Predicting atrial fibrillation and its complications

Abstract

Atrial fibrillation (AF) is a common cardiac arrhythmia associated with an increased risk of stroke and other complications. Identifying individuals at higher risk of developing AF in the community is now possible using validated predictive models that take into account clinical variables and circulating biomarkers. These models have shown adequate performance in racially and ethnically diverse populations. Similarly, risk stratification schemes predict incidence of ischemic stroke in persons with AF, assisting clinicians and patients in decisions regarding oral anticoagulation use. Complementary schemes have been developed to predict the risk of bleeding in AF patients taking vitamin K antagonists. However, major gaps in our ability to predict AF and its complications exist. Additional research should refine models for AF prediction and determine their value to improve population health and clinical outcomes, advance our ability to predict stroke and other complications in AF patients, and develop predictive models for bleeding events and other adverse effects in patients using non-vitamin K oral anticoagulants.

Atrial fibrillation (AF) is a common cardiac arrhythmia affecting >30 million persons worldwide.¹ In addition, AF increases the risk of stroke, heart failure, and overall mortality.² With the aging of the population, the incidence and prevalence of AF are expected to grow in future decades as well as the burden from its associated complications.³ The growing public health significance of AF has spurred efforts to identify individuals at higher risk of developing this arrhythmia and its complications. Identifying individuals more likely to develop AF could facilitate targeting of preventive interventions and screening programs, while risk stratification schemes in AF patients can assist clinicians and patients in treatment decisions.

This review article provides an overview of the expanding field of AF prediction. We have structured the text as follows. First, we describe available models for the prediction of incident AF in the community and their potential applications. Second, we summarize the main existing risk schemes for prediction of ischemic stroke in patients with AF, highlighting some limitations. Third, we discuss prediction of bleeding and other complications in patients with AF. We end by highlighting areas that demand additional research. To inform the review, we searched PubMed through December 31, 2015 using the search terms “atrial fibrillation” AND “prediction”. We considered publications mostly from the past 5 years, though we did not exclude frequently referenced older publications, and selected those considered relevant. Though related to the topic, we do not discuss the extensive work on prediction of AF recurrence after electrical cardioversion or catheter ablation, or on AF prediction in the context of specific clinical contexts, such as postoperative AF after cardiac surgery.

PREDICTION OF AF IN THE COMMUNITY

Risk scores and equations for AF prediction

Over the last few years, several risk scores and equations for the prediction of AF in the general population have been developed, published, and validated. Table 1 enumerates in chronological order the published scores, the variables included, the characteristics of the derivation and validation samples, if any, and the performance of the model (discrimination and calibration). Discrimination refers to the ability of the model to separate subjects who develop the outcome from those who do not, while calibration refers to the agreement between observed outcomes and predictions.⁴

Table 1.

Risk scores and equations for the prediction of atrial fibrillation in the community

