Predictive Accuracy of a Clinical and Genetic Risk Model for Atrial Fibrillation

Abstract

Background:

AF risk estimation using clinical factors with or without genetic information may identify AF screening candidates more accurately than the guideline-based age threshold of ≥65 years.

Methods:

We analyzed four samples across the United States and Europe (derivation: United Kingdom (UK) Biobank; validation: FINRISK, Geisinger MyCode Initiative, and Framingham Heart Study). We estimated AF risk using the Cohorts for Heart and Aging Research in Genomic Epidemiology AF (CHARGE-AF) score and a combination of CHARGE-AF and a 1,168-variant polygenic score (Predict-AF). We compared the utility of age, CHARGE-AF, and Predict-AF for predicting 5-year AF by quantifying discrimination and calibration.

Results:

Among 543,093 individuals, 8,940 developed AF within 5 years. In the validation sets, CHARGE-AF (c-index range 0.720–0.824) and Predict-AF (0.749–0.831) had largely comparable discrimination, both favorable to continuous age (0.675–0.801). Calibration was similar using CHARGE-AF (slope range 0.67–0.87) and Predict-AF (0.65–0.83). Net reclassification improvement (NRI) using Predict-AF versus CHARGE-AF was modest (NRI range 0.024–0.057), but more favorable among individuals aged <65 years (0.062–0.11). Using Predict-AF among 99,530 individuals aged ≥65 across each sample, 70,849 had AF risk <5%, of whom 69,067 (97.5%) did not develop AF, while 28,681 had AF risk ≥5%, of whom 2,264 (7.9%) developed AF. Of 11,379 individuals aged <65 years with AF risk ≥5%, 435 (3.8%) developed AF before age 65, with roughly half (46.9%) meeting anticoagulation criteria.

Conclusions:

AF risk estimation using clinical factors may prioritize individuals for AF screening more precisely than the age threshold endorsed in current guidelines. The additional value of genetic predisposition is modest, but greatest among younger individuals.

Introduction

Atrial fibrillation (AF) is associated with substantial morbidity including an increased risk of stroke.¹ Oral anticoagulation (OAC) has been shown to reduce the incidence of AF-related strokes.² Since AF can be asymptomatic, routine screening may identify individuals with undiagnosed AF who may benefit from OAC. Certain guidelines, including those from the European Society of Cardiology³ and Cardiac Society of Australia and New Zealand,⁴ provide a class I recommendation for AF screening among individuals aged ≥65 years. In contrast, the United States Preventive Services Task Force has concluded that evidence is insufficient to recommend for or against screening with electrocardiography.⁵

Given concerns about false positives and potential downstream adverse consequences of AF screening,⁶ efficient identification of individuals at greatest risk for AF is warranted. Several clinical risk factors are associated with increased risk of AF,⁷ and recent data have demonstrated that inherited predisposition to AF is associated with AF beyond clinical risk factors.^8,9 In this study, we compared the predictive utility of a risk-guided approach as compared to an age of at least 65 years – the age threshold included in guidelines which support AF screening^3,4 – for identifying individuals at risk for incident AF. We examined clinical risk factors alone and in combination with a polygenic risk score (PRS). We leveraged a diverse array of studies from multiple nations comprising prospective population-based cohorts (UK Biobank,¹⁰ FINRISK¹¹), a prospective community-based cohort (Framingham Heart Study¹²), and a healthcare-system based cohort (Geisinger MyCode Community Health Initiative¹³) – each with comprehensive clinical and genomic data – to contrast prediction approaches and enhance generalizability of the evaluation.

Methods

Detailed methods underlying the current study are described in Supplemental Methods I–III and Supplemental Tables I–II. In compliance with the Declaration of Helsinki, subjects provided informed consent and use of study data was approved by relevant local Institutional Review Boards. Use of UK Biobank data was approved under application 17488. UK Biobank data are publicly available to qualified researchers (https://www.ukbiobank.ac.uk/). The data processing scripts used to perform the main analyses described herein can be found on a persistent GitHub repository (https://github.com/shaankhurshid/predict_af).

