What is the association of hypothyroidism with risks of cardiovascular events and mortality? A meta-analysis of 55 cohort studies involving 1,898,314 participants

This meta-analysis follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) Checklist (Additional file 1). We searched PubMed and Embase from inception to 29 February 2016, using the following text and key words, both as MeSH terms and text words: hypothyroidism, subclinical hypothyroidism, thyroid diseases, thyroid function, thyroid status, coronary heart diseases, ischemic heart diseases, myocardial infarction, coronary atherosclerosis, cardiovascular diseases, death, mortality, heart failure, stroke, atrial fibrillation, arrhythmia, peripheral artery diseases. We searched articles published in any language and screened references from these studies to identify other relevant studies. YN and YJC performed the literature search independently. Differences were resolved by discussion.

To minimize differences, studies were required (1) to measure thyroid function and follow persons prospectively; to assess ischemic heart disease (IHD), mortality or other cardiovascular events; and to provide risk estimates or sufficient data to calculate risk estimates associated with hypothyroidism compared with euthyroidism; (2) to be cohort studies published as original articles; and (3) to be independent. For multiple publications of the same study design, we included the estimates from the most recent or informative reports.

Data were extracted by one reviewer (YN) and independently checked by another three reviewers (YJC, LJL, and HJH); the data included the first author’s name, publication year, location, female proportion, mean age, mean follow-up years, sample size, study design, study population, number of cardiovascular events and mortality, risk estimates, categories and definitions of hypothyroidism, thyroid treatment, and covariates adjusted for in the multivariable analyses. Primary authors were contacted by email if the study did not provide the data needed. We used the Newcastle-Ottawa scale to evaluate the quality of cohort studies. In the present analyses, we regarded a study awarded a total score of ≥6 as a high-quality study [11].

Hypothyroidism was categorized as SCH and OHypo. Their definitions varied across studies. Generally, SCH was defined as elevated thyroid-stimulating hormone (TSH) levels and normal free tetraiodothyronine (FT4) levels. OHypo was defined as elevated TSH but decreased FT4 levels [5]. In this analysis, hypothyroidism refers to the combination of SCH and OHypo. The primary exposure of interest is hypothyroidism (i.e., participants with an increased TSH level regardless of FT4 level).

The primary outcomes were IHD, cardiac mortality, cardiovascular mortality, and all-cause mortality as defined by International Classification of Diseases-10-CM (ICD-10-CM) diagnosis codes. For IHD (I20–I25), we included angina pectoris, myocardial infarction (MI), and chronic ischemic heart disease. For cardiac mortality, we included deaths caused by IHD, progressive heart failure (HF), arrhythmia, or cardiac arrest (I20–I25; I44–I50). For cardiovascular mortality, we included deaths from diseases of the circulatory system (I00–I99). We used both cardiac and cardiovascular mortality for the following two reasons: first, both of them had been discussed in different meta-analyses previously; second, the related data of both outcomes were available from considerable original cohort studies. Therefore, it is necessary to analyze both of them in one meta-analysis to achieve comprehensive comparison to the previous meta-analyses. The secondary outcomes were stroke, HF, MI, atrial fibrillation (AF), and total cardiovascular events.

The available risk estimates that were extracted were mostly hazard ratios (HRs), while those in partial studies were rate ratios (RRs) [12‐15], incidence rate ratios (IRRs) [16, 17], or odds ratios (ORs) [18]. We used the most comprehensively adjusted estimates in the reports. When risk estimates and confidence intervals (CIs) were not provided [19‐21], we calculated RRs and CIs from available data by using the Woolf method in Stata version 12.1 software (used for all analyses) [9, 22]. Since the absolute risk of IHD, cardiac mortality, cardiovascular mortality, all-cause mortality, stroke, HF, MI, AF, or total cardiovascular events is low, the four measures of association are expected to yield similar estimates of RR [23, 24]. Consequently, we estimated the pooled RRs and their CIs by using metan commands and random effects models based on the variance model developed by DerSimonian and Laird [9, 25]. We calculated the I ² statistic to assess heterogeneity among studies. The formula is {(Q-df)/Q × 100%}, where Q is the chi-squared statistic and df is its degrees of freedom [26]. We use the following criteria: I ² < 50%, low heterogeneity; 50–75%, moderate heterogeneity; >75%, high heterogeneity [26, 27].

Several predefined subgroup analyses were intended to explore potential sources of heterogeneity in primary outcomes (metan command); where enough studies (n ≥ 10) [28] were available, meta-regressions were conducted (metareg command). Specifically, we intended to conduct subgroup analyses by hypothyroid states and by study population to clarify whether adults with hypothyroidism and preexisting cardiovascular disease were at particularly high cardiovascular risk. Given previous findings of the possible protective effect of SCH in oldest old people [29], we intended to conduct subgroup analyses by mean age of populations. Predefined sensitivity analyses were performed to test the robustness of the results (metan command). When I ² ≥ 50% (moderate or high heterogeneity), we used the hetred command in Stata to evaluate the change in between-study heterogeneity as one or more outlier studies were excluded from the calculations, a method which was developed by Patsopoulos [30]. If available, we intended to perform additional subgroup analyses of main outcomes by TSH levels (TSH ≥ 10.0 and TSH < 10.0 mIU/L) in SCH.

