Genetically Predicted Insomnia in Relation to 14 Cardiovascular Conditions and 17 Cardiometabolic Risk Factors: A Mendelian Randomization Study.

Background This Mendelian randomization study aims to investigate causal associations between genetically predicted insomnia and 14 cardiovascular diseases (CVDs) as well as the potential mediator role of 17 cardiometabolic risk factors. Methods and Results Using genetic association estimates from large genome-wide association studies and UK Biobank, we performed a 2-sample Mendelian randomization analysis to estimate the associations of insomnia with 14 CVD conditions in the primary analysis. Then mediation analysis was conducted to explore the potential mediator role of 17 cardiometabolic risk factors using a network Mendelian randomization design. After correcting for multiple testing, genetically predicted insomnia was consistent significantly positively associated with 9 of 14 CVDs, those odds ratios ranged from 1.13 (95% CI, 1.08-1.18) for atrial fibrillation to 1.24 (95% CI, 1.16-1.32) for heart failure. Moreover, genetically predicted insomnia was consistently associated with higher body mass index, triglycerides, and lower high-density lipoprotein cholesterol, each of which may act as a mediator in the causal pathway from insomnia to several CVD outcomes. Additionally, we found very little evidence to support a causal link between insomnia with abdominal aortic aneurysm, thoracic aortic aneurysm, total cholesterol, low-density lipoprotein cholesterol, glycemic traits, renal function, and heart rate increase during exercise. Finally, we found no evidence of causal associations of genetically predicted body mass index, high-density lipoprotein cholesterol, or triglycerides on insomnia. Conclusions This study provides evidence that insomnia is associated with 9 of 14 CVD outcomes, some of which may be partially mediated by 1 or more of higher body mass index, triglycerides, and lower high-density lipoprotein cholesterol.

ischemic stroke

inverse variance weighted

Mendelian randomization

waist circumference

waist‐hip ratio

Clinical Perspective

What Is New?

Insomnia is the second most prevalent mental disorder.

Except for some major cardiovascular diseases (CVDs), the causal relationships between insomnia and the development of other CVDs such as arterial hypertension are still unclear, and the potential pathways involved in the association from insomnia to CVDs have not been studied.

This study performed Mendelian randomization analysis to investigate the causal associations between genetically predicted insomnia and 14 CVDs as well as the potential mediator role of 17 cardiometabolic risk factors.

What Are the Clinical Implications?

Insomnia was found to increase the risk of 9 of 14 CVD outcomes, including ischemic stroke, transient ischemic attack, thrombotic diseases, and 5 other CVDs (eg, coronary artery disease, heart failure, arterial hypertension).

Genetic predicted insomnia was associated with higher body mass index and triglycerides as well as lower high‐density lipoprotein cholesterol (without bidirectional causality), each of which may act as a mediator in the causal pathway from insomnia to several CVD outcomes.

These results are consistent with previous studies that suggest insomnia as an important causal risk factor for some major CVDs.

Insomnia is the second most prevalent mental disorder with an annual incidence of approximately 35% to 50% in the general population., In the past decade, there has been increasing evidence suggesting insomnia as an important risk factor of cardiovascular disease (CVD),, , , but evidence originating from observational studies (eg, the association between insomnia and hypertension, , , , , ) can sometimes be inconsistent, which may, at least in part, be explained by confounding or bias due to reverse causation. The American Heart Association published a scientific statement asking health organizations to develop evidence‐based sleep recommendations for a number of sleep disorders, including insomnia. Therefore, it is an urgent need to address the causal evidence to determine whether insomnia is in fact on the causal pathway for each CVD outcome. Although several Mendelian randomization (MR) studies have investigated that insomnia may be a causal risk factor for some major CVDs, including coronary artery disease (CAD), heart failure (HF), and ischemic stroke (IS),, , uncertainty persists about the causal role of insomnia for the development of other CVDs, such as arterial hypertension. Moreover, the potential pathways involved in the association from insomnia to CVDs have not been studied. Previous studies presented evidence that insomnia is associated with impaired glucose metabolism and several risk factors for CVD, such as body mass index (BMI) and waist‐hip ratio (WHR),, so cardiometabolic risk factors may act as potential mediators that lie in the pathway from insomnia to the risk of specific CVD outcome.

MR is a powerful approach to estimate the causal effect of an exposure on an outcome in observational data., The analytical method is less susceptible to bias due to confounders and evades potential bias by reverse causation through the use of genetic variants randomly allocated during conception, generally single‐nucleotide polymorphisms (SNPs), as instrumental variables for exposure., The following 3 key assumptions guarantee the altered genetic variants to be valid instrumental variables: (1) (Relevance) genetic variants are associated with the exposure; (2) (Independence) genetic variants are not associated with any confounder of the exposure‐outcome association; and (3) (Exclusion restriction) genetic variants are not associated with outcome conditional on exposure and confounders., Well‐powered genome‐wide association studies (GWAS) have identified hundreds of SNPs associated with insomnia; additionally, several consortia with large numbers of participants have made summarized data publicly available for many risk factors. These create the opportunity to conduct MR analysis in a 2‐sample strategy using summarized data to obtain more precise estimates of the causal effects. Moreover, compared with standard MR, the network MR method is increasingly being used to answer causal mediation questions within the MR framework. It calculated the causal effects of the exposure on a mediator and the mediator on the outcome; these 2 estimates can then be multiplied together to estimate the indirect effect., , The proportion mediated is calculated by dividing the indirect effect by the total effect; therefore, the network MR method can be used to investigate the effect of an exposure on the outcome operating through each mediator (indirect effect).