Results

Risk score performance for AF and stroke

Across all four cohorts, 543,093 individuals were free of prevalent AF and had complete clinical and genetic data for AF prediction. Patient flow through the UK Biobank derivation cohort is shown in Supplemental Figure I. The mean age ranged 47–59 years and females comprised 52–66%. Patient characteristics are summarized in Supplemental Table III. Characteristics of individuals excluded from the derivation cohort for missing data or prevalent AF are shown in Supplemental Table IV.

At 5 years, 8,940 individuals developed incident AF. Continuous age, CHARGE-AF, and Predict-AF were each associated with 5-year incident AF, although the within-cohort effect size was consistently highest using Predict-AF (Table 1). Discrimination using CHARGE-AF (UK Biobank c-index 0.767, 95% CI 0.762–0.772; validation cohort range 0.720–0.824) and Predict-AF (0.780, 95% CI 0.775–0.785; validation cohort range 0.749–0.831) were substantially greater than continuous age (0.719, 95% CI 0.714–0.724; validation cohort range 0.675–0.801, Table 1 and Figure 1). When compared to CHARGE-AF, Predict-AF consistently exhibited a significant but modest increase in discrimination (UK Biobank increase in c-index 0.013, 95% CI 0.011–0.015; FINRISK 0.029, 95% CI 0.018–0.041; Geisinger 0.007, 95% CI 0.005–0.009; Framingham 0.015, 95% CI 0.009–0.022). Calibration was similar using CHARGE-AF (UK Biobank calibration slope 1.10, 95% CI 1.07–1.13; validation cohort range 0.67–0.87) and Predict-AF (1.00, 95% CI 0.98–1.02; validation cohort range 0.65–0.83), and favorable relative to continuous age (1.00, 95% CI 0.97–1.03; validation cohort range 0.59–0.70, Table 1 and Supplemental Figure II). The distribution of AF risk scores stratified by AF status in the derivation set are depicted in Supplemental Figure III.

Table 1.

Association between AF risk scores and 5-year incident AF

Metric	UK Biobank	FINRISK	Geisinger	Framingham
N events/N total	7192/473681	229/10560	1336/55517	183/3335
Age
Hazard ratio (per 1-SD increase)	2.56 (2.48–2.64)	1.82 (1.60–2.08)	3.43 (3.21–3.65)	2.86 (2.38–3.45)
AIC*	183324	4148	27491	2799
R2 (95%CI)	0.028 (0.027–0.030)	0.024 (0.015–0.034)	0.073 (0.068–0.078)	0.075 (0.055–0.095)
c-index (95%CI)	0.719 (0.714–0.724)	0.675 (0.641–0.710)	0.801 (0.792–0.811)	0.750 (0.719–0.783)
calibration slope (95%CI)	1.00 (0.97–1.03)†	0.69 (0.55–0.84)	0.70 (0.66–0.73)	0.59 (0.48–0.69)
CHARGE-AF
Hazard ratio (per 1-SD increase)	3.04 (2.96–3.13)	2.22 (1.94–2.53)	4.08 (3.76–4.30)	3.31 (2.73–4.01)
AIC*	180712	4087	27171	2767
R2 (95%CI)	0.045 (0.043–0.047)	0.042 (0.029–0.055)	0.086 (0.079–0.093)	0.091 (0.069–0.11)
c-index (95%CI)	0.767 (0.762–0.772)	0.720 (0.689–0.751)	0.824 (0.813–0.835)	0.775 (0.744–0.805)
calibration slope (95%CI)	1.10 (1.07–1.13)	0.87 (0.73–1.01)	0.80 (0.76–0.83)	0.67 (0.56–0.78)
AF Polygenic Risk Score
Hazard ratio (per 1-SD increase)	1.34 (1.31–1.37)	1.48 (1.31–1.66)	1.28 (1.21–1.35)	1.49 (1.30–1.71)
AIC*	187086	4195	29081	2919
R2 (95%CI)	0.004 (0.004–0.005)	0.011 (0.004–0.018)	0.003 (0.002–0.005)	0.016 (0.004–0.027)
c-index (95%CI)	0.584 (0.577–0.590)	0.615 (0.579–0.650)	0.569 (0.554–0.584)	0.617 (0.574–0.660)
calibration slope (95%CI)	1.00 (0.93–1.07)	1.21 (0.84–1.58)	0.96 (0.76–1.16)	1.48 (0.97–1.99)
Predict-AF
Hazard ratio (per 1-SD increase)	3.21 (3.12–3.30)	2.46 (2.15–2.80)	4.19 (3.91–4.48)	3.68 (3.01–4.49)
AIC*	179857	4048	270510	2740
R2 (95%CI)	0.050 (0.048–0.053)	0.053 (0.038–0.068)	0.091 (0.085–0.097)	0.10 (0.080–0.13)
c-index (95%CI)	0.780 (0.775–0.785)‡	0.749 (0.718–0.781)‡	0.831 (0.822–0.840)‡	0.790 (0.761–0.819)‡
calibration slope (95%CI)	1.00 (0.98–1.02)†	0.83 (0.71–0.95)	0.73 (0.69–0.76)	0.65 (0.55–0.75)