Risk model	Variables	Derivation	Performance	Validation in external populations	Performance
FHS (10-year risk) 5	Age, sex, body mass index, systolic blood pressure, treatment for hypertension, PR interval, cardiac murmur, heart failure	4764 participants, 100% white, 55% women, 45–95 years of age, mean age 61	C-statistic (95%CI): 0.78 (0.76–0.80) χ2 = 4.2 (p = 0.09)	AGES:6 4238 participants, 100% white, 63% women, mean age 76	C-statistic (95%CI): 0.67 (0.64, 0.71) Recalibrated χ2 = 16.2 (p = 0.06)
				CHS:6 5410 participants, 16% African-American, 84% white, 60% women, 65 and older, mean age 75	C-statistic (95%CI): 0.68 (0.66, 0.70) in whites, 0.66 (0.61, 0.71) in African Americans Recalibrated χ2 = 46.1 (p < 0.001) in whites and 10.6 (p = 0.31) in African Americans
				ARIC:7 14,546 participants, 27% African-American, 73% white, 55% women, 45–64 years of age	C-statistic: 0.68 overall, 0.69 in whites, 0.65 in African Americans
				MESA:8 6663 participants, 38% white, 28% African-American, 22% Hispanic, 12% Chinese-American, 53% women, 45–84 years of age, mean age 62	C-statistic (95%CI): 0.75 (0.72, 0.77) overall, 0.75 (0.72, 0.78) in whites, 0.74 (0.70, 0.78 in non-whites χ2 = 57.4 (p < 0.001) overall, χ2 = 8.1 (p = 0.53) in whites, χ2 = 73.9 (p <0.001) in non-whites
ARIC (10-year risk) 7	Age, race, height, smoking, systolic blood pressure, treatment for hypertension, cardiac murmur, ECG-based left ventricular hypertrophy, ECG-based left atrial enlargement, diabetes, coronary heart disease, heart failure	14,546 participants, 27% African-American, 73% white, 55% women, 45–64 years of age	C-statistic: 0.78 χ2 = 10.0 (p = 0.35)	None
WHS (10-year risk) 11	Age, weight, height, systolic blood pressure, alcohol use, smoking	19,940 participants, 100% white, 100% women, median age 53	C-statistic (95%CI): 0.72 (0.68–0.75) χ2 = 8.1 (p = 0.43)	None
CHARGE-AF (5-year risk) 12		18,556 participants, 81% white, 19% African-American, 57% women, 46–94 years of age, mean age 65	C-statistic (95%CI): 0.77 (0.75–0.78) χ2 = 9.3 (p = 0.41)	AGES:12 4469 participants, 100% white, 60% women, mean age 76	C-statistic (95%CI): 0.66 (0.63, 0.70) Recalibrated χ2 = 12.6 (p = 0.18)
				Rotterdam Study:12 3203 participants, 100% white, 59% women, mean age 72	C-statistic (95%CI): 0.71 (0.66, 0.75) Recalibrated χ2 = 16.4 (p = 0.06)
				EPIC-Norfolk:13 24,020 participants, >99% white, 55% women, 39–79 years of age, mean age 59	C-statistic (95%CI): 0.81 (0.75, 0.85) χ2 = 142.2 (p < 0.001) Recalibrated χ2 = 13.3 (p = 0.15)
				MESA:8 6663 participants, 38% white, 28% African-American, 22% Hispanic, 12% Chinese-American, 53% women, 45–84 years of age, mean age 62	C-statistic (95%CI): 0.78 (0.74, 0.81) overall, 0.76 (0.72, 0.81) in whites, 0.78 (0.72, 0.83 in non-whites χ2 = 25.6 (p = 0.002) overall, χ2 = 14.6 (p = 0.10) in whites, χ2 = 12.3 (p = 0.20) in non-whites

AGES: Age, Gene/Environment Susceptibility-Reykjavik Study; ARIC: Atherosclerosis Risk in Communities Study; CHARGE-AF: Cohorts for Aging and Research in Genomic Epidemiology—Atrial Fibrillation; CHS: Cardiovascular Health Study; CI: Confidence interval; EPIC: European Prospective Investigation into Cancer and Nutrition; FHS: Framingham Heart Study; MESA: Multi-Ethnic Study of Atherosclerosis; WHS: Women’s Health Study

The first published risk score was derived in 4764 mostly white participants in the Framingham Heart Study (FHS), and used basic demographic and clinical variables to predict the 10-year risk of AF.⁵ The discrimination of the model, assessed with the C-statistic, was good (0.78, 95% confidence interval (CI) 0.76, 0.80). This score was subsequently validated in four different cohorts: the Age, Gene/Environment Susceptibility-Reykjavik (AGES) study, Atherosclerosis Risk in Communities (ARIC) study, Cardiovascular Health Study (CHS), and the Multi-Ethnic Study of Atherosclerosis (MESA).^6–8 In these external cohorts, the discrimination of the model was acceptable, ranging from 0.67 in African American participants in CHS to 0.75 in the racially diverse MESA cohort. In most populations, however, the model required recalibration to adjust the predicted probabilities to the actual risk of AF in the different cohorts. Independently, the ARIC study also developed a 10-year risk score for AF prediction among 14,546 study participants 45–64 years of age.⁷ In contrast to the FHS AF risk score, the ARIC model was based on a bi-racial cohort, including whites and African Americans. Given the well-established lower risk of AF among non-whites compared to whites,^{9, 10} attention to race in AF prediction is relevant and the application of scores developed in a specific racial/ethnic group to another should be done carefully. The discrimination of the ARIC model was similar to the FHS AF risk score (C-statistic 0.78). The ARIC model, however, has not been applied in any external cohorts and, therefore, its validity outside the ARIC population is uncertain. More recently, the Women’s Health Study (WHS), a cohort of mostly white, healthy women, derived and validated a 10-year model among 19,940 participants.¹¹ The model had good discrimination (C-statistic 0.72) and excellent calibration in the WHS cohort, but has not been validated in external populations and its applicability to men is unknown.

