Mendelian randomization study on atrial fibrillation and cardiovascular disease subtypes

Atrial fibrillation (AF) has been associated with numerous diseases. However, whether AF is a cause or consequence of these diseases is uncertain. To clarify, we assessed the causal role of AF on ischemic heart disease (IHD), stroke, other cardiovascular disease (CVD) subtypes, type 2 diabetes mellitus (T2DM), and late-onset AD using bi-directional two-sample Mendelian randomization (MR) among people primarily of European descent. Genetically predicted log odds of AF was associated with any stroke (odds ratio (OR) 1.22, 95% CI 1.18 to 1.27), particularly cardioembolic stroke and possibly subdural hemorrhage, with sensitivity analyses showing similar positive findings. Genetically predicted AF was also associated with arterial thromboembolism (1.32, 1.13 to 1.53), and heart failure (1.26, 1.21 to 1.30). No association of genetically predicted AF with IHD, T2DM, cognitive function, or late-onset AD was found. Conversely, genetically predicted IHD, heart failure and possibly ischemic stroke, particularly cardioembolic stroke, were positively associated with AF. Atrial fibrillation plays a role in any stroke, arterial thromboembolism, and heart failure, corroborating current clinical guidelines on the importance of preventing these complications by effective AF management. In addition, patients with IHD, heart failure or possibly ischemic stroke might be predisposed to developing AF, with implications for management.


Results
Association of genetically predicted AF with IHD, stroke, other CVD subtypes, T2DM, cognitive function, and late-onset AD. For genetic predictors of AF, we obtained 110 SNPs with the F-statistic = 89. 3. Nearly none of the SNPs for AF and the SNPs for various outcomes (except 1 SNP for AF with cognitive function) or exposures (except 1 SNP for cardioembolic stroke and 1 SNP for heart failure with AF) were identical or highly correlated. Genetically predicted AF was unrelated to education, Townsend deprivation index, smoking, alcohol drinking, and physical activity (Table 1). Figure 1 (and Appendix Table 1) shows genetically predicted AF was not associated with IHD, but was consistently positively associated with any stroke, and major stroke sub-types: ischemic stroke and possibly hemorrhagic stroke. In particular, genetically predicted AF was positively associated with cardioembolic stroke at Bonferroni-corrected significance and subdural hemorrhage at nominal significance, and perhaps intracerebral hemorrhage (although its confidence interval included the null value). No association of genetically predicted AF with large artery stroke, small vessel stroke or subarachnoid hemorrhage was found. Genetically predicted AF was also positively associated with arterial thromboembolism and heart failure. Sensitivity analyses using a weighted median, MR-Egger and/or MR-PRESSO showing similar positive findings. Genetically predicted AF was not associated with T2DM, cognitive function, or late-onset AD.
Association of genetically predicted IHD, stroke, other CVD subtypes, T2DM, cognitive function, and late-onset AD with AF. Figure 2 (and Appendix Table 2) shows genetically predicted IHD was positively associated with AF. Genetically predicted stroke (possibly any and ischemic, but not hemorrhagic) was also positively associated with AF at nominal significance. In particular, genetically predicted cardioembolic stroke and heart failure were positively associated with AF at Bonferroni-corrected significance, and perhaps small vessel stroke (although its confidence interval included the null value). No association of genetically predicted large artery stroke, subarachnoid hemorrhage, intracerebral hemorrhage or subdural hemorrhage with AF was found. Sensitivity analyses using a weighted median, MR-Egger and/or MR-PRESSO showing similar positive findings. No association of genetically predicted arterial thromboembolism, T2DM, cognitive function, or late-onset AD with AF was found. After excluding the overlapping SNP, similar patterns of association for AF with cognitive function and for cardioembolic stroke with AF were found (Appendix Table 3). Scatter plots showing associations for AF with these cardiovascular diseases and vice versa are shown in Appendix Figures 1  and 2.
Power calculations showed that this study had 80% power to detect OR ranging from 1.10 to 2.77 for the associations of AF with various diseases, and OR ranging from 1.10 to 1.71 for the associations of various diseases with AF (Appendix Table 4).