In this study, based on summary data obtained from large consortia and calculated based on individuals from UK Biobank, we performed a series of 2‐sample MR analyses to establish whether insomnia has a causal effect on the risks of 14 CVD outcomes. Then we conducted a mediation analysis to examine each of 17 cardiometabolic risk factors as a possible mediator in the causal relationship between genetically determined insomnia and each CVD outcome using a network MR design. A flow chart of our study design was provided in Figure 1.

Figure 1

The flow chart of study design and data sources for each analysis performed.

CKDGen indicates the Kidney Disease Genetics consortium; CVD, cardiovascular disease; eGFR, estimated glomerular filtration rate; ENGAGE, European Network for Genetic and Genomic Epidemiology; GIANT, the Genetic Investigation of Anthropometric Traits consortium; GWAS, genome‐wide association study; HbA1c, hemoglobin A1c; IVW, inverse variance weighted; MAGIC, the Meta‐Analyses of Glucose and Insulin‐Related Traits consortium; MR, Mendelian randomization; and MR‐PRESSO, Mendelian randomization‐pleiotropy residual sum and outlier.

Methods

The data used in these analyses are publicly available. The UK Biobank study was approved by the North West Multi‐centre Research Ethics Committee, and all its participants provided informed consent. The UK Biobank data were accessed through application 51470. The analyses of other publicly available data or summary statistics do not require additional ethical approval. All generated results were presented in the article and its supplement.

Genetic Instruments

Instrumental variables for insomnia were obtained from a GWAS meta‐analysis of 1 331 010 European‐descent individuals, the insomnia complaints were measured using questionnaire data, and the specific definition was provided in Table S1. This study identified 248 independent lead SNPs (r 2<0.1) located in 202 genomic risk loci (distinct genomic loci are >250kb apart) associated with insomnia at genome‐wide significance (2‐sided P value from the meta‐analysis of the GWAS results of insomnia: P<5×10−8), explaining 2.6% of the variance in insomnia (corresponding to an F statistic of 143.24). This means that assumption (1) is satisfied and avoids the bias caused by weak instrument in terms of the rule of thumb; see Data S1 for the details. The effect estimates (in units of log odds ratios [ORs]) and corresponding standard errors of these SNPs were extracted from Table S6 in Jansen et al (the details of those SNPs were provided in the Table S2). For all analyses in this study, SNPs were aligned to the same effect allele across the data sources before analyses, and we checked the effect allele frequencies for concordance.

Data Sources

Data Sources of CVD Outcomes

The summary statistics of genetic associations with the outcomes were derived based on the imputed genotype data of the UK Biobank, an ongoing large prospective cohort study. The UK Biobank recruited over 500 000 people aged 40 to 69 years, mean age 56.5 years, from Great Britain between 2006 and 2010, with 94% self‐reported European ancestry, 45.6% men, and median follow‐up time currently 11.1 years (https://www.ukbiobank.ac.uk/). In this study, we considered outcomes including the following 14 subtypes of CVD as Larsson et al did: cerebrovascular diseases (IS, transient ischemic attack, intracerebral hemorrhage, and subarachnoid hemorrhage), aortic aneurysms (abdominal and thoracic aortic aneurysm), thrombotic diseases (deep vein thrombosis and pulmonary embolism), and other CVDs (CAD, aortic valve stenosis, atrial fibrillation [AF], HF, peripheral vascular disease as well as arterial hypertension). The outcomes were defined based on electronic health records, hospital procedure codes, and self‐reported information confirmed by interview with a nurse. The specific definitions and sources of information can be found in Table S3. For quality control, we excluded participants with (1) non‐White British ancestry, (2) mismatch between genetic sex and self‐reported sex, (3) excess relatedness with kinship more than 10 putative third‐degree relatives, (4) poor‐quality genotyping with heterozygosity and missing rates > 1.5%, or (5) outliers with high heterozygosity and high missing rate (see Figure S1 for the flow chart of individual selection). Finally, a total of 424 811 UK Biobank participants who satisfied the inclusion criteria were included in this study. The mean age was 57.37 years (5th to 95th percentile: 43 to 69 years), and 45.65% of the population were men (descriptive statistics were reported in Table S4). For each CVD outcome, individuals suffering from any other 13 CVD outcomes or without genetic data were further excluded from the analysis’s data set (Figure S1). Then the genetic associations with each CVD outcome (on a log OR scale) were obtained from the identified White British individuals in UK Biobank using logistic regression controlling for 10 principal components, which can further control for population stratification.