The FHS, ARIC, and WHS risk scores and predictive models were derived in single cohorts, and restricted in terms of race/ethnicity (FHS), age (ARIC), or sex (WHS), which may reduce generalizability to other populations. To address this limitation, the Cohorts for Aging and Research in Genomic Epidemiology (CHARGE)-AF consortium derived a new predictive model pooling data from 18,556 participants in the FHS, CHS and ARIC studies to predict the 5-year risk of AF. This model was then validated in 7,672 participants from the AGES and Rotterdam studies, showing acceptable discrimination.¹² The CHARGE-AF model, which included demographic and clinical information readily available in clinical settings, had good discrimination in the derivation cohorts (C-statistic 0.77, 95%CI 0.75, 0.78) and acceptable in AGES (0.66, 95% 0.63, 0.70) and the Rotterdam study (0.71, 95% 0.66, 0.75).¹² The CHARGE-AF risk model has been validated in two additional cohorts. In the European Prospective Investigation into Cancer and Nutrition (EPIC)-Norfolk study, the model had excellent discrimination (C-statistic 0.81, 95% 0.75, 0.85) but it overestimated the risk of AF, requiring recalibration.¹³ Similarly, the CHARGE-AF model had good discrimination in the MESA cohort (C-statistic 0.78, 95%CI 0.74, 0.81), but also overestimated AF risk, particularly among those with the highest observed risk.⁸

Finally, some studies have suggested that scores derived for prediction of stroke in patients with AF, such as the CHADS₂ and CHA₂DS₂-VASc, could be also applied in AF prediction.^{14, 15} In fact, most of the elements included in these scores (age, diabetes, hypertension, heart failure, vascular disease) are well established risk factors for AF. However, these scores perform worse than AF-specific predictive models (as assessed with c-statistics, for example) and are not adequately calibrated to predict AF, failing to provide estimates of actual predicted AF risk over a particular time period.⁸

AF prediction beyond clinical variables

Extensions of these models have evaluated whether information on blood biomarkers, echocardiographic and electrocardiographic (ECG) measurements, or genetic variants would improve prediction of AF beyond the information provided by clinical variables.

Blood biomarkers

The predictive value of a diverse array of circulating biomarkers, including markers of inflammation (high-sensitivity C-reactive protein, fibrinogen),^16–18 atrial overload (atrial and B-type natriuretic peptides),^16–18 myocardial ischemia (high-sensitivity troponin T and I),^{17, 19, 20} cardiac fibrosis (galectin-3),²¹ and others (soluble ST2, growth differentiation factor-15),²⁰ has been assessed in the literature. Of these, only natriuretic peptides have consistently demonstrated added predictive value beyond information on clinical variables across multiple populations. For instance, in the CHARGE-AF pooled analysis, which included five separate cohorts, B-type natriuretic peptides but not C-reactive protein helped in risk reclassification of individuals, as measured by the net reclassification index (NRI).¹⁶ Similar observations have been made in the FHS,²² the Malmö Diet and Cancer Study,¹⁸ the MESA cohort,⁸ and in the Gutenberg Health Study.¹⁷

Electrocardiography

Some of the existing AF risk scores and models include ECG-derived variables, such as the PR interval in the FHS AF score, the ARIC score, and the CHARGE-AF model, or ECG-based left ventricular hypertrophy in the ARIC score and the CHARGE-AF model.^{5, 7, 12} Their added predictive value beyond clinical variables, however, is only marginal. Other ECG measurements considered as potential predictors of AF include P wave indices. A pooled analysis of the FHS and ARIC cohorts found that even though P wave indices such as P wave duration, area, and terminal force were associated with the incidence of AF, their contribution to risk prediction on top of established risk factors was minimal.²³ Information on atrial ectopy assessed through longer term heart rhythm monitoring, however, could improve AF prediction. An analysis of 1260 participants in the CHS cohort found that information on premature atrial contractions count from 24-hour Holter monitoring led to clinically significant improvements in AF prediction beyond the information provided by the FHS AF score (C-statistic of 0.65 in the FHS AF score alone vs 0.72 after adding atrial ectopy information to the statistical model).²⁴

Imaging

Information on cardiac structure and function obtained from echocardiographic studies, such as left atrial diameter, left ventricular function, left ventricular mass, or left ventricular wall thickness, have not demonstrated benefit in the prediction of AF once demographic and clinical information is considered.^{5, 25} Whether more novel measures of echocardiography-based left atrial function (e.g. left atrial strain by speckle tracking, tissue Doppler imaging-derived atrial conduction time)^{26, 27} or other cardiac imaging modalities (e.g. periatrial epicardial adipose tissue from computerized tomography)²⁸ can be used for AF prediction remains to be determined.

