Associations of risk factor burden and genetic predisposition with the 10-year risk of atrial fibrillation: observations from a large prospective study of 348,904 participants

Data from the UK Biobank (Application Number: 69550) were applied for the present analysis. We included all participants aged 40–69 years enrolled in 2006 to 2010 with complete baseline assessments, including information on demographic and socioeconomic factors. The detailed UK Biobank protocol is available elsewhere [19].

Participants who were diagnosed with AF at baseline (n=6748), had missing data for the risk factors of interest (n=46,400, mainly due to missing blood pressure and glycated hemoglobin), missing quality-controlled genotyping data (n=95,175), or withdrew from the study (n=1298) were excluded from the current analysis. The resulting sample included 352,804 participants (Fig. 1). The detailed information on quality-controlled genotyping data was presented in Additional file 1: Text S1.

Participants were divided into three categories: index ages 45, 55, and 65 years, according to the ages at assessment: 40–49 (n=84,506), 50–59 (n=117,520), and 60–69 (n=147,178) years. For example, at index age 55 years, the criteria for selecting the participants were that they did not develop AF and were still alive before 55 years old. If participants were recruited between 50 and 55 years old or between 55 and 59 years old, the follow-up times of those started at 55 years or at the specific age when participant attained, respectively. The same approach was applied for participants with index ages 45 years and 65 years.

Participants free of AF were followed up from the later dates of their index age until the first AF occurrence, death, loss to follow-up, end of the ten-year follow-up, or March 31, 2021, whichever occurred first.

The outcome of the present study was the incidence of AF including both atrial fibrillation and flutter. Physician diagnosis of incident AF occurring during the 10-year follow-up was identified through the primary care system, hospital inpatient records, and death registry [19]. The date of death was ascertained by linking to death registries of the National Health Service Information Centre. Ascertainment of AF in the UK Biobank (ICD code I48) was mainly extracted from the “first occurrence” of health outcomes, which incorporates the diagnoses of AF at different time points from recruitment. We also collected the cases directly from the following codes: (1) non-cancer illness code, self-reported (1471, 1483); (2) Operation code (1524), (3) diagnoses – main/secondary ICD10 (I48, I48.0–4, I48.9); (4) underlying (primary/secondary) cause of death: ICD10 (I48, I48.0–4, I48.9); (5) diagnoses – main/secondary ICD9 (4273); (6) operative procedures – main/secondary OPCS (K57.1, K62.1–4).

In the current study, risk factor burden for each participant was constructed using a series of risk factors for AF, which was in accordance with the widely used CHARGE-AF risk score and previous literature [7, 20, 21]. The risk factors under analysis include BMI, alcohol consumption, smoking status, blood pressure, diabetes mellitus, and history of myocardial infarction or heart failure. Each risk factor was classified into two (optimal and elevated) or three (elevated, borderline, or optimal) categories, as described in Table 1. Details of the assessment of covariates are shown in Additional file 1: Text S2 [22]. An overall risk factor burden was calculated and then categorized into three levels: optimal (all risk factors were optimal), borderline (the borderlines were presented but no elevated factors were present), and elevated (any one of the risk factors was elevated) [20].

Furthermore, a risk factor profile was constructed to reflect the risk factor burden with different numbers of borderline/elevated risk factors, which was categorized into 7 groups: “0/0,” “0/1,” “0/2,” “0/≥3,” “1/any,” “2/any,” and “≥3/any.”

To estimate the genetic predisposition of AF, we constructed a polygenic risk score for each participant, which was derived from the current optimal genetic risk variant list from the meta-analyses excluding the UK Biobank participants (n=165 SNPs, Additional file 1: Text S3, and Table S1) [15]. Briefly, most of the genome-wide significant risk variants for AF fall in genes that cause serious heart defects in humans (e.g., PITX2, TBX5) or near genes important for striated muscle function and integrity (e.g., CFL2, MYH7), which are crucial for the function of cardiac ion channels and calcium signaling. According to the number of risk alleles, we used imputed data to calculate the PRS through multiplying by the regression coefficient obtained from the previous study [23]: PRS = (β₁ × SNP₁ + β₂ × SNP₂ + … + β₁₆₅ × SNP₁₆₅). Furthermore, we classified each participant into three categories: low (lowest quartile), intermediate (mid two quartiles), and high (highest quartile) genetic AF risk groups.