Data Sources of Cardiometabolic Risk Factors

The genetic association estimates with 17 cardiometabolic risk factors, including anthropometric indexes (WHR adjusted for BMI, waist circumference [WC] adjusted for BMI, hip circumference adjusted for BMI, WHR, WC, hip circumference, and BMI), lipids (total cholesterol [TC], high‐density lipoprotein cholesterol [HDL‐C], low‐density lipoprotein cholesterol [LDL‐C], and triglycerides [TG]), glycemic traits (fasting glucose, fasting insulin, 2‐hour glucose, hemoglobin A1c), renal function (estimated glomerular filtration rate), and heart rate increase during exercise were taken from the GIANT (Genetic Investigation of Anthropometric Traits) consortium,, the GLGC (Global Lipids Genetic Consortium), the MAGIC (Meta‐Analyses of Glucose and Insulin‐related Traits Consortium),, , the CKDGen (Kidney Disease Genetics) consortium, and published GWAS study in Verweij et al, respectively. We constrained the population of these GWAS summary statistics data mainly from European ancestry to minimize the bias caused by population stratification. The basic characteristics of these consortia and the genetic association estimates were presented in Table S5.

Statistical Analysis

Primary analysis

Our primary analysis aimed to explore the causal associations of insomnia with 14 CVD outcomes. Because 1 of 248 SNPs (rs77641763 in chr9) was unavailable in the outcome data set (UK Biobank), 247 SNPs were used as genetic instrumental variables for insomnia.

For each CVD outcome, the overall causal estimate of insomnia on this CVD outcome was obtained using an inverse variance weighted (IVW) method (under a multiplicative random‐effects model), combining the variant‐specific Wald (ratio) estimators, estimated for each SNP through the SNP‐CVD association divided by the SNP‐insomnia association. This method provides the highest precision and retains maximal power under the assumption that all SNPs are valid instrumental variables, that is, the pleiotropic effects of the genetic variants (genetic variant affects the outcome via a different biological pathway from the exposure under investigation) are all zero. Moreover, it accounts for heterogeneity in the variant‐specific causal estimates. The results were converted to ORs expressed per genetically predicted 1‐unit‐higher log‐odds of liability to insomnia, corresponding to per 2.72‐fold increase in the prevalence of insomnia according to Burgess et al. As many MR analyses with multi‐outcomes did, to account for multiple testing and to preserve the type Ⅰ error of the global null hypothesis of all tested associations being in fact null, we used a Bonferroni‐corrected threshold of P<3.6×10−3 (α=0.05/14 outcomes) at the outcome level in our primary analysis. We note that these outcomes are related to each other; therefore, the tests are not completely independent of each other and the Bonferroni correction may be conservative. We reported the actual P values of each effect; the association between our threshold and 0.05 was considered suggestive evidence of association, requiring confirmation.

Mediation Analysis

We further conducted a mediation analysis to explore whether the chosen cardiometabolic risk factors mediate the causal pathway from insomnia to CVD outcome that insomnia is significantly associated with, using a network MR design. For each CVD outcome, this design consists of 3 different steps (Steps a‐c):

Step a: We estimated the causal effect of genetically determined insomnia on this CVD outcome, this step was in accordance with our primary analysis.

Step b: Up to 248 independent SNPs associated with insomnia at genome‐wide significance from Jansen et al were used to estimate the causal effect of genetically determined insomnia on each potential cardiometabolic risk factor, using the respective GWAS summary statistics described previously,, , , , , , , using the IVW approach (under a multiplicative random‐effects model). The estimated effects were SD or per unit change in risk factors expressed per genetically predicted 1‐unit‐higher log‐odds of liability to insomnia (per 2.72‐fold increase in the prevalence of insomnia). There was no sample overlap between insomnia GWAS study and consortia of all risk factors except for heart rate (~4%). The Bonferroni‐corrected threshold of P<3×10−3 (α=0.05/17 mediators) was used in this step.

Step c: For possible mediators that causal association was observed in Step b (BMI, HDL‐C, and TG), we estimated the causal effect of each mediator on this CVD outcome, respectively, using the IVW approach (under a multiplicative random‐effects model). The details of instrumental SNPs and the summary statistics we used for each mediator in Step c were shown in Data S2 and Table S6–S8, and SNP‐CVD outcome coefficients were calculated from the identified White British individuals in UK Biobank using logistic regression controlling for 10 principal components. There was no sample overlap between each mediator GWAS study and UK Biobank. The results were converted to ORs expressed per genetically predicted 1 SD increased of the mediator, and Bonferroni‐corrected threshold of P<5.6×10−3 (α=0.05/9 outcomes that insomnia significantly associated with) was used in this step.

If causal associations were observed in all 3 steps, the conclusion can be drawn that the specific cardiometabolic risk factor is a mediator in the pathway linking insomnia to this CVD outcome. The indirect effect of insomnia on this CVD outcome mediated through each mediator was estimated by multiplying the results from Step b and Step c. We finally divided the mediated effect by the total effect to estimate the proportion mediated by each mediator (see Data S3 for the calculation of proportions and Data S4 for the calculation of their 95% CIs).