If there were at least 10 studies included in each outcome of the meta-analysis, publication bias or small study bias was assessed with a funnel plot by using Egger’s regression asymmetry test (metabias6 command) [31, 32]. Statistical tests were two-sided and used a significance level of P < 0.05.

Of 3889 reports identified, we excluded 3775 studies that were unrelated to our theme and 59 studies after detailed evaluation. Fifty-five cohort studies met the inclusion criteria. Figure 1 shows details of the study selection.

Additional file 2: Table S1 shows the main characteristics of the 55 included studies. Thirty-two studies were population-based studies. Twenty-three studies were convenience samples (i.e., particular patient groups), of which 14 studies comprised cardiac patients. Nineteen studies analyzed with all hypothyroidism, 34 with SCH, and 8 with OHypo. Cohort studies included in each outcome are mainly prospective and mostly from Europe or North America. Thirty-seven of the 55 studies were published in the last 5 years (2012–2016). Of the 55 cohort studies, 49 studies were awarded a total score of ≥6 (Additional file 2: Table S2).

Additional file 2: Table S3 shows detailed information on adjustments in risk estimates of each cohort study. Most risk estimates were adjusted for age (47 studies) and sex (40 studies), which are well known as potential confounders of both thyroid disease status and risk of cardiovascular events. Forty studies (72.7%) reported an adjusted estimate for at least one of the main cardiovascular risk factors (body mass index (BMI), smoking, cholesterol, hypertension, and diabetes). Seven studies’ risk estimates were adjusted for replacement therapy (RT).

Thirteen studies with 615,596 individuals and 16,862 events analyzed with IHD. Twelve studies’ risk estimates were adjusted for age, ten adjusted for sex, and more than half adjusted for smoking, BMI, hypertension, or diabetes, but only three adjusted for RT.

Overall, patients with hypothyroidism experienced a higher risk of IHD compared to euthyroidism (RR: 1.13; 95% CI: 1.01–1.26), with evidence of low heterogeneity (I ²: 40.2%; P = 0.066; Fig. 2). Table 1 shows the results of sensitivity analyses. Risk estimates changed little after analyzing with fixed effects models, or after inclusion of studies with ≥10,000 participants, studies without RT at baseline, or high-quality studies. No publication bias was found (Egger test: P = 0.65; Additional file 2: Figure S1).

A priori subgroup analyses were conducted across key study characteristics (Table 2); higher risk of IHD associated with hypothyroidism was consistently observed in some of these analyses. The estimates did not differ significantly between population-based and convenience samples studies. There was no significantly higher risk of IHD events in cardiac patients with hypothyroidism. However, this should be considered seriously for only three studies included. The estimates also did not differ significantly across other subgroups. In additional subgroup analyses by TSH levels in patients with SCH compared with euthyroidism, we observed a significantly higher risk of IHD in SCH with TSH level ≥10.0 mIU/L (RR: 1.32; 95% CI: 1.00–1.74; Additional file 2: Table S4). Univariate meta-regressions showed that these a priori factors did not affect the overall effects (Table 3).

Of seven studies analyzed with cardiac mortality, there were 1101 events among 34,922 participants. All studies adjusted for age and sex, six adjusted for smoking, less than half adjusted for other main cardiovascular risk factors, and only one study adjusted for RT.

Evidence synthesis for cardiac mortality showed a statistically strong association with hypothyroidism (RR: 1.96; 95% CI: 1.38–2.80; Fig. 3), with moderate heterogeneity (I ²: 66.2%; P = 0.007). Sensitivity analysis by hetred command in Stata found that two outlier studies might be the source of heterogeneity [33, 34]. When the two outlier studies were removed, there was evidence of low heterogeneity in the five remaining studies (I ²: 37.9%; P = 0.17; Table 1). The heterogeneity may have been due to the smaller sample size of the outlier studies than that of the other five studies (Fig. 3).

Subgroup analyses did not find the source of heterogeneity. However, we found that the risk of cardiac mortality was also significantly higher in SCH (Table 2). Besides, we observed that cardiac patients with hypothyroidism might have a higher risk of cardiac mortality than cardiac patients with euthyroidism (RR: 2.22; 95% CI: 1.28–3.83). Sensitivity analyses with large cohorts or high-quality studies obtained similar results. Considering only seven studies in this section, we did not conduct meta-regression and funnel plots.

Seventeen studies were included for cardiovascular mortality, involving 158,017 participants and 5273 events. Fourteen studies adjusted for age, thirteen adjusted for sex, more than half adjusted for cardiovascular risk factor, and only four adjusted for RT.