Baseline characteristics including risk factors and PRS were reported as a frequency (proportion) and were compared across different index ages using the chi-square test. To investigate the potential interactions between risk factor burden and PRS on both additive and multiplicative scales within strata of index age and sex, a new term with nine categories representing nine combinations (3×3) of risk factor burden levels and PRS levels was created.

To determine an additive interaction, both the attributable proportion (AP) and relative excess risks due to interaction (RERI), and their 95% confidence intervals (CIs) with the reference group of optimal risk factor burden/profiles and low PRS were calculated, and these measures represented the combined excess risk in both exposed groups [24]. The P-values for the additive scale in each index age were adjusted by FDR (false discovery rate). For the multiplicative scale, likelihood tests were performed to compare the model with and without the multiplicative interaction term.

The 10-year risk for incident AF from index ages of 45, 55, and 65 years up to a maximum of 10 years was calculated for each risk factor burden/profile and each risk factor separately, and then we performed subgroup analyses stratified by sex and PRS. To account for the competing death risk and to avoid inflating the cumulative incidence, a modified Kaplan-Meier estimate with age as the time scale was used to compute the 10-year risks of AF and associated 95% CIs [25]. Z ratio tests using low level as the reference were compared to the 10-year risks in the elevated and borderline risk groups.

Multivariable Fine and Gray models for each index age were fitted to predict the 10-year risk of AF (full details in Additional file 1: Text S4) [26, 27]. Calibration plots and C-index were used to evaluate the performance of the models. The predicted 10-year risks of AF for 16 risk profiles were presented in different sexes and PRS separately. All analyses were conducted using SAS software, version 9.4 (SAS Institute), and R software 4.1.1 (R Development Core Team, Vienna, Austria).

A total of 348,904 participants were included in statistical analysis defined by index ages of 45 (n=84,206), 55 (n=117,520), and 65 (n=147,178) years (Table 2 and Fig. 1). Among participants at index ages 45, 55, and 65 years, 924 (1.1%), 3883 (3.3%), and 12,924 (8.8%) developed AF during follow-up, respectively. Within each respective group, 45.6%, 43.6%, and 46.9% were men and 90.1%, 94.2%, and 96.9% were white. As index age increased, the proportion of participants that reported smoking decreased, while blood pressure, prevalent diabetes, and history of MI or HF increased. Elevated blood pressure was the most common risk factor (33.8% at age 45 years, 52.5% at 55 years, and 70.9% at 65 years). The PRS of AF ranged from 7.10 to 11.67 with a median of 9.29. Differences in baseline characteristics between low, intermediate, and high PRS at different index ages are presented in Additional file 1: Table S2-4.

The overall 10-year risks of AF were 0.67% (95% CI: 0.61% to 0.73%) for index age 45 years, 2.05% (95% CI: 1.96% to 2.13%) for index age 55 years, and 6.34% (95% CI: 6.21% to 6.46%) for index age 65 years. The 10-year risk of AF increased gradually at each index age with an increase in each single risk factor profile. History of MI or HF had the strongest association with a 10-year risk of AF at each index age, followed by diabetes mellitus and alcohol consumption (Additional file 1: Table S5).

Table 3, Fig. 2, and Additional file 2: Figure S1-S2 show the 10-year risks of AF for index ages 45, 55, and 65 years, respectively, with the elevated, borderline, or optimal risk factor burden. The 10-year risk was the lowest for the participants with optimal factors, whereas the risk was higher for participants with elevated factors for index age 45, 55, and 65 years (Table 3). The 10-year AF risks were all higher in men than in women (Additional file 1: Table S6 & S7).

Table 4 showed the combined health impact of risk factor burden and PRS on AF incidence. At each index age, compared to those with optimal risk factor burden or low PRS, participants with higher risk factor burden or higher PRS generally had a higher risk of AF. Furthermore, compared with those who had optimal risk factor burden and low PRS, participants with an elevated risk factor burden and high PRS had the highest risk of developing AF at index age 45 years (HR: 7.32; 95% CI: 3.02 to 17.70), 55 years (HR: 8.41; 95% CI: 4.37 to 16.20), and 65 years (HR: 4.12; 95% CI: 2.76 to 6.14).