Sensitivity Analysis

A series of sensitivity analyses were performed to examine the robustness of the results in the primary analysis and each step for mediation analysis. First, the presence of substantial heterogeneity among Wald estimates based on individual SNPs would indicate the presence of invalid instruments. In this study, the magnitude of heterogeneity among variant‐specific Wald estimates was evaluated using the I 2 statistic, which is defined as the percentage of the total variation in the estimates explained by heterogeneity rather than sampling error if all genetic variants are valid instruments and relationships between all variables (genetic variants, exposure, and outcome) are linear and homogeneous for all individuals in the population, independent of the number of estimates., Moreover, leave‐one‐out sensitivity analysis was also performed to assess the reliance of the MR results on each particular variant.

Second, we performed a range of robust methods making different consistency assumptions. We examined the potential bias from invalid instruments when using multiple genetic variants through weighted median and mode‐based estimation, which rely on weaker assumptions about the invalid instruments and are more robust to outliers. The former method will provide a consistent estimate if at least 50% of the weights come from valid instrumental variables, and the latter assumes more variants estimate the true causal effect than estimating any other quantity, but have been shown to have low precision in some studies and simulations. In addition, we applied the MR‐Egger regression method, which can obtain a valid estimate of the causal effect even if all the genetic instruments are invalid under the Instrument Strength Independent of Direct Effect assumption, but with low power. Moreover, the non‐zero intercept indicates the existence of directional pleiotropy; therefore, MR‐Egger intercept test can help to assess the validity of instrumental variable assumption (3). MR pleiotropy residual sum and outlier method was also used to detect and correct for horizontal pleiotropic through the removal of SNPs that are most likely to display horizontal pleiotropic effects (outliers). As suggested by Slob and Burgess, these 4 methods can adequately assess the evidence of the causal effects of each exposure on the outcome and detect the sensitivity of the results to different patterns of violations of instrumental variable assumptions; consistency of results across methods strengthens an inference of causality.

Additionally, we performed a bidirectional MR analysis to examine whether each selected mediator can casually affect insomnia (bidirectional causality) by using mediator‐associated independent SNPs as instrumental variables (Data S2 and Table S6–S8), the summary statistics of insomnia were obtained from Jansen et al. If there was no evidence that one genetically predicted mediator causally affects insomnia, then we applied regression‐based multivariable MR as an additional sensitivity analysis of Step c for mediation analysis. We estimated the causal effect of this mediator on each CVD outcome, respectively, adjusting for the genetic effects of the instruments on insomnia obtained from Jansen et al, as suggestive by Carter et al., We additionally performed multivariable MR to consider the role of multiple mediators simultaneously and to investigate the direct causal effect of insomnia on each CVD outcome not mediated by these 3 mediators; see Data S5 for the details.

In our analyses, we considered arterial hypertension as a CVD outcome. As blood pressure traits (including systolic blood pressure, diastolic blood pressure [DBP]) are also important cardiometabolic risk factors, we additionally performed analyses to explore whether systolic blood pressure and DBP mediate the causal pathway from insomnia to the other 8 insomnia‐associated CVD outcomes using a network MR design. The details were provided in Data S6.

Replication Analysis

We conducted replication analysis by varying different data sets to further assess the robustness of our results. For the primary analysis, we estimated that approximately 30% of individuals from GWAS on insomnia were also included in the UK Biobank. Thus, we further performed a replication analysis using summary data of 4 CVD outcomes, including IS, CAD, AF, and HF, from previously published GWAS studies, , , that have no or limited sample overlap with the GWAS of insomnia. Full details of each GWAS can be found in Table S9. We also validated these results with UK Biobank individual‐level data in a 2‐sample setting (see Data S7). For mediation analysis, we performed replication analysis using a most recent data (the ENGAGE (European Network for Genetic and Genomic Epidemiology) 1000 Genome Consortium, http://diagram‐consortium.org/2015_ENGAGE_1KG/,, see Table S10 for the details) as the replication data set for 6 anthropometric and lipid traits in Step b. For Step c, we also performed replication analysis using summary statistics of 4 CVD outcomes from published GWAS studies (Table S9). All replication analysis were conducted using IVW in addition to similar heterogeneity assessment approaches and 4 robust methods for sensitivity analysis.

All statistical tests were 2 sided. Analyses in this study were performed using PLINK2 and R version 4.0.2 together with the R package MendelianRandomization, MRPRESSO.

Results

Causal Association Between Genetically Predicted Insomnia and CVD Outcomes

We found that genetically predicted insomnia was significantly positively associated with 10 of the 14 outcomes using the IVW method; the ORs ranged from 1.13 (95% CI, 1.08–1.18) for AF to 1.24 (95% CI, 1.16–1.32) for HF (Figure 2). Additionally, there was suggestive evidence of positive associations between genetically predicted insomnia and intracerebral hemorrhage (OR, 1.18; 95% CI, 1.02–1.37; P=0.03) and subarachnoid hemorrhage (OR, 1.23; 95% CI, 1.06–1.43; P=0.005), whereas statistically nonsignificant positively associations were observed between insomnia and abdominal aortic aneurysm (OR, 1.14; 95% CI, 1–1.30; P=0.06) or thoracic aortic aneurysm (OR, 1.02; 95% CI, 0.79–1.31; P=0.9) (Figure 2).

Figure 2

Associations between genetically predicted insomnia and 14 cardiovascular diseases (CVDs).

