Estimating Vitamin K Antagonist Anticoagulation Benefit in People with Atrial Fibrillation Accounting for Competing Risks: Evidence from 12 Randomized Trials

Abstract

Background:

Patients with atrial fibrillation (AF) have a high mortality rate that is only partially attributable to vascular outcomes. The competing risk of death may affect expected anticoagulant benefit. We determined if competing risks materially affect the guideline-endorsed estimate of anticoagulant benefit.

Methods:

We conducted a secondary analysis of 12 RCTs that randomized patients with AF to Vitamin K antagonists (VKAs) or either placebo or antiplatelets. For each participant, we estimated the absolute risk reduction (ARR) of VKAs to prevent stroke or systemic embolism using two methods—first using a guideline-endorsed model (CHA₂DS₂-VASc) and then again using a Competing Risk Model that uses the same inputs as CHA₂DS₂-VASc but accounts for the competing risk of death and allows for non-linear growth in benefit. We compared the absolute and relative differences in estimated benefit and whether the differences varied by life expectancy.

Results:

7933 participants (median age 73 years, 36% women) had a median life expectancy of 8 years (IQR 6, 12), determined by comorbidity-adjusted life tables. 43% were randomized to VKAs. The CHA₂DS₂-VASc model estimated a larger ARR than the Competing Risk Model (median ARR at 3 years, 6.9% [IQR 4.7%, 10.0%] vs. 5.2% [IQR 3.5%, 7.4%], p<0.001). ARR differences varied by life expectancies: for those with life expectancies in the highest decile, 3-year ARR difference (CHA₂DS₂-VASc model – Competing Risk Model 3-year risk) was −1.3% (95%CI −1.3% to −1.2%); for those with life expectancies in the lowest decile, 3-year ARR difference was 4.7% (95% CI 4.5% to 5.0%).

Conclusion:

VKA anticoagulants were exceptionally effective at reducing stroke risk. However, VKA benefits were misestimated with CHA₂DS₂-VASc, which does not account for the competing risk of death nor decelerating treatment benefit over time. Overestimation was most pronounced when life expectancy was low and when benefit was estimated over a multi-year horizon.

INTRODUCTION

Anticoagulants are the mainstay of preventative therapy for millions of older adults with atrial fibrillation. While anticoagulants reduce the risk of ischemic stroke and systemic embolism, they also increase the risk of bleeding. To help patients and clinicians weigh the risks and benefits of treatment, clinical guidelines and decision support tools endorse using the CHA₂DS₂-VASc score to estimate a patient’s annual risk of ischemic stroke or systemic embolism without treatment.^1–5 The ACC/AHA/HRS consensus guidelines recommend that this baseline risk be used to estimate a patient’s expected absolute risk reduction by applying the relative risk reduction from a meta-analysis of randomized trials.^1,6 Guidelines recommend anticoagulant therapy when the absolute risk reduction exceeds a threshold. The premise is that treatment benefits outweigh risks when an individual’s estimated event risk exceeds this threshold.

While transparent, this approach makes two assumptions that can affect the accuracy of expected benefit. First, this approach does not account for the competing risk of death. A competing risk is an alternative outcome that occurs before, and necessarily precludes, the event of interest (e.g., cancer death before stroke from atrial fibrillation), thus limiting the absolute benefit achievable by anticoagulant treatment.^7–11 Competing risks are germane given the high all-cause mortality rate following a new diagnosis of atrial fibrillation—multiple studies estimate 20-25% mortality in the first year.^12–14 Second, this approach assumes that the therapeutic benefit continues to increase at a constant rate over time—that is, the risk of stroke over two years is twice the one-year risk and thus the absolute risk reduction over two years is twice the one-year absolute risk reduction. Both issues are readily addressed by estimating benefit using a competing risk model; however, it is unknown if doing so will materially affect absolute risk reduction estimates attributable to anticoagulants.

We used patient-level data from 12 randomized trials of vitamin K antagonists for atrial fibrillation to determine if a competing risk model affects the measurement of absolute stroke risk reduction. First, we compared the guideline-endorsed approach to measuring absolute risk reduction (i.e., CHA₂DS₂-VASc score) to a competing risk model (i.e., Fine-Gray model¹⁵). Second, we determined if differences in expected stroke risk reduction between the guideline-endorse CHA₂DS₂-VASc approach and a competing risk model varied by life expectancy.