The summary RR of cardiovascular mortality associated with hypothyroidism was 1.11 (95% CI: 0.97–1.28; Fig. 4), with low heterogeneity (I ²: 11.1%; P = 0.324). In sensitivity analyses, the risk estimates did not change after analyzing with studies removing RT at baseline or with high-quality studies. No publication bias was found (Egger test: P = 0.92). In meta-regression and subgroup analyses, the pooled RRs of each group or subgroup were almost similar to the total.

All-cause mortality occurred in at least 112,156 patients among 1,036,389 participants from 40 studies. Thirty-five studies adjusted for age, thirty-one adjusted for sex, most adjusted for cardiovascular risk factors, and only five adjusted for RT.

The pooled RR value of all-cause mortality was 1.25 (95% CI: 1.13–1.39; Fig. 5), with high heterogeneity (I ²: 86.9%; P < 0.001). In subgroup and meta-regression analyses, we observed evidence for heterogeneity by study population (P = 0.009). The RR for all-cause mortality in convenience samples was significantly higher than that in population-based studies; also, the risk of all-cause mortality in cardiac patients with hypothyroidism was significantly higher than that in cardiac patients with euthyroidism (RR: 1.51; 95% CI: 1.26–1.81). We observed that SCH was also associated with a higher risk of all-cause mortality (RR: 1.32; 95% CI: 1.13–1.54).

In sensitivity analyses, the pooled RR value changed little after inclusion of large cohorts, studies without RT at baseline, or high-quality studies. Sensitivity analyses by hetred command in Stata found eight outlier studies [16, 18, 19, 29, 33, 35‐37]. When they were removed, the pooled RRs of the remaining studies were still significant (RR: 1.30; 95% CI: 1.22–1.39; I ² = 26.4%). No publication bias was found (Egger test: P = 0.44).

For stroke, nine studies were included, reporting 27,196 events among 761,993 participants. All studies adjusted for age, eight adjusted for sex, most adjusted for cardiovascular risk factors, and only one adjusted for RT. The overall RR of stroke was 1.09 (95% CI: 0.96–1.24; Fig. 6) with moderate heterogeneity (I ²: 52.3%; P = 0.03). After excluding one outlier study [16] with population >10,000, RR changed slightly with no heterogeneity (I ²: 0.0%; P = 0.43).

Of eight studies analyzed with HF, there were 23,033 events among 599,876 participants. Seven studies adjusted for age, all adjusted for sex, most adjusted for cardiovascular risk factors, and only two adjusted for RT. Evidence synthesis for HF did not show a significant association with hypothyroidism (RR: 1.13; 95% CI: 0.98–1.30; Fig. 6), with moderate heterogeneity across studies (I ²: 54.8%; P = 0.03). After excluding one outlier study [5], the heterogeneity decreased with the pooled RR still insignificant. We performed an additional subgroup analysis of risk of HF by mean age (Additional file 2: Table S5). The risk estimates did not significantly change in each subgroup.

Seven studies involving 588,182 participants were analyzed reporting a total of 13,263 MI. Six studies adjusted for age, four adjusted for sex, most adjusted for cardiovascular risk factors, and none adjusted for RT. Hypothyroidism was significantly associated with higher risk of MI (RR: 1.15; 95% CI: 1.05–1.25; Fig. 7), with no heterogeneity (I ²: 0.0%; P = 0.43). The same results were obtained after using fixed-effect models, inclusion of large cohorts, studies without RT at baseline, or high-quality studies.

For AF, four studies were included, reporting 18,259 events among 616,608 individuals. All studies adjusted for age and sex, most adjusted for cardiovascular risk factors, and only one adjusted for RT. The pooled RR value of AF was 1.02 (95% CI: 0.71–1.46; Fig. 7), with moderate heterogeneity (I ²: 70.6%; P = 0.02). After excluding one outlier study [38] of HIV-infected patients with a large population size, no heterogeneity was found (I ²: 0.0%; P = 0.85), but the RR was still not statistically significant.

Eight studies involving 579,969 participants were analyzed reporting a total of 45,545 total cardiovascular events. Six studies adjusted for age, six adjusted for sex, most adjusted for cardiovascular risk factors, and none adjusted for RT. The pooled RR for total cardiovascular events was 1.16 (95% CI: 0.97–1.38; Fig. 7), with moderate heterogeneity (I ²: 73.4%; P < 0.001). Similar results were observed in the sensitivity analyses.

The strengths of this meta-analysis include strict inclusion criteria, many more cohort studies included than ever before, and predefined sensitivity and subgroup analyses. There is no publication bias, which favors stability of the findings. Moreover, we do not restrict hypothyroid states and study population, thus realizing a full-scale assessment of the effects of hypothyroidism.

There are some limitations. Firstly, heterogeneity is observed in some outcomes. However, we have found all sources of heterogeneity, and sensitivity analyses confirm the stability of our results. Secondly, the absence of IPD might influence the accuracy of the results. Thirdly, thyroid function testing was mostly performed at baseline, which would weaken the accuracy of the results regarding the possible change of thyroid function during follow-up. Another limitation is the different TSH values used across studies to define thyroid disease.