The meta-analysis was performed in seven cohorts of the AFGen consortium: the Atherosclerosis Risk in Communities (ARIC) study; the Framingham Heart Study (FHS); the Multi-Ethnic Study of Atherosclerosis (MESA); the Prevention of Renal and Vascular End-stage Disease (PREVEND) study; the PROspective Study of Pravastatin in the Elderly at Risk (PROSPER) study; the Rotterdam Study (RS-I and RS-II); and the Study of Health in Pomerania (SHIP). The selection of the cohorts is visualized in Fig 1. Detailed cohort descriptions are referenced in the Supplementary Notes in S1 File.

Participants were of European ancestry, with known AF status and a resting heart rate measured at baseline with a 10-second 12-lead electrocardiogram (ECG). Individuals with prevalent AF at baseline were excluded. We did not exclude individuals with negative dromotropic medication or pacemakers. Additionally, genome-wide genotyping was performed for the common genetic variants used in the resting heart rate polygenic risk score (PRS) (S1 Table in S1 File). Written informed consent by all participants and approval of the ethical review boards of all participating cohorts were provided and specifications can be found in the referenced cohort descriptions in the supplementary data.

AF was assessed by each cohort, as referenced in the supplementary notes in S1 File. The AF status of participants was established by evidence on follow-up ECG recordings and/or by diagnostic codes for AF [International Statistical Classification of Diseases and Related Health Problems (ICD)].

Resting heart rate was measured as a continuous covariate by a resting ECG recorded during baseline evaluations. Information was collected on recognized AF risk factors, such as hypertension, diabetes, heart failure, myocardial infarction and body mass index (BMI). Age was set as age at the time of the baseline ECG. Follow-up duration was defined as the time between the baseline ECG and diagnosis of incident AF or censoring. Detailed information is described for each cohort in the Supplementary Notes in S1 File.

Genotyping methodologies within the AFGen consortium were described previously [17]. A weighted PRS was calculated for each participant, incorporating genetic variants previously associated with resting heart rate [1]. Specifically, the PRS was the sum of the dose of each effect allele multiplied by the allele effect size (beta coefficient) (listed in S1 Table in S1 File). The heart rate PRS was used as the genetic instrument in the Mendelian randomization. Pleiotropy was assessed by adjusting for heart rate in the regression analyses of resting heart rate PRS and incident AF.

Mendelian randomization analyses were stratified for instrumental variable-free resting heart rate. The instrumental variable-free heart rate is the heart rate without the effect of the heart rate PRS [18]. First, multivariable-adjusted Cox proportional hazard regressions of resting heart rate as a risk factor for incident AF were performed in each cohort. However, no time to event data was available for one of the cohorts (SHIP), for which logistic regression was performed. The odds ratio from SHIP was used as an approximation of the hazard ratio and meta-analyzed with the other hazard ratios.

Heart rate was included in the models of the regressions using linear and quadratic terms together with age, sex, eigen vector, and center if appropriate. From the beta-coefficients of the linear and quadratic terms, the hazard ratio (HR) and resting heart rate were derived for each cohort. Meta-analysis of the beta-coefficients of the linear and quadratic terms were used to derive a meta-analyzed value of the HR and resting heart rate.

The association between genetically-determined resting heart rate and incident AF was investigated by performing Mendelian randomization studies in three strata, with approximately equal numbers of individuals in the three groups of heart rate. The instrumental variable-free heart rate distribution is used to stratify heart rate and avoid an association between heart rate PRS and incident AF based on the association between the heart rate PRS and heart rate [18]. The genetic variants were tested for pleiotropic effects by performing a regression between resting heart rate PRS and incident AF, adjusted for heart rate. A significant result suggests pleiotropy. We used the ratio method of Mendelian randomization. Specifically, the causal beta is estimated as the ratio of (1) a beta-coefficient representing the association between resting heart rate PRS and incident AF to (2) a beta coefficient representing the association between the heart rate PRS and resting heart rate. The resulting beta coefficients, when exponentiated, express the causal hazard ratio of AF per bpm increase in resting heart rate (Fig 2). The standard error of the three causal beta coefficients was calculated by use of Taylor series expansion [19]. Bonferroni corrected statistical significance was defined as p < 0.017 based on testing three strata. All analyses were performed using the R package (version 3.1.6; R Foundation for Statistical Computing, Vienna, Austria). Meta-analyses were performed with the function metagen () of the meta R package. Forest plots were created with the forest () function of the metaphor R package. Cox regressions were performed with the coxph function () of the survival R package.