Results were obtained from the multiplicative random‐effects inverse‐variance weighted method. Sample size denotes the total number of individuals in the analysis data set for each CVD outcome (see Figure S1 for flow chart of individual selection). Estimates represent odds ratios (OR) expressed per genetically predicted 1‐unit‐higher log‐odds of liability to insomnia (per 2.72‐fold increase in the prevalence of insomnia); I 2 statistic quantifies the amount of heterogeneity among estimates based on individual single‐nucleotide polymorphisms. Circles show the estimated ORs, and the sizes of these circles represent the precision of these estimates.

There was no evidence of SNP that has a strong influence on the estimations of causal association (Table S11). Although we noted that MR‐Egger and mode‐based estimates were sometimes different from the other MR methods, for example, pulmonary embolism and aortic valve stenosis, most OR estimates were consistent using different MR approaches (Table S12). The intercept of MR‐Egger regression for 247 SNPs in each CVD outcome was not statistically significant except for intracerebral hemorrhage (Intercept=−0.03; 95% CI, −0.06 to −0.01; P=0.01). After excluding SNPs because of their pleiotropic effects, the analyses of the remaining SNPs did not materially change the OR of insomnia on intracerebral hemorrhage (OR, 1.15; 95% CI, 1–1.34; P=0.057). However, results of replication analysis showed that genetically predicted insomnia was still significantly positively associated with 9 of the 10 outcomes identified, except for aortic valve stenosis (OR, 1.25; 95% CI, 0.93–1.68; P=0.14) (Table S13–S16).

Causal Association Between Genetically Predicted Insomnia and Cardiometabolic Risk Factors

The causal estimates between genetically predicted insomnia and 17 cardiometabolic risk factors using the IVW method were displayed in Figure 3. Genetically predicted insomnia was significantly positively associated with 4 risk factors, including BMI (0.07; 95% CI, 0.04–0.10), TG (0.06; 95% CI, 0.04–0.09), WC (0.06; 95% CI, 0.03–0.09), and WHR (0.05; 95% CI, 0.03–0.07). However, no association was observed with insomnia for the latter 2 indexes after the adjustment for BMI (WC adjusted for BMI [0; 95% CI, −0.03 to 0.02] and WHR adjusted for BMI [0.02; 95% CI, −0.01 to 0.04]). Moreover, genetically predicted insomnia was significantly negatively associated with HDL‐C (−0.06; 95% CI, −0.09 to −0.03).

Figure 3

Associations between genetically predicted insomnia and 17 cardiometabolic risk factors.

Results were obtained from the multiplicative random‐effects inverse‐variance weighted method. Estimated effects were SD or per unit change in risk factor expressed per genetically predicted 1‐unit‐higher log‐odds of liability to insomnia (per 2.72‐fold increase in the prevalence of insomnia). I 2 statistic quantifies the amount of heterogeneity among estimates based on individual SNPs. N SNPs was the number of instrument SNPs we used for Mendelian randomization analysis of the association of genetically predicted insomnia on each risk factor. BMI indicates body mass index; eGFR, estimated glomerular filtration rate; HbA1c, hemoglobin A1c; HDL‐C, high‐density lipoprotein cholesterol; HIP, hip circumference; HIPadjBMI, HIP adjusted for BMI; LDL‐C, low‐density lipoprotein cholesterol; SNP, single‐nucleotide polymorphism; TC, total cholesterol; TG, triglycerides; WC, waist circumference; WCadjBMI, WC adjusted for BMI; WHR, waist‐hip ratio; and WHRadjBMI, WHR adjusted for BMI.

There was no evidence of SNP that has a strong influence on the estimations of causal associations (Table S17). Associations estimated by the weighted median, mode‐based, MR‐Egger, and MR‐PRESSO method were broadly similar to our main results for BMI, TG, and HDL‐C, with little evidence for the presence of pleiotropy (Table S18). However, the association estimated by mode‐based and MR‐Egger method was generally underpowered with wide CIs. Our replication studies for BMI, TG, and HDL using the ENGAGE 1000 Genome data also provided consistent results (Tables S19–S20). There was consistently no statistically significant association between genetically predicted insomnia with TC, LDL‐C, glycemic traits, renal function as well as heart rate increase during exercise. Finally, 3 of 17 cardiometabolic risk factors including BMI, TG, and HDL‐C were suggested to be possible mediators from insomnia to CVD outcomes and were chosen to conduct further analyses.

Causal Association Between Genetically Predicted BMI, TG, HDL, and CVD Outcomes

We evaluated further whether each of BMI, TG, and HDL causally associated with each of 9 selected CVD outcomes. To begin with, we used 74 of 78 independent SNPs (R 2=2.2%) associated with BMI (P<5×10−8) as instrumental variables. Under a Bonferroni‐corrected threshold of P<0.0056, a genetically predicted 5 kg/m2 (1 SD) increase of BMI was significantly associated with 7 of 9 CVD outcomes using the IVW method, with ORs ranging from 1.52 (95% CI, 1.37–1.67) for arterial hypertension to 2.14 (95% CI, 1.78–2.58) for HF (Figure S2). There was no evidence of SNP that has a strong influence on the estimations of causal associations and presents of pleiotropy (Tables S21–S22), except for the analyses of BMI on CAD. After excluding SNPs due to the potential pleiotropic effects, analyses of the remaining SNPs yielded a similar OR of 1.57 (95% CI, 1.37–1.79; P=2.89×10−11). The results of sensitivity analysis and replication analysis (Tables S23–S24) were broadly similar to our main results for BMI.