^*Akaike Information Criterion, a penalized likelihood metric, where lower values indicate improved model fit after adjustment for model complexity

^†Values after bias correction using bootstrapping: 0.9997 (age), 1.0057 (PRS), 1.00027 (Predict-AF).¹⁴

^‡Difference in c-index versus CHARGE-AF: UKBB 0.013 (0.011–0.015); FINRISK 0.029 (0.018–0.041); Geisinger 0.007 (0.005–0.009); Framingham 0.015 (0.009–0.022)

Figure 1.

Discrimination of 5-year incident AF using AF risk scores versus continuous age. Depicted are time-dependent receiver operating characteristic curves for Predict-AF (red), CHARGE-AF (blue) and age (green), each as continuous variables, for 5-year incident AF. Curves higher and to the left demonstrate superior predictive performance. The area under each curve is equivalent to test discrimination. The hashed diagonal line represents a completely uninformative test.

At 5 years, 3,011 individuals developed incident stroke. Continuous age, CHARGE-AF, and Predict-AF were associated with 5-year incident stroke (Supplemental Table V). Discrimination of stroke using CHARGE-AF (UK Biobank c-index 0.704, 95% CI 0.693–0.714; validation cohort range 0.761–0.772) and Predict-AF (0.704, 95% CI 0.693–0.714; validation cohort range 0.752–0.767) were substantially greater than continuous age (0.679, 95% CI 0.668–0.689; validation cohort range 0.742–0.747).

High AF risk versus age ≥65 years

We then compared ≥5% 5-year AF risk versus age ≥65 years as a potential alternative criterion for selecting individuals for AF screening. Using Predict-AF, overall reclassification was favorable in each cohort except for the UK Biobank (UK Biobank net reclassification improvement [NRI] −0.049, 95% CI −0.053 to −0.039; validation cohort range 0.054–0.13, Table 2).

Table 2.

Net reclassification analyses using estimated 5-year AF risk ≥5% versus age ≥65 years