Discussion
This first two-sample MR study assessing the bi-directional relationships between AF and CVD, T2DM, cognitive function, and late-onset AD suggests that AF may have a bi-directional relationship with any or ischemic stroke, particularly cardioembolic stroke. Further, AF may be a consequence of IHD that causes possibly hemorrhagic www.nature.com/scientificreports/ stroke, particularly subdural hemorrhage, arterial thromboembolism, and heart failure, but not T2DM, cognitive function, or late-onset AD. Our MR findings corroborate current clinical guidelines on the importance of preventing arterial thromboembolism, any stroke, and heart failure for effective AF management 6,7 , given these diseases are likely complications of AF. These findings also suggest people with IHD, heart failure and possibly ischemic stroke may be more susceptible to develop AF, given these diseases are likely the causes of AF, with implications for management of AF.
Overall, our MR findings support the causal role of AF in stroke and its subtypes. Consistent with the MR of the bi-directional association between ischemic stroke and AF 23 , we found ischemic stroke, particularly cardioembolic stroke (and possibly small vessel stroke), predisposes to AF or vice versa. There was no association of hemorrhagic stroke and its subtypes with AF, which might be underpowered due to few cases in Western populations and future replication is warranted when a larger GWAS of hemorrhagic stroke and its subtypes becomes available. We add that AF may cause any stroke and its major subtypes, with both ischemic stroke (especially cardioembolic stroke) and possibly hemorrhagic stroke (subdural hemorrhage, and perhaps intracerebral hemorrhage) being potential consequences. Further, this study provides evidence that AF predisposes to higher risk of arterial thromboembolism, and heart failure, corresponding to current clinical guidelines on the prevention of AF complications 6,7 . In contrast, this study clarifies that AF unlikely predisposes to IHD, but actually IHD may increase AF, coherent with the RCT showing better prognosis (fewer strokes) in people with AF and IHD when receiving anticoagulants 18 . Consistent with the previous MR study on the association of T2DM with AF 24 , we found no association between AF and T2DM (bi-directionally) using genetic variants of T2DM identified from a large sample. In addition, our MR suggests AF unlikely affects cognitive function or, consistent with another MR study 26 , late-onset AD, substantiating the null effect in an open-label trial of warfarin on cognitive function 20 . Our findings are less consistent with observational studies reporting AF associated with any or ischemic (not hemorrhagic) stroke, IHD 9 , cognitive decline 11 , and late-onset AD 12,13 , whereas people with T2DM develop more AF 10 . However, observational studies in patients are difficult to interpret because of selection bias.
Taken together, these findings suggest a potentially causal link from IHD and possibly ischemic stroke (especially cardioembolic stroke, and perhaps small vessel stroke) to AF, as well as AF to ischemic (especially cardioembolic stroke) and possibly hemorrhagic stroke (subdural hemorrhage, and perhaps intracerebral hemorrhage), arterial thromboembolism, and heart failure, but not T2DM or late-onset AD. AF may be the consequence of damage to the heart via atrial ischemia/infarction from IHD 27,28 , as well as dysregulation of heart rate Table 1. Association of genetically predicted atrial fibrillation with socio-economic position (education, and Townsend deprivation index) and lifestyle (smoking, alcohol drinking, and physical activity) from the UK Biobank using Mendelian randomization (MR). CI, confidence interval; IVW, inverse variance weighting; MR, Mendelian randomization, SNP, single nucleotide polymorphism; WM, weighted median. a MR-PRESSO estimate was obtained by excluding 4 outliers (rs12245149, rs2040862, rs35005436, rs3820888) for education, 1 outlier (rs6596717) for Townsend deprivation index, 3 outliers (rs11191116, rs1458038, rs6596717) for physical activity. www.nature.com/scientificreports/ via impaired autonomic nervous system from stroke 29,30 . AF may contribute to thrombosis and hence arterial thromboembolism because more turbulent blood flow resulting from an irregular heart beat could damage endothelial structure/function and cause blood stasis, hence promoting coagulation 31 . In particular, AF likely facilitates the coagulation cascade rather than platelet aggregation, considering the failure of aspirin to reduce stroke risk in AF 32 or coagulation biomarkers among patients with AF 33 . Further, AF was not only related to ischemic stroke (particularly cardioembolic stroke), but also to possibly hemorrhagic stroke (subdural hemorrhage, and perhaps intracerebral hemorrhage) in this MR study. AF is known to increase cardioembolic stroke, considering abnormal atrial contraction may result in blood stasis within the left atrium and hence embolism, which could subsequently translocate to the brain 34 . For stroke prevention, anticoagulant prescription is recommended among AF patients, but control