METHODS

Study design and participants

We used patient-level data from the Atrial Fibrillation Investigators (AFI) database which contains patient-level data from 12 published clinical trials where patients were randomized to full-dose vitamin K antagonists, antiplatelets, or placebo. We focused on trials that established the efficacy of oral anticoagulants; therefore, we did not include trials that compared two different oral anticoagulants. We included the following trials: Atrial Fibrillation, Aspirin, and Anticoagulation Study 1 (AFASAK-1),¹⁶ AFASAK-2,¹⁷ Boston Area Anticoagulation Trial for Atrial Fibrillation (BAATAF),¹⁸ Birmingham Atrial Fibrillation Treatment of the Aged Study (BAFTA),¹⁹ Canadian Atrial Fibrillation Anticoagulation (CAFA),²⁰ European Atrial Fibrillation Trial (EAFT),²¹ Primary Prevention of Arterial Thromboembolism in Atrial Fibrillation (PAATAF),²² National Study for Prevention of Embolism in Atrial Fibrillation (NASPEAF),²³ the Stroke Prevention in Atrial Fibrillation 1 (SPAF-1),²⁴ SPAF-2,²⁵ SPAF-3,²⁶ and Stroke Prevention in Non-rheumatic Atrial Fibrillation (SPINAF).²⁷ We did not include patients with mitral stenosis and patients in SPAF-1, EAFT, PATAF, and SPAF-3 who were deemed ineligible to receive vitamin K antagonists (trial details in Table S1 and Table S2). Because of the sensitive nature of the data collected for this study, requests to access the dataset from qualified researchers trained in human subject confidentiality protocols may be sent to the corresponding author.

Participant characteristics

Research coordinators and physicians collected patient characteristics before therapy initiation. While specific features varied from study to study, common elements included a history of stroke or transient ischemic attack, hypertension or systolic blood pressure ≥160 mmHg or use of antihypertensives, diabetes, angina, myocardial infarction, peripheral vascular disease, smoking, and congestive heart failure, and body mass index. History of myocardial infarction was not collected in NASPEAF and peripheral vascular disease was not collected in AFASAK1, BAATAF, or BAFTA. Because history of myocardial infarction and peripheral vascular disease were missing for all participants in specific trials, we assumed they were missing at random and imputed them in 20 datasets using chained equations.²⁸ We excluded <1% of participants who were missing data otherwise collected in a given trial (Figure S1).

Treatment exposure

We examined all patients based on their treatment allocation (i.e., intention to treat). Because studies have shown that antiplatelet and low-dose warfarin are ineffective thromboprophylaxis in AF,⁶ we categorized all trial participants as being randomized to full-dose vitamin K antagonists or control. Patients randomized to placebo, antiplatelets, low-dose warfarin, or low-dose warfarin with aspirin were considered controls. AFASAK2, PAATAF, and SPAF3 used low-dose warfarin and reported a mean international normalized ratio (INR) of < 1.5 supporting their categorization as a control. While NASPEAF also had a low-dose vitamin K antagonists arm, the mean INR was 2.0 which is considered therapeutic¹; therefore, we excluded participants randomized to the NASPEAF low-dose vitamin K antagonists arm.

Outcome ascertainment

The primary outcome was ischemic stroke or systemic embolism. We detail outcome definitions by trial in Table S3. In general, trials defined ischemic stroke as a focal neurological deficit lasting >24 hours. All trials except AFASAK-1 required a CT or MRI showing the absence of blood. Systemic embolism was collected as an outcome in all but SPINAF and, by and large, defined as an embolism to internal organs or limbs and required evidence via angiography, surgery, or autopsy. Patients were evaluated at 3- to 6-month intervals or when a clinical outcome event was suspected. Except in AFASAK-1, a central committee, blinded to intervention allocation, adjudicated all clinical events.

Life expectancy

We estimated the life expectancy of each participant at the time of trial enrollment using the life table method.²⁹ We started with sex- and enrollment year-specific life tables from the U.S. Centers for Disease Control and Prevention.³⁰ These tables, generated from population data, predict annual mortality rates stratified by age, sex, and year. The life table method uses annual mortality rates to calculate life expectancy. We used the life table method to estimate each participant’s life expectancy by adjusting the annual mortality rate for the additional mortality risk associated with their comorbidities at the time of trial enrollment (Supplemental Methods 1, Table S4).