	UK Biobank			FINRISK			Geisinger			Framingham
		CHARGE-AF			CHARGE-AF			CHARGE-AF			CHARGE-AF
	Age	≥5% risk	<5% risk	Age	≥5% risk	<5% risk	Age	≥5% risk	<5% risk	Age	≥5% risk	<5% risk
AF events	≥65	854 (11.9%)	2294* (31.9%)	≥65	62 (27.1%)	19* (8.3%)	≥ 65	665 (49.8%)	16* (1.2%)	≥65	132 (72.1%)	4* (2.2%)
	<65	185† (2.6%)	3859 (53.7%)	<65	19† (8.3%)	129 (56.3%)	< 65	224† (16.8%)	431 (32.3%)	<65	20† (10.9%)	27 (14.8%)
AF non events	≥65	10325 (2.2%)	75099† (16.1%)	≥65	1038 (10.0%)	576† (5.6%)	≥ 65	6603 (12.2%)	674† (1.2%)	≥65	1099 (34.9%)	70† (2.2%)
	<65	1767* (0.4%)	379298 (81.3%)	<65	262* (2.5%)	8455 (81.8%)	< 65	3691* (6.8%)	43213 (79.8%)	<65	251* (8.0%)	1732 (54.9%)
	NRI+	−29.4% (−30.9 to −27.9)		NRI+	0.25% (−5.6–4.1)		NRI+	15.6% (14.4–17.5)		NRI+	8.8% (3.3–10.3)
	NRI−	15.7% (15.6–15.8)		NRI−	3.0% (2.7–3.9)		NRI−	−5.6% (−5.7 to −5.4)		NRI−	−5.7% (−7.0 to −5.0)
	NRI	−0.14 (−0.15 to −0.13)		NRI	0.033 (−0.028–0.080)		NRI	0.10 (0.086–0.12)		NRI	0.030 (−0.021–0.036)
	UK Biobank			FINRISK			Geisinger			Framingham
		Predict-AF			Predict-AF			Predict-AF			Predict-AF
	Age	≥5% risk	<5% risk	Age	≥5% risk	<5% risk	Age	≥5% risk	<5% risk	Age	≥5% risk	<5% risk
AF events	≥ 65	1400 (19.5%)	1748* (24.3%)	≥65	68 (29.7%)	13* (5.7%)	≥65	665 (49.8%)	16* (1.2%)	≥65	131 (71.6%)	5* (2.7%)
	<65	435† (6.0%)	3609 (50.2%)	<65	29† (12.7%)	119 (52.0%)	<65	288† (21.6%)	367 (27.5%)	<65	24† (13.1%)	23 (12.6%)
AF non events	≥65	17620 (3.8%)	67804† (14.5%)	≥65	1030 (10.0%)	584† (5.7%)	≥65	6694 (12.4%)	583† (1.1%)	≥65	1073 (34.0%)	96† (3.0%)
	<65	5118* (1.1%)	375947 (80.6%)	<65	428* (4.1%)	8289 (80.2%)	<65	4806* (8.9%)	42098 (77.7%)	<65	251* (8.0%)	1732 (54.9%)
	NRI+	−18.3% (−18.7 to −17.3)		NRI+	7.4% (4.9–18.0)		NRI+	20.4% (18.9–21.8)		NRI+	10.3% (7.9–12.5)
	NRI−	13.4% (13.4–13.5)		NRI−	1.5% (1.3–2.5)		NRI−	−7.8% (−7.9 to −7.6)		NRI−	−4.9% (−6.1 to −4.0)
	NRI	−0.049 (−0.053 to −0.039)		NRI	0.089 (0.062–0.20)		NRI	0.13 (0.11–0.14)		NRI	0.054 (0.029–0.076)

^*Denotes inappropriate reclassification

^†Denotes appropriate reclassification

NRI=Net reclassification improvement; NRI+=Event reclassification improvement; NRI−=Non-event reclassification improvement

NRI values calculated using the Kaplan-Meier estimator to account for censored survival data¹⁵

The predominant benefit of utilizing estimated AF risk was appropriate downclassification of AF non-events in the UK Biobank (non-event reclassification improvement [NRI-] 13.4%, 95% CI 13.4–13.5) and FINRISK (NRI- 1.5%, 95% CI 1.3–2.5), versus appropriate upclassification of AF events in Geisinger (event reclassification improvement [NRI+] 20.4%, 95% CI 18.9–21.8) and Framingham (NRI+ 10.3%, 95% CI 7.9–12.5). Of 99,530 individuals aged ≥65 years, 28,681 had predicted AF risk ≥5% using Predict-AF, of whom 2,264 (7.9%) developed AF. Of 70,849 individuals aged ≥65 years with predicted AF risk <5%, 69,067 (97.5%) did not develop AF (Table 2 and Figure 2). Of the other 1,782 individuals (2.5%) who did develop AF, 1,039 (58.3%) had a guideline-based indication for OAC. Of 11,379 individuals aged <65 with predicted AF risk ≥5%, 435 (3.8%) developed AF before prior to age 65, with roughly half (46.9%) having a guideline-based indication for OAC (Supplemental Figure IV). Among individuals with predicted AF risk ≥5%, the 5-year cumulative risk of AF was similar among those aged <65 years (8.41%, 95% CI 6.91–9.88; validation cohort range 5.65–9.09) versus ≥65 years (7.53%, 95% CI 7.15–7.91; validation cohort range 6.38–11.60). The cumulative risks of AF and stroke stratified by estimated AF risk versus age are shown in Supplemental Tables VI–IX and Supplemental Figures IV–VII.