Genetics

Recent research has identified several common genetic variants associated with the risk of AF.²⁹ The added value of information on these genetic variants to predict AF has been explored in at least two different populations. The WHS cohort found that a genetic risk score, calculated with information on 12 single nucleotide polymorphisms previously associated with AF, significantly improved prediction, measured with change in C-statistic and continuous NRI, beyond a clinical risk score in approximately 20,000 women: the C-statistic increased from 0.72 to 0.74, while the continuous NRI was 0.49 (95%CI 0.30–0.67).¹¹ In a similar analysis among 27,471 participants of the Malmö Diet and Cancer Study, however, a genetic risk score also based on 12 single nucleotide polymorphisms only minimally improved risk prediction (C-statistic changed from 0.735 to 0.738).³⁰ Notably, none of these analyses considered information on natriuretic peptides, which are possibly the strongest biomarkers for AF risk. Future studies should evaluate whether genetic information improves our ability to predict AF on top of clinical variables and established AF circulating biomarkers.

Applications of models for AF prediction

Available risk scores, though imperfect, may play a role in identifying individuals at higher risk of developing AF, particularly the externally validated FHS and CHARGE-AF models. A follow-up question is whether this information has any clinical or public health implications. We think of two major areas in which these scores could be useful: as aids for selection of high-risk participants to screening programs and primary prevention trials, and as benchmarks for the testing of potential novel biomarkers of AF risk.

The interest in developing screening programs for identification of asymptomatic AF is growing.³¹ AF is responsible for a substantial proportion of strokes, and in a number of cases, stroke is the first clinical manifestation of AF.³² Identifying individuals with asymptomatic AF offers a unique preventive opportunity if AF diagnosis is followed by adequate antithrombotic therapy. Restricting screening programs to individuals more likely to have AF—as identified by one of the validated risk scores—would make those programs more cost-effective. A similar rationale can be applied to the selection of participants for primary prevention trials of AF. Currently, there are no established interventions for the primary prevention of AF. Trials testing such interventions will have to be conducted in subgroups at higher risk of AF, which will lead to more efficient designs.

Validated risk scores, particularly those including circulating natriuretic peptides as predictors, can also be used as benchmarks against which novel biomarkers purported to improve AF prediction can be compared. In this era of “precision medicine,” rigorous comparisons with extensively validated risk scores are needed to avoid the hype that frequently surrounds the discovery of novel markers of disease. For example, as summarized above, adequate testing against a model including natriuretic peptides (BNP or NT-proBNP) showed that inflammatory markers such as CRP, despite being associated with increased risk of AF in observational studies, are not particularly useful in AF prediction.^{8, 16}

PREDICTING STROKE IN PERSONS WITH AF

Cardioembolic stroke is a major complication of AF. Fortunately, oral anticoagulation, either using vitamin K antagonists or the newer direct oral anticoagulants, has consistently demonstrated a reduction in the risk of ischemic stroke in persons with AF.³³ Oral anticoagulation, however, is not exempt from complications. Thus, identifying patients with a higher stroke risk, more likely to benefit from oral anticoagulation, is imperative in AF treatment. Over the last 15 years several scores and risk stratification schemes have been developed (Table 2).^34–37 Of these, the CHADS₂ and CHA₂DS₂-VASc scores have gained the most acceptance, being included in AF treatment guidelines from different professional organizations across the world.^{33, 38, 39} Predictors consistently associated with the risk of stroke in AF patients in the different risk scores include older age, female sex, prior stroke, hypertension, diabetes, and heart failure.

Table 2.

Risk stratification schemes for stroke prediction in AF patients

Risk score	CHADS2 34	CHA2DS2-VASc 35	Framingham 36	ATRIA 37
Variables	Heart failure Hypertension Age Diabetes Prior stroke / TIA	Heart failure Hypertension Age Diabetes Female sex Prior stroke / TIA Vascular disease	Age Female sex Systolic blood pressure Prior stroke / TIA Diabetes	Age Female sex Diabetes Heart failure Hypertension Proteinuria Reduced kidney function or ESRD Prior stroke