On an additive scale, positive interactions were observed for risk factor burden (optimal versus elevated levels) with the PRS at each index age (P_{additive-scale} < 0.05, Table 4, Additional file 1: Table S8). For example, for participants at index age 55 with elevated risk factor burden and high PRS, the RERI and AP were 2.44 and 0.29, suggesting that a 2.44 relative excess risk was due to the additive interaction. This accounted for 29% of the risk of AF in participants who had elevated risk factor burden and high PRS. The values of RERIs and APs tended to decrease as index age increased in participants with high PRS and risk factor burden. On a multiplicative scale, significant interactions were found between risk factor burden and PRS in relation to AF risk at index age 65 years (P_{multiplicative-scale} < 0.05). Similar results were also observed in the subgroup of sex (Additional file 1: Table S9 and S10).

Among participants without AF at age index age 45, 55, and 65 years, those in optimal risk burdens and low PRS had 10-year risks of AF of 0.26%, 0.43%, and 2.27% separately, whereas those in the elevated risk burdens and high PRS had 10-year risks of AF of 1.65%, 4.22%, and 11.10%, respectively (Table 3).

For individuals at index ages 55 and 65 with optimal risk factor burden and high PRS of AF, the 10-year risks were 2.10% and 6.02%, with values higher than those in other elevated clinical risk factor burden strata at the same index age. At the index age of 45 years, compared with the participants with the optimal burden and high PRS, participants with an elevated risk factor burden and low/intermediate PRS had higher 10-year AF risk. At the same strata of risk factor burden, gradients of increased AF risk were observed with increasing risk of PRS. Stratification analyses by sex were conducted, yielding results that were generally consistent with our main findings (Additional file 1: Table S6 & S7).

Additional file 1: Table S11 shows the distributions of risk factor profiles (number of elevated/borderline risk factors) at age index age 45, 55, and 65 years. The participants with optimal risk factors decreased from 9.4% to 4.9% and 2.0% with the index age changed from 45 to 55 and 65 years, respectively.

As age increased, the number of participants with one or more borderline risk factors decreased, and gradually transitioned to having one to three elevated risk factors. Significant interactions were observed between risk factor profile and PRS in relation to AF risk (Additional file 1: Table S12). Table 5 and Additional file 1: Tables S13 and S14 show the 10-year risk among the participants with elevated or multiple borderline risk factors when the participants were separated into low, intermediate, and high PRS, respectively. The results were similar to the 10-year risks associated with risk factor burden. Especially, the 10-year risks were higher in men with optimal risk factor profiles and high PRS compared to those with elevated risk factor profiles (≥3/any) and low PRS at the index age of 65 years. In contrast, the opposite results were observed in the index age 45 years.

Additional file 1: Table S15-S17 presented three multivariable predictive models of the 10-year risk of AF at index ages 45, 55, and 65 years. The 10-year risk of AF is shown in Fig. 3 and Additional file 2: Figure S3-S6 for 16 different risk profiles for both men and women.

Higher predicted 10-year risks of AF were observed in men and in those with high PRS. Participants with high PRS, history of MI or HF, elevated alcohol consumption, a BMI of 35 or higher, and hypertension treatment had the greatest 10-year risks in both men and women. The C-indexes of the predicted model at index ages 45, 55, and 65 years were 73.5% (95% CI: 71.7% to 75.2%), 71.0% (95% CI: 70.1% to 71.9%), and 67.1% (95% CI: 66.2% to 68.1%), respectively. Additional file 2: Figures S7-S9 show the calibration plots of the predictive models at index ages 45, 55, and 65 years.

The Framingham Heart Study estimated short-term and lifetime AF risk among 9764 participants without AF at ages 55 to 75 years [21]. Though instrumental in value, cohort studies are needed to further refine the joint effect of both polygenic risk and other risk factors. Our updated 10-year risks, estimated by the contribution of both risk factors burden and genetic predisposition, developed from a substantially larger prospective cohort study, offers superior generalizability to contemporary populations and helps identify high-risk populations [10].

While previous literature has reported that the 10-year risk of AF is 2.0%, 4.5%, and 7.6% among participants in the index age of 55 years with low, intermediate, and high polygenic risk [21], the observed lower 10-year risk in our study may be attributable to a relatively healthier cohort in UK Biobank. Furthermore, the incidence of AF increases rapidly after the age of 40, and the 10-year AF risk varies with age [7].