Figure 2.

Distribution of predicted 5-year AF risk stratified by AF status among individuals aged ≥65 years. Depicted are frequency distributions of estimated 5-year AF risk among individuals with incident AF (top panels) versus without incident AF (bottom panels) for individuals aged ≥65 years (i.e., with an existing indication for AF screening).^3,4 Lighter shades depict CHARGE-AF, while darker shades depict Predict-AF. The dotted vertical line shows the 5% risk threshold. In the bottom panels, colored areas depict individuals correctly classified as low risk, in whom deferral of AF screening may be reasonable. In the top panels, colored areas depict individuals correctly classified as high risk, in whom AF screening may be most appropriate. Grayed areas represent individuals misclassified at the 5% risk threshold (i.e., had high predicted AF risk but did not develop incident AF [bottom], or had low predicted AF risk but developed incident AF [top]). Shifting the dotted vertical line along the x-axis illustrates the effect of changing the risk threshold used to identify individuals for AF screening. Values above the 95^th percentile of risk may not be depicted for graphical purposes.

A decision curve demonstrated net benefit using a risk- as opposed to age-based approach across the entire spectrum of plausible AF screening costs and benefits (Supplemental Figure VIII), and suggested that use of AF risk across a range of thresholds could result in avoidance of unnecessary AF screenings (Supplemental Figure IX). Results of exploratory analyses assessing the comparative yield of AF screening guided by risk as opposed to age are shown in Supplemental Table X and Supplemental Figure X.

Adding genetic to clinical risk

Improvement in AF discrimination using Predict-AF (UK Biobank c-index 0.761, 95% CI 0.752–0.770; validation cohort range 0.744–0.807) versus CHARGE-AF (0.745, 95% CI 0.736–0.754; validation cohort range 0.704–0.794) was more prominent among individuals aged <65 years (Supplemental Figure XI). Overall, reclassification using Predict-AF versus CHARGE-AF was modest but favorable (NRI 0.089, 95% CI 0.082–0.094; validation cohort range 0.024–0.057). Specifically, Predict-AF appropriately upclassified AF events compared to CHARGE-AF (UK Biobank NRI+ 11.1%, 95% CI 10.5–11.7; validation cohort range 1.6–7.2), but at the cost of some inappropriate upclassification of AF non-events (UK Biobank NRI- −2.3%, 95% CI −2.3 to −2.2; validation cohort range −2.2–0.83, Supplemental Table XI and Supplemental Figure XII). Favorable reclassification was also generally more prominent among individuals aged <65 years (UK Biobank NRI 0.074, 95% CI 0.052–0.10; validation cohort range 0.062–0.11), with greater appropriate upclassification of AF events (UK Biobank NRI+ 8.3%, 95% CI 6.1–11.2; validation cohort range 8.1–11.7), and similar or less inappropriate upclassification of AF non-events (UK Biobank NRI- −0.87%, 95% CI −0.90 to −0.87; validation cohort range −2.4% to −0.028%, Supplemental Table XII).

Secondary analyses

In secondary analyses, results were similar using an age threshold of ≥70 years (Supplemental Table XIII). The association between 5-year AF risk and Predict-AF was modestly stronger in women (Supplemental Table XIV). A score defined as the simple sum of CHARGE-AF and the PRS performed worse than Predict-AF (Supplemental Table XV). Discrimination using AF risk quartiles demonstrated similar performance (Supplemental Table XVI). Results of analyses restricted to individuals of European ancestry were also similar (Supplemental Table XVII). The ability of AF risk to identify individuals at higher risk for AF and stroke persisted when assessed solely among individuals with CHA₂DS₂-VASc ≥2 (male) or ≥3 (female) at the time of AF diagnosis (Supplemental Table XVIII). Analyses utilizing the optimal diagnostic threshold of AF risk using Youden’s statistic are presented in Supplemental Table XIX.

