study selection: A) first step; B) second step

Characteristics of included studies

Data were collected from six different databases, primarily based in the United States^{23-26, 28-33}, with the exception of one study²⁷, which used data from the Korean National Health Insurance Service (Table 1). All studies focused on adult populations, with male representation ranging between 40% and 70%. Hypertension prevalence ranged from 25.2% to 39.8%; diabetes prevalence was lower, ranging from 2.1% to 20.4%. The history of e-cigarette, cigarette, dual, and never use was collected by self-report questionnaire for all studies. The questions used for defining each type of exposure for each study are presented in Supplementary file Table 3.

Among 11 included studies, 8 ^{23-25,27,29-32}, 4 ^23,28,30,33 and 5 studies^{23,26,30,32,33} reported outcomes as composite CVD, MI, and stroke, respectively. Definitions and methods of outcome verification for each study are described in Supplementary file Table 4. Most of the included studies verified these outcomes by self-reported questionnaire. However, provision of adjusted ORs for meta-analysis were only provided by 6 studies for composite CVD^{25,27,29,30-32}, 2 studies for MI^30,33, and 4 studies for stroke^26,30,32,33. Details of covariate adjustment used in each study are provided in Supplementary file Table 5. Data for pooling are provided in Supplementary file Tables 10–13.

Risk of bias assessment

Only one study was classified as having low risk of bias²⁷, while nearly all studies^23-26,28-33 (10 out of 11) were judged to have a high risk of bias or poor quality, primarily due to the domain of outcome measurement (Supplementary file Tables 6 and 7). This domain was rated as high risk of bias because the CVD outcomes were assessed through self-reported questionnaires, which may not represent a confirmed CVD diagnosis, given the limitations associated with self-reported data.

Composite CVD endpoints

Adjusted ORs for seven exposure categories were evaluated using data from six studies^{25,27,29,30-32} (Figure 2). Common covariates in the adjusted model included age, sex, ethnicity, body mass index, education, income level, comorbidity status, and substance abuse.

Figure. 2

Network configuration of the composite cardiovascular outcome considering adjusted effect of exposures. The size of the nodes is proportional to the number of included studies, while the thickness of the edges corresponds to the inverse standard errors of effect sizes comparing two exposures. The numbers within the edges indicate the number of studies for each comparison

Relative effects from seven exposures were estimated. Compared to non-use, all exposures with the exception of former e-cigarette use, were significantly associated with composite CVD endpoints. The pooled AORs (95% CI) were: 1.31 (1.05–1.62) for e-cigarette use, 1.67 (1.37–2.03) for dual use, 1.46 (1.03–2.08) for former dual use, 1.57 (1.30–1.88) for cigarette use, and 1.29 (1.05–1.59) for former cigarette use (Table 2 and Figure 3). Additionally, current dual users had a significantly higher risk of CVD compared to current e-cigarette users, former cigarette users, and former e-cigarette users, with adjusted ORs of 1.28 (1.03–1.59), 1.30 (1.04–1.61), and 1.89 (1.23–2.90), respectively, although this risk did not significantly differ for comparisons with current cigarette and former dual users. Furthermore, current cigarette use was associated with a 1.20 (0.97–1.48) times higher risk of CVD relative to current e-cigarette use, although this failed to reach the significance threshold. Among the exposures of interest, dual use was ranked as having the highest risk of CVD (SUCRA=0.91, P-score=0.91), followed by cigarette (SUCRA=0.80, P-score=0.80), former dual use (SUCRA= 0.68, P-score=0.67), e-cigarette (SUCRA=0.47, P-score=0.47) , and former cigarette smoking (SUCRA=0.45, P-score=0.45).