within the therapeutic range is essential to reduce thrombotic risk when under-dose or bleeding risk when overdose 35 . Vitamin K antagonists (e.g. warfarin) have been shown to promote hemorrhagic stroke, rather than other anticoagulants (e.g. direct factor Xa inhibitor) 36,37 . As such, AF may also increase risk of subdural and possibly intracerebral hemorrhage due to anticoagulant elicited bleeding risk. In addition, AF was unrelated to IHD or T2DM in our study, indicating atherosclerosis or hyperglycemia may be unlikely to be involved. The specific association of AF with ischemic stroke but not IHD lends credence to the relevance of coagulation rather than atherosclerosis, coherent with few coagulation factors being linked with IHD in a two-sample MR study 38 . Despite the lack of association of AF with IHD, AF was associated with heart failure and vice versa. AF may generate cardiomyopathy 39 and hence increase the risk of developing heart failure 40 . Conversely, heart failure may induce atrial remodeling and hence predispose to AF 41 . Considering the pathophysiology of AF is complex and multifactorial, this first MR study adds etiological insights for further examination so as to dissect out causes and consequences of AF. www.nature.com/scientificreports/ Some limitations have to be considered. First, we obtained independent SNPs reaching genome-wide significance that predict AF from the largest GWAS, with F-statistic greater than 10 and sufficient power to detect small effect sizes on various diseases (OR ranging from 1.10 to 2.77). Conversely, unlike SNPs that predicted IHD, any and ischemic stroke, T2DM, cognitive function, and late-onset AD, fewer SNPs that predicted rarer outcomes including hemorrhagic stroke, arterial thromboembolism and heart failure were available based on the relatively smaller number of participants having these conditions from the UK Biobank whose participants were generally healthier than the underlying population 42 . However, our F-statistic for all these SNPs > 10 indicates low possibility of weak instrument bias. Second, considering the pathophysiology between AF and CVD, T2DM and late-onset AD remains to be elucidated, we did not exclude some possible pleiotropic effects because they may arguably be potential mediators (e.g. BMI) such that removing these SNPs might not produce robust causal estimates. Nonetheless, the null findings remain similar before and after considering any statistical evidence of pleiotropic outliers based on the weighted median, MR-Egger and MR-PRESSO 43 . Third, we used several sensitivity analyses to validate the IVW results. The positive association of any and ischemic stroke with AF using IVW was in the opposite direction using MR-Egger. MR-Egger is sensitive to outliers that may reverse the sign of the estimates especially the limited number of SNPs 44 , whereas the weighted median and MR-PRESSO methods, which take outliers into account, obtained results that were more consistent with the IVW approach. Fourth, selection bias might bias estimates concerning late onset conditions that share etiology with common conditions that cause death prior to recruitment 45 , possibly attenuating estimates for late-onset conditions, so that estimates may be conservative. Fifth, sample overlap in two-sample MR might bias in the direction of confounding, particularly for weak instruments 46 . However, most instruments had acceptable F-statistics ranging from 22.1 to 163.6. We also used the largest available non-overlapping GWAS for IHD. Sixth, the AF GWAS does not account for medication usage. AF patients with higher risk of developing stroke may be prescribed anticoagulants, however www.nature.com/scientificreports/ medication use does not confound genetic associations although not adjusting for medication use might impair precision. Seventh, although this study showed no association of AF with cognitive function or late-onset AD, we cannot rule out that AF may affect cognitive decline, which warrants further investigation. Finally, the applicability of our findings based on largely on people of European descent to other populations including Chinese needs further investigation, considering the relevance of a causal factor may vary by setting. From a clinical perspective, our findings suggest AF may predispose to both ischemic (particularly cardioembolic stroke) and possibly hemorrhagic stroke (subdural hemorrhage), arterial thromboembolism, and heart failure. These findings emphasize the importance of addressing the current underutilization of effective treatment for AF (anticoagulants) 47 , which could help prevent serious cardiovascular complications. Further, IHD, heart failure and possibly ischemic stroke (particularly cardioembolic stroke) could predispose to AF, with corresponding implications for management. The null finding of AF on late-onset AD casts doubt on the relevance of AF treatment in late-onset AD. Together with the potential of technological innovation in early detection of often asymptomatic AF 48,49 , better understanding of AF etiology will contribute to tackling the growing societal burden of AF in a more targeted and effective manner.