Then 85 of 86 independent SNPs (R 2=5.9%) as well as 51 independent SNPs (R=4.6%) associated with HDL‐C and TG, respectively, were used as instrumental variables. According to the results of IVW method, a genetically predicted 1 SD increase of HDL‐C was negatively associated with higher risk of CAD (OR, 0.82; 95% CI, 0.73–0.91), peripheral vascular disease (OR, 0.80; 95% CI, 0.69–0.94) and arterial hypertension (OR, 0.86; 95% CI, 0.80–0.93) (Figure S3). A genetically predicted 1 SD increase of TG was positively associated with higher risk of CAD (OR, 1.44; 95% CI, 1.29–1.61), HF (OR, 1.33; 95% CI, 1.17–1.51), and arterial hypertension (OR, 1.18; 95% CI, 1.09–1.27) (Figure S4). There was no evidence of SNPs that have a strong influence on the estimations of causal associations (Table S25–S26). After excluding SNPs suggested to have potential pleiotropic effects (Table S27–S28), analyses of the remaining SNPs did not materially change the ORs of HDL‐C on CAD (OR, 0.68; 95% CI, 0.6–0.79; P=2.52×10−8), HDL‐C on arterial hypertension (OR, 0.78; 95% CI, 0.72–0.84; P=3.48×10−10), TG on CAD (OR, 1.39; 95% CI, 1.25–1.54; P=3.88×10−10), and TG on arterial hypertension (OR, 1.16; 95% CI, 1.08–1.24; P=2.88×10−5). However, results of replication analysis did not support that a genetically predicted 1 SD increase of HDL‐C was significantly associated with a higher risk of CAD (OR, 0.88; 95% CI, 0.8–0.98; P=0.015) under a Bonferroni‐corrected threshold of P<0.0125; other replication analysis results were accordant with our main results (Table S29–S32).

The results of bidirectional MR showed no evidence of causal association of genetically predicted BMI, HDL‐C, or TG on insomnia (Table S33). Thus, insomnia does not tend to lie on the causal pathway from each mediator to each CVD outcome (ie, insomnia is not the mediator of risk factor‐CVD association). Therefore, for all 3 mediators, regression‐based multivariable MR was conducted as an additional sensitivity analysis to estimate the causal effect of the mediator on each CVD outcome, respectively, adjusting for the genetic effect of the instruments on insomnia. The results of multivariable MR shown in Figure S5–S7 were highly accordant to our main results.

According to the results of network MR analysis, we found that some or all of BMI, TG, and HDL might act as mediators in the causal pathway from insomnia to deep vein thrombosis, pulmonary embolism, CAD, AF, HF, peripheral vascular disease, and arterial hypertension. The proportion of the total effect of insomnia on each CVD outcome that each mediator accounts for was provided in the Table. The results of the direct causal effect of insomnia on each CVD outcome not mediated by these 3 mediators were shown in Data S5 and Table S34.

The Potential Mediator Role of Blood Pressure Traits

Finally, when we consider blood pressure traits (Table S35) as cardiometabolic risk factors rather than consider arterial hypertension as a CVD outcome, our results showed that genetically predicted insomnia was significantly positively associated with DBP (0.41; 95% CI, 0.24–0.58; P=1.75×10−6); nevertheless, no association was observed with insomnia and DBP after adjustment for BMI (0.11; 95% CI, −0.05 to 0.27) (Table S36). In addition, a genetically predicted 1 mm Hg increase of DBP was positively associated with a higher risk of IS (OR, 1.09; 95% CI, 1.06–1.13), CAD (OR, 1.08; 95% CI, 1.06–1.09), AF (OR, 1.04; 95% CI, 1.02–1.06), and HF (OR, 1.07; 95% CI, 1.04–1.09) (Table S37) after Bonferroni correction, and the results of bidirectional MR showed no evidence of causal association of genetically predicted DBP on insomnia (Table S38). The sensitivity analysis yielded a similar pattern of effects; there was no evidence of the presence of pleiotropic effect (Table S36 and Table S38). In addition, we found that no SNP had an influential effect on the results for each blood pressure trait (Table S39). Therefore, DBP might act as a mediator in the causal pathway from insomnia to IS, CAD, AF, and HF. The proportion of the total effect of insomnia on each CVD outcome that DBP accounts for was provided in Table S40.

Discussion

Principal Findings

In this study, using publicly available summary statistics from large consortia and data from UK Biobank, we performed a series of 2‐sample MR analyses to systematically assess the causal roles of genetically predicted insomnia for a wide range of cardiovascular conditions. We further explored 17 cardiometabolic risk factors as possible mediators in the causal relationship between genetically determined insomnia and each CVD outcome. Our results provided consistent evidence that genetically determined insomnia was causally associated with increased risk of IS, transient ischemic attack, thrombotic diseases, and 5 other CVDs (CAD, HF, AF, peripheral vascular disease, arterial hypertension). Additionally, we concluded that genetically predicted insomnia was associated with higher BMI and TG as well as lower HDL‐C, each of which may act as a mediator in the causal pathway from insomnia to several CVD outcomes. In addition, we found very little evidence to support a causal link between genetically predicted insomnia with abdominal aortic aneurysm, thoracic aortic aneurysm, TC, LDL‐C, glycemic traits, renal function as well as heart rate increase during exercise. Finally, we found no evidence of a causal association of genetically predicted BMI, HDL‐C, or TG on insomnia.