ESRD: End-stage renal disease; TIA: transient ischemic stroke

Despite current recommendations suggesting the use of the CHA₂DS₂-VASc score for stroke risk stratification in AF patients, the evidence supporting this score above others is mixed. Studies that have compared multiple risk scores in the same population have not found meaningful and consistent differences between the various scores. An analysis of 2720 patients with AF from Olmsted County, Minnesota, in the United States, tested nine different stratification schemes, reporting similar performance across the models.⁴⁰ Other studies in larger populations have reported one scoring system does better than the rest, but these differences tend not to be clinically relevant and lack consistency across populations.^{41, 42} Overall, all scores offer limited discrimination ability (c-statistic <0.7) and, in the end, decisions to use one approach over another should balance simplicity, availability of data required for the score calculation, and predictive ability.⁴³ Finally, though some published studies suggest that circulating biomarkers (e.g. creatinine, troponin, natriuretic peptides, D-dimer) may improve stroke prediction in AF patients, few of these observations have been replicated in multiple populations.^44–46

PREDICTING BLEEDING AND OTHER COMPLICATIONS IN PERSONS WITH AF

Even though oral anticoagulation in AF reduces the risk of ischemic stroke and systemic thromboembolism, this benefit is accompanied by an increased bleeding risk. Identifying individuals at higher risk of bleeding complications when using oral anticoagulants may facilitate personalized treatments. To date, at least four risk scores for the assessment of bleeding risk in patients with AF treated with vitamin K antagonists have been published (summarized in Table 3).^47–50 Variables consistently associated with increased bleeding risk in this patient population include older age, renal disease, and a history of prior bleeding, with anemia, cancer, and hypertension included in several of the scores. As with the scores used for stroke risk stratification, the existing bleeding predictive models have only moderate discrimination and do not differ significantly from each other when applied to the same population.^51–53 Of note, these scores have been developed for the prediction of bleeding among persons with AF using vitamin K antagonists. Given the different bleeding profile of the newer direct oral anticoagulants (NOACs) compared to vitamin K antagonists, future research should develop bleeding risk scores specifically for patients using NOACs.

Table 3.

Risk stratification schemes for bleeding prediction in AF patients using vitamin K antagonists

Risk score	HEMORR2HAGES 47	HAS-BLED 48	ATRIA 49	ORBIT-AF 50
Variables	Age Hepatic / renal disease Hypertension Prior bleeding Stroke Alcohol abuse Anemia Cancer Reduced platelet count / function Genetic factors Fall risk	Age Abnormal renal / liver function Hypertension Prior bleeding Stroke Drugs/alcohol Labile INR	Age Renal disease Hypertension Prior bleeding Anemia	Age Abnormal kidney function Prior bleeding Anemia Antiplatelet use (heart failure) (cancer) (COPD) (hip fracture / osteoporosis) (smoking)

INR: International Normalized Ratio

AF is associated with an increased risk of other cardiovascular complications beyond ischemic stroke, such as heart failure or myocardial infarction.⁵⁴ A limited number of published studies have developed models for the prediction of these and other endpoints in persons with AF. For example, investigators from the FHS developed a predictive model including information on age, body mass index, left ventricular hypertrophy, diabetes, cardiac murmur, and prior myocardial infarction.⁵⁵ Though this and other models can help identify individuals at a particular higher risk of AF-related complications, the clinical implications of the information provided from those models are unclear since they do not translate directly into specific changes in the management of these patients.

NEXT STEPS IN PREDICTION OF AF AND ITS COMPLICATIONS

As summarized in this review, over the last 15 years we have made great advances in our ability to predict AF and its complications. However, a number of areas require additional research.

The discriminative ability of models for prediction of AF needs improvement. Additional work should determine whether novel biomarkers, imaging techniques, and genomic markers can refine the existing predictive models. Also, current models predict 5- or 10-year risk of AF in the community. Future efforts may be directed to the development of models for the prediction of lifetime risk of AF and models that predict AF using information obtained from the electronic health records. The usefulness of these predictive models to implement preventive strategies and screening programs that lead to improved population health and clinical outcomes has not been determined.

The CHA₂DS₂-VASc, HAS-BLED and other scores are widely recommended for the risk stratification of persons with AF. However, multiple studies conducted in diverse settings have demonstrated that these scores have limited ability to separate those who will develop ischemic stroke or other AF-related complications from those who will not. In order to treat the right patients at the right time with the right therapies, additional work needs to build upon the existing scores and improve them.

Finally, with the advent of new therapies for the prevention of cardioembolic stroke in AF, such as NOACs and left atrial appendage closure devices, new models for the prediction of complications among persons receiving these novel treatments need to be developed.