Figure. 3

Pooled adjusted odds ratios (ORs) and 95% confidence intervals (CIs) from random-effect network meta-analysis of composite cardiovascular outcomes comparing exposures and never use. Square size around point estimates is proportional to the precision

Myocardial infarction

For pooled adjusted ORs (Supplementary file Figure 1), e-cigarette use was not significantly associated with risk of MI (AOR=0.86; 95% CI: 0.44–1.66), in contrast to dual use, cigarette use, and former cigarette use represented by significantly increased odds of MI when compared to never use with AORs (95% CI) of 3.30 (2.15–5.08), 2.43 (1.69–3.49), and 1.59 (1.03–2.45), respectively. In addition, the odds of MI were significantly higher for dual users than those of e-cigarette, and former cigarette users with pooled AORs of 3.85 (1.91–7.73), and 2.08 (1.29–3.35), respectively. Cigarette users also had significantly higher odds of MI than e-cigarette users (Table 2). Although MI risk was greater for dual users compared to cigarette users with an AOR of 1.36 (0.87–2.13), this was not significant. Regarding SUCRA and P-scores, dual use ranked highest for MI risk (SUCRA=0.98, P-score=0.98), followed by cigarette use (SUCRA=0.76, P-score=0.76), and former cigarette use (SUCRRA=0.50, P-score=0.49).

Stroke

Four studies^26,30,32,33 were included for pooled adjusted ORs (Supplementary file Figure 2). Cigarette and dual use were significantly associated with stroke, when compared to never use with pooled AORs (95% CI) of 1.84 (1.29–2.63) and 2.11 (1.41–3.15), respectively. However, the odds of stroke were not significantly higher for cigarette and dual users than those of e-cigarette users (Table 2). In addition, the odds of stroke in e-cigarette users did not differ significantly compared to non-users, with an AOR (95% CI) of 1.44 (0.92–2.24). According to SUCRA and P-score, dual use ranked highest for stroke risk (SUCRA=0.92 and P-score=0.93), followed by cigarette use (SUCRA=0.77 and P-score=0.76), e-cigarette use (SUCRA=0.45 and P-score=0.45), and former cigarette smoking (SUCRA=0.28 and P-score=0.29).

Inconsistency, transitivity, and publication bias

Study characteristics across all those included were comparable (Table 1), suggesting the transitivity assumption was unlikely to be violated. Global τ² and I² were 0.037 and 94.6% for composite CVD, 0.052 and 45.6% for MI, and 0.089 and 64.6% for stroke, respectively. These values indicate high heterogeneity for the composite CVD outcome and moderate heterogeneity for MI and stroke outcomes, corresponding with Q statistics (P-value) of 260.45 (<0.001), 5.52 (0.138), and 22.57 (0.004), respectively. No evidence of inconsistency was identified by the design-by-treatment interaction model for adjusted effect pooling of the composite CVD, MI, and stroke outcomes with P-values of 0.575, 0.138, and 0.286, respectively. In addition, node-splitting using a back-calculation method did not indicate local inconsistency for almost all the comparisons (Supplementary file Table 8). No asymmetrical pattern was observed in the comparison-adjusted funnel plots for all outcomes, suggesting no publication bias (Supplementary file Figures 3–5). The Egger’s test P-values were 0.512, 0.123 and 0.876 for the composite CVD, MI, and stroke outcomes, respectively.

Confidence in network meta-analysis

The confidence of evidence synthesis from the NMA was rated as low to very low for most comparisons across all outcomes (Supplementary file Table 9). The primary reason for downgrading was within-study bias.

Limitations

There were some other limitations that should be acknowledged. First, most of the included studies relied on participants’ self-reported information regarding tobacco exposure and clinical outcomes that might be subject to misclassification and inaccurate reporting of both exposure and outcomes. These issues might introduce bias to the study findings. Second, the covariates considered in the multivariate analysis varied across individual studies, meaning that the adjusted OR may differ between studies. Moreover, this study is not an individual patient meta-analysis, where effect pooling can be performed using a statistical model that incorporates a uniform set of covariates. Third, this study did not assess the effects of different types of tobacco products due to the limited number of included studies and the lack of data for direct comparisons. Lastly, because the data were primarily derived from cross-sectional studies, the analysis can only identify associations and cannot infer the causal relationships between e-cigarettes and CVD. In addition, it is also possible that patients switched from cigarettes to e-cigarette use following a CVD diagnosis.