Comparisons With Other Studies

Although caution should be taken when comparing our results with other observational studies, because of the variations in measurement and definition of insomnia, nonetheless, our findings for CVDs from the primary analysis are generally consistent with those based on observational studies, which suggested that insomnia is an important risk factor for CVDs including hypertension,, CAD, AF,, , and HF, . The null results for genetically predicted insomnia and abdominal aortic aneurysm are consistent with findings from a cohort study of 22 444 men and 10 982 women showing no association between insomnia and later development of abdominal aortic aneurysm, but we cannot rule out the possibility that our analyses were underpowered to detect such a relatively modest association. Our results are also similar to previously reported MR studies of insomnia and risk of CAD,, , , HF, and IS. For AF, the OR of our analyses was similar to the result of Larsson et al (1.04; 95% CI, 1.01–1.07; P=7.28×10−3), but they did not conduct a significant conclusion under a stricter Bonferroni‐corrected threshold (). We are not aware of any observational or MR study of insomnia in relation to other cardiovascular conditions. In this study, we found genetically predicted insomnia was associated with an increased risk of transient ischemic attack, deep vein thrombosis, pulmonary embolism, and peripheral vascular disease.

Associations between insomnia and lipid markers (TC, TG, LDL‐C, and HDL‐C) reported by observational studies have been inconsistent., , , , , In addition, a previous observational study has shown that insomnia is not associated with a higher BMI, insomnia symptoms were found to have a close association with high hemoglobin A1c in Japanese men, insomnia scores were found to have a strong negative correlation with estimated glomerular filtration rate. Our findings for cardiometabolic risk factors are generally contradictory to those based on observational studies but similar to previously reported MR studies on insomnia and BMI as well as WHR. In this study, we found strong and consistent evidence that genetically predicted insomnia was associated with higher BMI, WHR, WC, and TG as well as lower HDL‐C. Moreover, there was consistent no statistically significant association between genetically predicted insomnia with TC, LDL‐C, glycemic traits, estimated glomerular filtration rate as well as heart rate increase during exercise. The contradiction may reflect reverse causation or confounding effects of previous observational studies.

In addition, our results that genetically predicted BMI was significantly associated with 7 of 9 CVD outcomes were consistent with those results from MR analysis in Larsson et al previously. Several MR studies have investigated the associations between genetically predicted lipid traits and the risk of some major CVDs. For instance, Hindy et al reported that genetically elevated TG did not associate with IS or any of its subtypes; for HDL‐C, only weak evidence of association with 1 subtype of IS (small artery occlusion stroke) was found. A higher genetically predicted TG was observed to be causally associated with a higher risk of hypertension and CAD, and lower genetically predicted HDL‐C was observed to be causally associated with a higher risk of hypertension and no significant association with CAD. Our finding was consistent with previous MR studies, which showed that a genetically predicted 1 SD increase of HDL‐C was negatively associated with a higher risk of peripheral vascular disease and arterial hypertension. A genetically predicted 1 SD increase of TG was positively associated with a higher risk of CAD, HF, and arterial hypertension.

Potential Mechanisms

The mechanisms underpinning the association are unclear but may involve that insomnia increases the activity of hypothalamic‐pituitary‐adrenal axis, , , and increases the systemic inflammation, ; these provide a biologic rationale for insomnia to possibly lead to increased risk of CVD. In addition, studies also demonstrated that subjects with insomnia had increased sympathetic nervous system activity, which is an integral part of cardiovascular homeostasis and plays a critical role in the pathogenesis of hypertension, CAD, and HF. Moreover, there might be potential mechanisms to explain the link between insomnia and dyslipidemia as well as the increase of BMI. Genetically determined insomnia usually contributes to reduced sleep duration. People with reduced sleep duration tend to show a preference for high energy‐density fatty food and have a higher BMI via reducing leptin and elevating ghrelin., Moreover, higher BMI and intake of fat rich food are able to increase the risk of dyslipidemia, which may further lead to the increased risk of CVD.

Strengths and Limitations

In this study, we performed a series of 2‐sample MR analyses, which reduced the possibility that the observed associations were biased by unmeasured confounding, reverse causation, and measurement error. We conducted mediation analysis using a network MR design, which can be used to decompose the direct and indirect effects, and can overcome some of the strong assumptions required for traditional causal mediation methods (eg, no unmeasured confounding and no measurement error in the exposure or mediator, which are unlikely to be met in many scenarios, , ) to interpret the results as causal. Moreover, most of these analyses were performed using summarized data, which considerably increases the power to detect a causal effect. We used multiple independent genetic instruments that explained a large variance of exposure, further validated the relevance assumption by the rule of thumb, and tested the exclusion assumption by MR‐Egger intercept test and the MR‐PRESSO test. Although we cannot entirely rule out the possibility of the violations of MR assumptions (eg, the independence assumption), most of our results were corroborated by both sensitivity analysis that makes allowance for violation of MR assumptions and the replication analysis, which further strengthens the inference of causality. In addition, we have made sure that most of our summary data estimates of SNP‐exposure and SNP‐outcome associations were gleaned from 2 independent but large homogeneous (European descent) study populations, which minimized the bias from population stratification.