Methods
Data source. Association of genetically predicted AF with IHD, stroke, other CVD subtypes, T2DM, cognitive function, and late-onset AD. We obtained genetic predictors of AF from the largest genome-wide association study (GWAS) by Nielsen et al. (2018) (n = 60,620 cases and 970,216 controls) almost entirely in people of European descent (98.6%) adjusted for age, sex, study-specific covariates, and if available, principal components 50 . AF was identified mainly based on clinical diagnosis codes in medical records (ICD-9 and ICD-10 codes) supplemented with 12-lead electrocardiogram at the examination. Genetic associations with IHD were from CARDIo-GRAMplusC4D 1000 Genomes-based GWAS 51 , with stroke from MEGASTROKE (any, ischemic, large artery, cardioembolic, small vessel) 52 and the UK Biobank SAIGE study (hemorrhagic, subarachnoid, intracerebral, subdural) 53 , with other CVD subtypes from the UK Biobank SAIGE study (arterial thromboembolism) 53 and HERMES GWAS (heart failure) 54  Association of genetically predicted IHD, stroke, other CVD subtypes, T2DM, cognitive function, and late-onset AD with AF. Genetically predicted IHD were obtained from CARDIoGRAMplusC4D 1000 Genomes-based GWAS (n = 60,801 cases and 123,504 controls) among people primarily of European descent (77%), adjusted for genomic control 51 . Genetic predictors of any and ischemic stroke were from MEGASTROKE (any: n = 40,585 cases, ischemic: n = 34,217 cases, large artery: n = 4373 cases, cardioembolic: n = 7193 cases, small vessel: n = 5386 cases, and 406,111 controls) (mean age 67.4 years, 41.7% women from the full trans-ethnic studies including Europeans), adjusted for age, sex and study-specific covariates and corrected for genomic control 52 . Genetic predictors of hemorrhagic stroke were from the UK Biobank SAIGE study (hemorrhagic: n = 1796 cases, subarachnoid: n = 812 cases, intracerebral: n = 700 cases, subdural: n = 259 cases, and 399,017 controls) among people of British white descent, adjusted for birth year, sex, and four principal components 53 . The UK Biobank recruited 503,317 adults (94% European ancestry) intended to be aged 40 to 69 years between 2006 and 2010 42 . Genetic predictors of arterial thromboembolism were from the UK Biobank SAIGE study (n = 921 cases and 400,595 controls) 53 . Genetic predictors of heart failure were from Heart Failure Molecular Epidemiology for Therapeutic Targets (HERMES) GWAS (n = 47,309 cases and 930,014 controls) (mean age: cases 71.4 years and controls 52.4 years), adjusted for age and sex, and if available, principal components and genomic control 54 . Genetic predictors of T2DM among people of European descent only were obtained from DIAbetes Meta-ANalysis of Trans-Ethnic association studies (DIAMANTE) (n = 74,124 cases and 824,006 controls) (mean age: cases 55.2 years and controls 52.7 years; proportion of women: cases 49.6% and controls 48.0%), adjusted for study-specific covariates and principal components, and corrected for genomic control 55 . Genetic predictors of cognitive function were from Davies et al. (2018) 56 (n = 300,486) among people without stroke or dementia of European descent (age range 16-102 years), adjusted for age, sex, population stratification, and cohort-specific covariates. Genetic predictors of late-onset AD were from Jansen et al. (2019) (n = 79,145 late-onset AD cases, 47,793 proxy cases without late-onset AD but with family history of late-onset AD and 328,320 proxy controls without late-onset AD or family history of AD) among people of European descent (mean age of onset for late-onset AD and proxy cases = 74.1 years and mean age of last contact for proxy controls = 68.5 years), adjusted for sex and/or age or study-specific covariates 57  Statistical analysis. MR has three fundamental assumptions: (a) relevance requires the genetic variants are associated with the exposure; (b) independence refers to the genetic variants being unrelated to confounders of the exposure-outcome association; and (c) exclusion-restriction means the genetic variants affect the outcome only via the exposure 58 . We selected genetic predictors as single nucleotide polymorphisms (SNPs) genome-wide significantly (P < 5 × 10 −8 ) and independently (r 2 < 0.001) associated with each exposure; except a less stringent level of significance (P < 5 × 10 −6 ) for any stroke and its subtypes, and arterial thromboembolism due to the lack of genome-wide significant SNPs. Independent variants (r 2 < 0.001) were selected using the "clump_data" function of the MR-Base R package (http:// www. mrbase. org/). Non-bialleleic or indel genetic variants or those without a rs number were excluded. Proxy SNPs (r 2 ≥ 0.8) in Europeans obtained from LDLink 59 were used for any SNP unavailable for the outcome. Palindromic SNPs coded A/T or C/G were aligned on effect allele frequency, for minor allele frequency ≤ 42%. Any overlapping between SNPs for AF and SNPs for various outcomes (and vice versa) were checked; if the SNPs were identical or highly correlated (linkage disequilibrium r 2 ≥ 0.8), as sensitivity analyses, we repeated the analyses after excluding these SNPs. We computed the F-statistic for SNPs www.nature.com/scientificreports/ on the exposure to assess instrument strength (i.e., relevance). We identified any association of genetic predictors of atrial fibrillation with possible confounders (education, Townsend deprivation index, smoking, alcohol drinking, and physical activity) in the UK Biobank to assess whether the genetic predictors were unconfounded (i.e., independence). For MR estimation, as the main analysis, inverse variance weighting (IVW) with multiplicative randomeffects, which assumes balanced horizontal pleiotropy (i.e., random positive or negative pleiotropy exist with zero average pleiotropic effect), was used 60 . Compared with IVW with additive random-effects, it is less subject to weak instrument bias (by downweighing estimates from SNPs with weaker SNP-exposure associations) and is relatively robust to outliers (by assigning weights to estimates based on standard error of SNP-outcome associations) 61 . We combined SNP-specific Wald estimates (which were calculated as the genetic association with the outcome divided by genetic association with the exposure) using inverse variance weighting (IVW) with multiplicative random-effects, from which odds ratio (OR) or beta coefficients (mean differences) with 95% confidence intervals (CIs) and Cochran's Q-statistic and P-value for heterogeneity were presented 61 . As sensitivity analyses to assess horizontal pleiotropy (i.e., exclusion-restriction), we used three complementary methods: (1) A weighted median which requires at least 50% of the information from valid SNPs. (2) MR-Egger which allows all SNPs to be invalid provided that the InSIDE (Instrument Strength Independent of Direct Effect) assumption holds, from which an intercept with P < 0.05 indicates the presence of pleiotropy and a higher I 2 value indicates the 'no measurement error' assumption holds 62 . (3) Mendelian Randomization Pleiotropy RESidual Sum and Outlier (MR-PRESSO) identifies potentially pleiotropic outliers and provides estimates after excluding these outlier SNPs 43 . Power calculations were performed to estimate the minimally detectable effect size for MR of AF with a power of 80% and a significance level of 0.05 63 .
For assessing the bi-directional associations of AF with cardiovascular outcomes, to adjust for multiple comparisons, a Bonferroni-corrected significance level of 0.0031 (0.05/16) was considered to account for testing inter-relationships between AF and eight traits (i.e., IHD, ischemic stroke, hemorrhagic stroke, arterial thromboembolism, heart failure, T2DM, cognitive function, and late-onset AD).
We conducted statistical analyses using R version 4.0.1 (R Foundation for Statistical Computing) with the MendelianRandomization and MRPRESSO R packages unless specified.
Ethics approval. Each study has been specifically approved by the Ethical Committees of the original studies and all the participants provided a written informed consent. This analysis of publicly available summary data does not require ethical approval. Procedures were performed in accordance with the Declaration of Helsinki.

Data availability
Data are available in public, open access repositories with the repository names and/or URL listed in