Analysis

Our first analytic goal was to determine if using a competing risk framework generates estimates of stroke risk reduction different from those of the guideline-endorsed CHA₂DS₂-VASc model. To accomplish this and speak directly to guideline-recommended practice, we estimated absolute risk reduction using the CHA₂DS₂-VASc score as the ACC/AHA/HRS guidelines recommend.^1,2 Specifically, we assigned each patient an off-treatment risk of ischemic stroke or systemic embolism corresponding to their CHA₂DS₂-VASc score. Rates come from the 2012 study by Friberg et al., which used the Swedish Atrial Fibrillation cohort to validate off-treatment thromboembolic rates corresponding to each CHA₂DS₂-VASc score.³¹ These rates are used in patient-facing decision tools, online calculators, and decision analytic models.^5,32,33 To calculate the absolute risk reduction, we multiplied the off-treatment stroke rate by 0.64, the guideline-cited efficacy of vitamin K antagonists.^1,6 Using this procedure for each patient, we estimated the annual absolute risk reduction; this precise method is endorsed by the ACC/AHA/HRS Atrial Fibrillation management guidelines to estimate benefit.¹ Because patients and physicians prefer to make anticoagulant decisions using a 1-to-5-year time horizon,⁵ we extrapolated this annual reduction over five years, accounting for the declining at-risk population (Supplemental Methods 2).^34,35 The same approach is also used in decision aids.⁵

Next, we estimated the absolute risk reduction using the Fine-Gray extension of the Cox proportional hazards model, treating death unrelated to ischemic stroke or systemic embolism as a competing event.¹¹ We fit a Fine-Gray model where time to ischemic stroke or systemic embolism is a function of age (<65 years, 65-74 years, >75years), gender, congestive heart failure, diabetes, hypertension, prior stroke or transient ischemic attack, and vascular disease stratified by randomization to vitamin K antagonists. These are the same predictors used in the CHA₂DS₂-VASc score. We verified proportional hazards assumptions by visual examination of the cumulative incidence plots and using Schoenfeld residuals. Then, we used the resulting treatment-stratified models to estimate the cumulative incidence of ischemic stroke or systemic embolism for each participant given their covariates at each study time point, assuming first they had been randomized to vitamin K antagonists and then assuming they had been randomized to control (i.e., predicted values).^36,37 The difference between the two estimates represented the ARR for a given patient at a given time point. We determined the misestimation of the CHA₂DS₂-VASc method as the difference between the ARR estimated by CHA₂DS₂-VASc and the ARR estimated by the Competing Risk Model. We used the paired t-test to determine if the two methods produced statistically different estimates of benefit at each year after randomization.

Our second analytic goal was to determine if life expectancy predicted magnitude in the CHA₂DS₂-VASc model misestimation. To achieve this, we determined the association between life expectancy and misestimation of the CHA₂DS₂-VASc method over a 3-year horizon. We chose 3 years since it is the midpoint between the 1-to-5 year horizon preferred by patients and physicians and has been used in prior anticoagulation decision analyses.^5,38 We examined misestimation of the CHA₂DS₂-VASc method by decile of life expectancy at trial enrollment, hypothesizing that the misestimation would be greater at lower life expectancies.

We performed all statistical analyses using SAS 9.4 (Cary, NC). The study protocol was approved by Institutional Review Boards at UCSF (21-34930) and MGH (2022P001783).

RESULTS

Patient characteristics and overall event rates

This study included 7933 patients from 12 randomized trials where 3407 (43%) were randomized to vitamin K antagonists (Table 1). The median age was 73 years at enrollment, 36% were women, and the median CHA₂DS₂-VASc score was 3 [IQR 2, 4]. At enrollment, the median life expectancy was 8 years [IQR 6, 12] (Table S5). Most patients (83%) ended the follow-up period without a clinical event (median 731 days; IQR 415, 1025) (Table 2). In these trials, 530 (7%) patients’ first clinical event was an ischemic stroke or systemic embolism (median 334 days; IQR 120, 580). Additionally, 630 (8%) patients died before a stroke or systemic embolism (median 457 days; IQR 216, 772).

