For the meta-analysis, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) checklist was used (13). The trail was registered with PROSPERO identifier CRD42022330612. We performed a comprehensive search in PubMed and Embase (from inception to June 1, 2022) for all full-text articles using the following text and keywords in combination: corticosteroid, fluticasone, budesonide, mometasone, prednisone, prednisolone, methylprednisolone, hydrocortisone, betamethasone, glucocorticoid, triamcinolone, flunisolide, beclomethasone, ciclesonide, flunisolide, inhaled corticosteroids (ICS), cardiovascular disease, cardiovascular events, stroke, cerebrovascular accident, cerebrovascular disease, coronary heart disease, myocardial infarction, coronary artery disease, ischemic heart disease, ischaemic heart disease, acute coronary syndrome, and heart failure. The search was limited to humans, but no restrictions were applied based on language, gender, or location. References from included studies and correlated reviews were manually searched to identify additional relevant studies. Duplicate data were removed for studies reporting information from the same cohort for the same outcome. When duplicate studies revealed data from different outcomes, they were included in the pooled analysis.

To minimize differences between studies, the inclusion criteria were as follows: (1) studies presented the use of GCs (users and non-users), including cohort, case-control studies, and randomized controlled trials; (2) studies reported outcomes of CVDs with the minimum information necessary to estimate the relative risk (RR). Moreover, we excluded review articles, letters to editors, and case reports. The decision for inclusion of each study was made independently by two authors. Conflicts between reviewers were resolved by consensus.

Two authors independently extracted relevant data such as the year of publication, geographical location, source of study, study design, baseline patient characteristics, the definition of exposure, the outcome of CVDs, and the cumulative and daily dose of GCs use. We assessed the quality of observational studies using the Newcastlee-Ottawa scale (14). We assessed the quality of randomized controlled trials using the modified Jadad score (15), with a higher score indicating higher quality of the study.

The primary outcome of this study was MACE, including CHD, stroke, HF, and cardiovascular death. Additionally, we examined CHD, HF, and stroke all-cause death as secondary outcomes.

The primary measure of the association between GCs and CVDs in each study was relative risks (RR). Because the incidence of CVDs is rare, hazard ratio (HR) and odds ratio (OR) were treated as equivalent estimates of RR (16–18). We extracted the adjusted estimates when both unadjusted and adjusted estimates were provided.

Random effects were used to calculate pooled RRs (95% CI) to account for heterogeneity among studies, and the results are treated as conservative estimates because CIs are wider. Cochran’s Q and I2 statistics were used to quantify heterogeneity across estimates, with an I2 greater than 50% indicating significant heterogeneity (19). Meta-regression and stratified analyses were used to assess the potential sources of between-study heterogeneity. Studies were stratified based on study design (case-control, cohort, or randomized controlled trial), events (<1,000 or >=1,000), type of GCs (inhaled or not inhaled), location (Asia, Europe, or North America), age (<65 or >=65 years), male proportion (<50% or >=50%), time periods of GCs use (ever or current), and whether risk profiles were adjusted. Moreover, the treatment of GCs changed in 2010. The European League Against Rheumatism (EULAR) recommended adding glucocorticoids to DMARD monotherapy for rheumatoid arthritis as the initial treatment (20). We therefore selected 2010 as the cut-off point for subgroup analysis of the year of publication. Funnel plot, Begg’s test, and Egger’s test were used to evaluate publication bias (21).

In the dose-response analysis, the generalized least squares (GLST) regression model and restricted cubic splines were used to assess pooled dose-response relation between GCs and MACE. Both linear models and nonlinear models were fitted, and the results presented 95% CIs (22, 23). While exposure was reported as categorical data with a range, the mean or median exposure was extracted for each category, and the midpoint was used. Similarly, when the lowest or highest categories were open-ended, the width of the category was assumed to be the same as the adjacent category when estimating the midpoint.

We calculated the absolute difference in risk of GCs treatment as [(RR −1)* I₀]. RR indicates pooled RRs and I₀ was the cumulative incidence of events among patients without GCs (24).

All statistical analyses were performed using STATA 15.0 (STATA Corporation, College Station, TX). A two-sided p-value <0.05 was considered statistically significant.