Our study has several limitations. First, insomnia was determined based on self‐reported questionnaires, which may not be completely accurate. In addition, it should be cautious when considering our results, which may vary when using different definitions of insomnia. Insomnia we used is a binary exposure. Some studies have pointed that an MR estimate with a binary exposure and binary outcome is difficult to interpret as a specific causal effect., Second, in our primary analysis, the number of cases was few for some CVD outcomes, such as intracerebral hemorrhage, subarachnoid hemorrhage, and abdominal aortic aneurysm. We thus cannot rule out that we may have overlooked weak associations due to insufficient power. Third, the reason SNP‐outcome estimates were obtained from UK Biobank is the availability of adequate outcome variables, such as arterial hypertension, aortic valve stenosis, which lacks published GWAS study. Nevertheless, there is around 30% sample overlap between UK Biobank and insomnia GWAS study, which may have introduced some bias in the MR estimates of primary analysis. However, we used a strict Bonferroni‐corrected threshold to control the possible increase of type I error rate caused by sample overlap. In addition, we conducted replication studies using summary data from previously published GWAS studies that have no or limited sample overlap with the GWAS of insomnia and UK Biobank individual‐level data as sensitivity analysis, the same way as Siedlinski et al did, to validate our results. To be noted, our genetic instruments were strongly related to the exposures (F statistic≈143.24), implying that bias from participant overlap is relatively small according to Burgess et al. Fourth, this study evaluated only whether the liability to insomnia is associated with CVD, so we cannot rule out that there are other causal pathways leading to insomnia that causes CVD.

Conclusions

This MR study provides evidence that genetically predicted insomnia is associated with increased risks of 9 CVD outcomes, some of them may be partially mediated by one or more of higher BMI, TG, and lower HDL‐C.

Sources of Funding

This work was supported by the National Natural Science Foundation of China (Grant number 81773547 and 82003557), the National Key Research and Development Program of China (Grant number 2020YFC2003500), the Shandong Provincial Natural Science Foundation of China (Grant number ZR2019ZD02), and Shandong Provincial Key Research and Development project (Grant number 2018CXGC1210).

Disclosures

None.

Data S1–S7

MEGASTROKE project authors

Tables S1–S40

Figures S1–S7

References 96–106

Click here for additional data file.

Table 1

The Proportion of the Total Effect of Insomnia on Each Cardiovascular Disease That Each Mediator Accounts For

Exposure (X)	Mediator (M)	Outcome (Y)	TEXY	βXM	ORMY	NIEXY (95% CI)	Proportion (95% CI)
Insomnia	BMI	Deep vein thrombosis	1.15	0.07	1.79	0.041 (0.02– 0.061)	29.16% (14%–44.32%)
Insomnia	BMI	Pulmonary embolism	1.16	0.07	1.66	0.035 (0.014–0.057)	23.9% (8.53%–39.27%)
Insomnia	BMI	Coronary artery disease	1.22	0.07	1.53	0.03 (0.014–0.046)	14.97% (7.44%–22.5%)
Insomnia	triglycerides	Coronary artery disease	1.22	0.06	1.44	0.022 (0.01–0.034)	11% (5.03%–16.98%)
Insomnia	BMI	Atrial fibrillation	1.13	0.07	1.56	0.031 (0.015–0.048)	25.47% (11.67%–39.27%)
Insomnia	BMI	Heart failure	1.24	0.07	2.14	0.053 (0.026–0.081)	24.76% (12.93%–36.59%)
Insomnia	triglycerides	Heart failure	1.24	0.06	1.33	0.017 (0.006, 0.028)	7.95% (2.5%–13.41%)
Insomnia	BMI	Peripheral vascular disease	1.23	0.07	1.73	0.038 (0.013–0.063)	18.53% (6.5%–30.57%)
Insomnia	HDL‐C	Peripheral vascular disease	1.23	−0.06	0.8	0.013 (0.001–0.025)	6.47% (0.53%–12.4%)
Insomnia	BMI	Arterial hypertension	1.14	0.07	1.52	0.029 (0.014–0.044)	22.37% (12.07%–32.67%)
Insomnia	HDL‐C	Arterial hypertension	1.14	−0.06	0.86	0.009 (0.002–0.016)	6.91% (1.75%–12.06%)
Insomnia	triglycerides	Arterial hypertension	1.14	0.06	1.18	0.01 (0.003–0.017)	7.58% (2.42%–12.74%)

BMI indicates body mass index; βXM, effect of the exposure on the mediator; HDL‐C, high‐density lipoprotein cholesterol; TE, total effect of the exposure on the outcome expressed in odds ratios (OR) scale; NIE, natural indirect effect of exposure on the outcome in log OR scale; and Proportion, the proportion of the total effect of exposure on outcome that mediator accounts for.