Discussion

In a half-million individuals across four independent cohorts, we observed that risk models were superior predictors of AF than age. Specifically, low estimated AF risk identified nearly 70,000 individuals aged ≥65 years who did not develop AF but who have a guideline-based indication for AF screening on the basis of age,^3,4 whereas high estimated risk identified over 400 individuals developing AF before age 65 years without a conventional indication for screening. The addition of genetic risk to clinical risk resulted in modest improvements in AF discrimination, although more substantive increases were observed among younger individuals. Overall, our results suggest that consideration of individual AF risk may improve selection of AF screening candidates as compared to the current guideline-based age threshold of 65 years.^3,4

Our results provide a contemporary benchmark of AF risk prediction performance, and suggest that estimated AF risk is consistently more accurate than age in identifying individuals at high risk for AF and stroke. Specifically, we observe that a decision to pursue AF screening using an age-based threshold alone may result in screening of a large number of individuals at low risk for AF and incorrectly result in failure to screen certain younger individuals at elevated risks of both AF and stroke. Although simple, age-based thresholds may be inefficient and expose low risk individuals above the screening threshold to false positive results, and high-risk individuals below the threshold to risks of stroke. In contrast, a risk-based threshold may offer a tradeoff in which deployment of screening may be more efficient and personalized, but at the expense of increased complexity. To this end, we anticipate that increasing availability of linked electronic health records may offset complexity by facilitating automated AF risk estimates,^16,17 and note that use of composite risk scores to guide screening is already performed for conditions such as osteoporosis.¹⁸ Practically, our evaluation demonstrates that in some study samples (e.g., UK Biobank, FINRISK), many participants meeting existing indications for AF screening are in fact at very low risk of developing AF. Using estimated AF risk to exclude such individuals may enrich the screening pool for those at highest short-term AF risk, increasing the probability that a positive screen reflects true AF, and reducing unwarranted exposure of low-risk individuals to screening – a key consideration given imperfect test characteristics of current modalities.¹⁹ Indeed, even some of the most accurate screening modalities (e.g., 12-lead ECG, surface monitors) exhibit false positive rates of up to 5%,^19,20 which may become significant at population scale.⁶

Our findings demonstrate substantial variability in the distribution of estimated AF risk across populations, suggesting that the optimal AF risk thresholds may vary based on the context of the planned intervention (e.g., screening). Indeed, the dominant pattern of reclassification using estimated AF risk versus age varied by sample. In samples with lower baseline AF risk (e.g., UK Biobank, FINRISK), the prevailing benefit of AF risk estimation was identification of low risk individuals aged ≥65 years in whom screening may be deferred. Conversely, in samples with greater AF-related comorbidity, AF risk estimation primarily facilitated identification of individuals aged <65 years at high risk for AF in whom screening may be considered. As a result, although we empirically defined high risk as 5-year AF probability ≥5%, we do not specifically advocate for a uniform risk threshold. Decision curve analyses suggest that a risk-based approach provides greater benefit than an age-based approach across a wide spectrum of AF risk thresholds. Future work is needed to determine optimal risk thresholds, which should account for patient comorbidities, accuracy of predicted AF risk in the target population, availability of financial and other resources to support screening, and the characteristics of the screening test under consideration.²¹

Importantly, AF risk can identify individuals younger than 65 years who are at high risk for both AF and stroke, although the relative yield may be limited. Use of Predict-AF identified over 400 individuals who developed AF prior to age 65 years, about half of whom had sufficiently high stroke risk to merit OAC at AF diagnosis. Since current guidelines²² do not recommend population-based AF screening in individuals aged <65 years, use of estimated AF risk may provide an opportunity to identify younger individuals at high AF risk who would otherwise not be screened. Indeed, individuals at high AF risk using Predict-AF had similar short-term AF incidence regardless of whether they were above or below age 65 years. At the same time, although reclassification indices were generally favorable, use of AF risk also resulted in thousands of younger individuals classified as high risk who did not develop AF within five years or prior to age 65. Future work is warranted to determine whether predicted AF risk coupled with a screening strategy having very high specificity can facilitate appropriate AF detection at an acceptable false positive rate among individuals aged <65 years at elevated AF risk.