Previous studies have largely focused on middle-aged and older adults (aged ≥ 55 years old) to investigate the 10-year risk of AF, and it is uncertain of the applicability to younger individuals (age ≤ 55). Indeed, prospectively identifying these relatively young individuals at risk for AF might be beneficial, allowing for early intervention measures [28]. Our study included younger participants (index age 45 years) and observed that the joint effect of elevated risk factors burden/profiles and high inherited predisposition on AF incidence declined with advancing age. Furthermore, among the younger age group, compared to genetic predisposition, risk factor burden/profiles have a larger effect on AF development. Our results suggest that a more favorable risk burden/profile might delay the onset of AF risk, especially among young people and men. Similar to a previous study [21], participants with high polygenic risk had twice the 10-year risk than those with low PRS, underscoring the independent impact of an inherited predisposition to AF risk.

Three AF multivariable predictive models had moderate discrimination (C-index 67.1% to 73.5%) but suboptimal calibration. The performances of our models were comparable to other models reported in prior studies [20, 29]. In our predictive models, the strongest risk factors were sex, history of MI or HF, and PRS. Both the 10-year and predicted risks of AF were higher in men than women, as previously noted [30]. One possible explanation might be that women had more favorable risk burdens compared to men. Furthermore, considering the main outcomes in the present study, which include both AF and flutter, risk factor burden and genetic predisposition might have an independent prognostic impact in terms of different types of AF [31].

Compared to the performances of previously used models, we included ethnicity as an important covariate in the AF prediction models due to the accumulated evidence that white individuals have a higher AF risk than those who are not white [32]. Furthermore, the preponderance of AF cases in our population (56%, 71%, and 82% in the index age of 45, 55, and 65 years, respectively) developed in participants with a history of MI, HF, hypertension, or diabetes, suggesting that comorbidities contributed more to AF risk in older individuals. Additionally, lifestyle risk factors including BMI, alcohol consumption, and smoking, which are modifiable, also played important roles in 10-year risk and our predictive models. Alcohol consumption led to the highest 10-year risk of AF among lifestyle risk factors. There is growing evidence on the adverse effects of alcohol on left atrial function, electrical remolding, and structure. These effects would contribute to AF [33]. Considering the close association of risk factors with the risk of AF, these modifiable factors may be better targets for AF prevention [34].

Furthermore, we hypothesize that our prediction models could be used to further inform targeted screening of individuals at risk, which may be more cost-effective than routinely screening all patients aged ≥65 years [35, 36]. The risk factor burden of AF is an easily interpretable and accessible tool of structured management, which can be readily measured with existing and future data. Moreover, with the increasing ubiquity of genomic data, both in the clinical setting and via direct-to-consumer testing, the current optimal variant list of a PRS for AF could potentially provide a more accurate personalized risk assessment [15, 21]. For individuals at an index age of 45, 55, and 65 with elevated risk factor burden and high PRS, the 10-year risks of AF were 1.65%, 4.22%, and 11.10%, respectively, which could potentially be used as thresholds. The application of the prediction models could be used to inform thresholds for screening in individuals with high predicted 10-year risks, potentially leading to earlier detection, guidance on preventive approaches, and individualized counseling.

The strengths of this study include the uniquely large sample size, including younger participants; long-term follow-up; detailed information on behavioral, clinical profile, and genetic data; the large number of AF cases included in the analysis; and the estimation of 10-year risks incorporating both risk burden/profiles and genetic predisposition with adjustment for competing risks. Prediction models and potential thresholds were developed in our study to help implement early and targeted interventions.

Our study also has some limitations. Importantly, biomarkers, electrocardiographic data, anemia, electrolyte imbalance, and other potential risk factors for AF were not included in our study [37]. Nonetheless, some studies have pointed out that electrocardiogram data may not improve the performance of predictive models [8, 38]. Some AF events caused by acute events, such as trauma, surgery, etc., may have been missed since the majority of events were ascertained using hospital inpatient data, thus the true risk of AF might be underestimated since undiagnosed AF is common [3]. In addition, measurements of risk factor burden/profiles were attained at baseline and could not reflect changes over time. Participants recruited in the UK Biobank are healthier and with a higher socioeconomic status than the general UK population, which might result in healthy volunteer bias [39]. Given that an adequate number of participants with various levels of exposure were evaluated with high internal validity, this might not have an impact on the valid estimates of relationships [39]. The UK Biobank does not record close relationships or not release the kinship coefficients of all participants estimated from genetic data, which prevents us from implementing mixed-effects models that account for the relatedness among the samples. Finally, our study used only one cohort to predict 10-year risks and predictive models of AF, and whether our findings can be generalized to other ethnic groups needs further investigation.