First, regression analyses of resting heart rate PRS and instrumental variable-free resting heart rate (β₁) and regression analyses of resting heart rate PRS and incident AF (β₂) are calculated using the following formula’s; β1=ΔRestingheartrateΔRestingheartratePRS and β2=ΔloghazardrateofIncidentAFΔRestingheartratePRS. Beta coefficients of three resting heart rate strata were calculated separately. Subsequently, β3=β2β1 was calculated for each stratum. The hazard ratio (ℯ ^β3) of each stratum indicates the causal association between resting heart rate and incident AF. Abbreviations: PRS = Polygenic risk score, HR = Hazard ratio.

A total of 38,981 individuals were studied (mean age 59±10 years, 54% women). The mean resting heart rate was 67±11 bpm on baseline ECGs. A total of 4,779 (12%) individuals developed AF during a mean follow-up period of 13±5 years. Table 1 shows baseline characteristics of participating cohorts.

Observed associations between resting heart rate and incident AF are plotted in Fig 3. The meta-analyzed association between resting heart rate and incident AF showed a U-shaped curve (p = 0.028). Thus, both lower and higher resting heart rates were positively associated with incident AF.

The heart rate PRS was associated with resting heart rate in all strata (< 65 bpm, p < 0.001; 65–75 bpm, p < 0.001; >75 bpm, p < 0.001) (S1 Fig). The association between the heart rate PRS and incident AF was not significant (<65 bpm, p = 0.027; 65–75 bpm, p = 0.017; >75 bpm, p = 0.293) (S2 Fig). Pleiotropy was assessed by performing regression analyses of resting heart rate PRS and incident AF, adjusted for resting heart rate (<65 bpm, p = 0.052; 65–75 bpm, p = 0.082; >75 bpm, p = 0.778) (S3 Fig). Thus, the effect of the resting heart rate PRS on incident AF may be completely due to resting heart rate. Moreover, absence of pleiotropy was confirmed when the three p values of S3 Fig were combined using Fishers method (p = 0.076).

Causal effects, estimated by the ratio method of Mendelian randomization (Fig 2), showed an inverse relation between resting heart rate and incident AF (HR per 5 bpm, 0.82; 95% confidence interval (CI), 0.73–0.95; p = 0.010) for heart rates <65 bpm). The other strata of instrumental variable-free resting heart rates were not causally associated with incident AF (65–75 bpm, HR 0.82, 95% CI 0.59–1.05, p = 0.13; >75 bpm, HR 1.16, 95% CI 0.95–1.34, p = 0.15) (Table 2).

Our study is one of the largest studies to specifically investigate causal evidence for the observed non-linear association between resting heart rate and incident AF. Moreover, the Mendelian randomization design is less susceptible to confounding, reverse causation, and selection bias, compared to observational analyses. However, several limitations should be discussed.

First, although the 10-second ECG is a standard method for measuring resting heart rate, unusual circumstances in combination with the short measurement time (10 seconds) can distort the resting heart rate.

Second, asymptomatic AF may have gone undetected in some individuals, leading to incorrect AF status (i.e., false negative classification). Considering the large number of participants, the effect of false-negative classification results may be limited.

Third, genetic variants associated with heart rate may have biased the heart rate PRS through (unknown) associations with AF or AF related risk factors. However, pleiotropic effects of heart rate were reduced to a minimum by adjusting the association between the heart rate PRS and incident AF for heart rate.

Fourth, information on negative dromotropic medication or pacemaker rhythm was not available for all participants. Individuals with higher heart rates may be more likely to be treated with heart rate lowering medications, which also reduces the risk of AF. Although participants with negative dromotropic medication or pacemaker rhythm were not excluded, the heart rate was still strongly associated with the constructed heart rate PRS (S1 Fig).

Fifth, the data of PROSPER showed a weaker association between heart rate and heart rate PRS, which may be due to the higher mean age of the population. Exclusion of individuals with myocardial infarction did not result in different data. However, the association between resting heart rate and heart rate PRS of all cohorts together was still significant.

Sixth, although our study is one of the largest studies to investigate causal evidence for the observed non-linear association between resting heart rate and incident AF, results could have been subjected to limited power. In the future more research is needed to confirm our conclusions.

Finally, the generalizability of our findings beyond European ancestry and the age range in the cohorts is uncertain.