Our findings suggest that genetic information does not substantively improve AF risk prediction, except potentially among younger individuals. In the current study, adding a 1,100-variant PRS (which performed identically to a 6.7 million-variant LDpred PRS in the UK Biobank) to the CHARGE-AF clinical model resulted in a modest increase in AF discrimination (0.01–0.03). Our findings are consistent with past observations suggesting small improvements in AF risk prediction using genetic scores in smaller samples.^8,9 More recently, Mars et al. demonstrated more meaningful net reclassification (0.10) using a 6.2 million-variant LDpred PRS developed using Finnish linkage disequilibrium structure and deployed in the FINRISK population,¹¹ although the improvement in c-index was again small (0.009). Importantly, we did observe greater improvements in discrimination among individuals aged <65 years (0.02–0.04), resulting in event reclassification of 8–12%. Such findings are consistent with previous observations that early-onset AF is particularly heritable.²³ Notably, utilization of genetic information to guide screening would require an efficient mechanism for genomic data ascertainment, and a harmonized platform to facilitate integration with clinical risk. Economic analyses are warranted to estimate whether ascertainment and utilization of genetic information among younger individuals at risk for AF may be cost-effective.

Our findings must be interpreted in the context of study design. First, participants were predominantly of European ancestry, which limits generalizability to other populations. Second, although we submit that improved performance for predicting clinically apparent AF would translate to a similar benefit when applied prospectively to improve detection of undiagnosed AF, we acknowledge that the clinical and genetic risk profile of undiagnosed AF is unknown and may differ. Third, although certain guidelines recommend AF screening based on age >65 or ≥65 years, some screening studies have utilized higher age thresholds which we cannot fully evaluate given the age composition of our study.^3,4,24 Fourth, given our intent to assess the value of absolute AF risk estimation to guide screening, we utilized the baseline AF rate in each sample when calculating AF probabilities.⁷ Prospective use of our approach would therefore be best informed by prior knowledge of the average AF risk in the population of interest. Fifth, we utilized a 1,168-variant pruning-and-thresholding PRS¹² in all analyses. Although recent PRS methods have suggested greater performance, we note that such analyses generally do not account for the effect of clinical risk factors on the comparison.^25–27 Indeed, our PRS had slightly lower discrimination (0.583) as compared to an LDpred PRS comprising 6.7 million variants (0.589), but both scores performed identically when added to CHARGE-AF in the derivation set. Nevertheless, we cannot rule out better performance using additional PRS methods,^25,26 or with re-derivation, tuning, or selection of PRS within each population. Sixth, given our intent to assess models stratifying risk of future AF, we excluded individuals with prevalent AF. Since early-onset AF is more likely to be prevalent, and early-onset AF may be particularly heritable, current design may limit our ability to evaluate the impact of inherited predisposition to AF among individuals at risk for AF occurring at very young ages. Seventh, we cannot establish causality or eliminate residual confounding.

In summary, in over a half-million individuals across four independent cohorts, we found AF risk exhibits superior predictive utility as compared to age, assessed either continuously or using the guideline-based threshold of ≥65 years, for identifying individuals likely to develop AF and stroke. Provided that appropriate resources and infrastructure exist, use of estimated AF risk may efficiently prioritize individuals for AF screening by identifying individuals at greatest AF risk and distinguishing individuals at low AF risk who are less likely to benefit.

Nonstandard Abbreviations and Acronyms

AF

atrial fibrillation

CHARGE-AF

Cohorts for Heart and Aging Research in Genomic Epidemiology atrial fibrillation

NRI

net reclassification improvement

NRI+

event reclassification improvement

NRI−

non-event reclassification improvement

OAC

oral anticoagulation

PRS

polygenic risk score