Twenty-three studies (17 cohort studies, five case-control studies, and one randomized controlled trial) with 883,877 cases reported the outcome of MACE. Overall, GCs use was significantly associated with increased MACE risk compared with no use (RR = 1.27, 95% CI: 1.15–1.40, Figure 1). Heterogeneity analyses suggested considerable heterogeneity across studies (Cochran’s Q = 190.64, I² = 88.5%, P < 0.001). Therefore, subgroup analysis and meta-regression were performed in different subgroups across various main study characteristics and clinical factors to identify potential sources of heterogeneity (Table 1). The difference in the type of study, number of events, location, baseline disease, age, the time periods of GCs use (current or ever), and whether risk profiles were adjusted or publication year did not contribute to the heterogeneity. In contrast, the type of GCs and male proportion were associated with heterogeneity. Notably, increased risk of MACE associated with GCs was not observed in some of the subgroup analyses, including a subgroup of study type (case-control study, randomized controlled trial), the number of events (<1,000), location (Asia), baseline disease (chronic obstructive pulmonary disease/asthma, etc.), Age (>=65 years), Male proportion (>=50%), time periods of GCs use (ever), adjusted for risk profiles (no), publication year (<2010). In the subgroup analysis, we found that current GCs users had a significantly higher risk of developing MACE than non-GC users (RR = 1.37, 95% CI: 1.27–1.49), while the ever GCs users did not. Furthermore, the risk of MACE was lower in patients who used inhaled GCs (RR = 0.82, 95% CI: 0.69–0.97). Also, from nine cohort studies that reported on person-years in GCs users and non-GCs users, we calculated an absolute MACE risk increase of 13.94 (95% CI: 10.29–17.58) cases per 1,000 person-years in GCs users. Moreover, results from Funnel plots (Supplementary Figure S2A), Egger’s test (P = 0.686), and Begg test (P = 0.751) revealed no evidence of publication bias.

Furthermore, we performed an analysis to assess the correlation between cumulative and the daily dose of GCs use and MACE. The types of GCs included in the dose-response analysis were not inhaled (most were oral), and the dosage was converted into a prednisolone-equivalent dose (Supplementary Table S3). As depicted in Figure 2A, both nonlinear dose-response and linear dose-response analyses revealed that the risk of MACE increased with increasing GCs cumulative dose. In addition, in the linear dose-response analysis, we concluded that the MACE risk increased by 10% (95% CI: 6%–15%) for each additional gram of GCs cumulative dose. A similar trend was observed in the analysis of the daily dose-response relation between GCs and MACE outcomes (Figure 2B). Likewise, the increment per 10 micrograms daily risk was estimated to be 63% (95% CI: 46%–83%).

For the risk of CHD, 20 studies (including 622,097 patients) contained information on GCs use, and GCs were associated with an increased risk of CHD (RR = 1.25, 95% CI: 1.11–1.41, Figure 3). Substantial heterogeneity was detected (Cochran’s Q = 111.56, I² = 83.0%, P < 0.001), and the type of GCs, male proportion, and publication year showed potential sources of between-study heterogeneity (Table 1). Additionally, we calculate an absolute risk of CHD increase with 4.16 (95% CI: 2.58–5.74) cases per 1,000 person-years among GCs users. Neither the funnel plot (Supplementary Figure S2B) nor Egger (P = 0.586) and Begg (P = 0.922) tests showed publication bias.

Seven studies about the relationship between GCs and HF contain data from 14,811,957 patients. A significantly increased risk of HF was observed, for GCs use (RR = 1.92, 95% CI: 1.51–2.45, Figure 3), with substantial heterogeneity (Cochran’s Q = 60.99, I²= 90.2%, P < 0.001). In studies included, the Yao’s study contains over 14 million patients, which represents over 90% of the subjects studied. Thus a separate analysis excluding this study was performed, and we also found that GCs were associated with an increased risk of HF (RR = 1.943, 95% CI: 1.50–2.52). In the subgroup analysis, we found the differences in the type of study, age, and time periods of GCs use would contribute to heterogeneity (Table 1). Further, the absolute risk of HF increased by 3.14 (95% CI: 3.38–3.51) cases per 1,000 person-years in GCs users. No publication bias was indicated (Funnel plot in Supplementary Figure S2C, Egger, P = 0.773, Begg, P = 0.548).

Figure 3 shows the results from the random effects model combining the RR for stroke, and no significant correlation was observed with GCs (RR = 1.10, 95% CI: 0.95–1.28). The heterogeneity analyses suggested moderate heterogeneity across these studies (Cochran’s Q = 59.00, I² = 78.0%, P < 0.001). Differences in study type, events, GCs type, age, male proportion, time periods of GCs use, whether adjusted for risk profiles, and publication year were not potential sources of heterogeneity (Table 1). Funnel plot (Supplementary Figure S2D), Egger (P = 0.671), and Begg (P = 0.827) suggested an absence of publication bias.

All-cause death was reported in 15 studies involving 661,256 participants. Overall, GC use was not associated with all-cause death (RR = 0.98, 95% CI: 0.76–1.25, Figure 4), with significant between-study heterogeneity (Cochran’s Q = 406.70, I² = 96.6%, P < 0.001). We also conducted a subgroup analysis and meta-regression in different subgroups (Table 1). GCs type and population age could be sources of heterogeneity. Moreover, we found that inhaled GCs were associated with decreased risk of all-cause death (RR = 0.80, 95% CI: 0.77–0.82). Also, previous GCs use and current GCs use were not associated with all-cause death. Besides, GCs use decreased all-cause death risk in some subgroup analyses, including studies of case-control, North American, chronic obstructive pulmonary disease/asthma, and age >=65 years. The funnel plot (Supplementary Figure S2F) showed no evidence of publication bias with Egger, P = 0.452, Begg, P = 0.276.