Table 1:

Characteristics of patients with atrial fibrillation in 12 randomized trials

Characteristic	Full cohort (n=7933)	Control* (n=4526)	Vitamin K antagonist (n=3407)	Standardized mean difference
Age, years (median [IQR])	73 [67, 78]	73 [67, 78]	73 [67, 79]	0.02
Sex
Male	5040 (64%)	2879 (64%)	2161 (63%)	<0.01
Female	2893 (36%)	1647 (36%)	1246 (37%)
Diabetes
No	6780 (85%)	3885 (86%)	2895 (85%)	0.03
Yes	1153 (15%)	641 (14%)	512 (15%)
Hypertension
No	4049 (51%)	2353 (52%)	1696 (50%)	0.04
Yes	3884 (49%)	2173 (48%)	1711 (50%)
Congestive heart failure
No	5533 (70%)	3103 (69%)	2430 (71%)	0.06
Yes	2400 (30%)	1423 (31%)	977 (29%)
Prior stroke
No	6413 (81%)	3663 (81%)	2750 (81%)	0.01
Yes	1520 (19%)	863 (19%)	657 (19%)
Angina
No	6566 (83%)	3750 (83%)	2816 (83%)	0.01
Yes	1367 (17%)	776 (17%)	591 (17%)
Prior myocardial infarction				0.17
No	6514 (82%)	3794 (84%)	2720 (80%)
Yes	876 (11%)	506 (11%)	370 (11%)
Missing†	543 (7%)	226 (5%)	317 (9%)
Peripheral vascular disease				0.02
No	5109 (64%)	2930 (65%)	2179 (64%)
Yes	429 (5%)	235 (5%)	194 (6%)
Missing‡	2395 (30%)	1361 (30%)	1034 (30%)
Body mass index kg/m2 (median [IQR])
BMI among those with data	26 [24, 29]	26 [24, 29]	27 [24, 30]	0.06
Missing§	2070 (26%)	1200 (27%)	870 (26%)
Smoking status				0.10
Never smoker	2427 (31%)	1347 (30%)	1080 (32%)
Former smoker	1843 (23%)	1003 (22%)	840 (25%)
Current smoker	790 (10%)	492 (11%)	298 (9%)
Missing‖	2873 (36%)	1684 (37%)	1189 (35%)
CHA2DS2-VASc score
Score (median [IQR])	3 [2, 4]	3 [2, 4]	3 [2, 4]	0.02
Missing#	2938 (37%)	1587 (35%)	1351 (40%)
Trial
AFASAK1	1002 (13%)	668 (15%))	334 (10%)	0.39
AFASAK2	677 (9%)	507 (11%)	170 (5%)
BAATAF	420 (5%)	208 (5%)	212 (6%)
BAFTA	973 (12%)	485 (11%)	488 (14%)
CAFA	375 (5%)	188 (4%)	187 (5%)
EAFT	661 (8%)	438 (10%)	223 (7%)
NASPEAF	543 (7%)	226 (5%)	317 (9%))
PATAF	364 (5%)	244 (5%)	120 (4%)
SPAF1/SPAF2	1306 (16%)	753 (17%)	553 (16%)
SPAF3	1044 (13%)	521 (12%)	523 (15%)
SPINAF	568 (7%)	288 (6%)	280 (8%)

IQR – Interquartile range; CHA₂DS₂-VASc - congestive heart failure, hypertension, age, diabetes, stroke, and vascular disease score; AFASAK - Atrial Fibrillation, Aspirin, and Anticoagulation Study; BAATAF - Boston Area Anticoagulation Trial for Atrial Fibrillation; BAFTA - Birmingham Atrial Fibrillation Treatment of the Aged Study; CAFA - Canadian Atrial Fibrillation Anticoagulation; EAFT - European Atrial Fibrillation Trial; PAATAF - Primary Prevention of Arterial Thromboembolism in Atrial Fibrillation; NASPEAF - National Study for Prevention of Embolism in Atrial Fibrillation; SPAF - Stroke Prevention in Atrial Fibrillation; SPINAF - Stroke Prevention in Non-rheumatic Atrial Fibrillation

^*Control includes those assigned placebo, aspirin, low-dose warfarin, or low-dose warfarin an aspirin. Patients enrolled in SPAF3 and AFASAK2 and assigned to low-dose warfarin had a mean internal normalized ratio (INR) of < 1.5 supporting their categorization as a control. Patients enrolled in NASPEAF and randomized to low-dose vitamin K antagonists had a mean INR of 2.0 and were therefore not included as a control.

^†History of myocardial infarction was not available for NSPEAF trial participants

^‡History of peripheral vascular disease was not available for AFASAK1, BAATAF, or BAFTA trial participants

^§Height and weight were not available for BAATAF or AFASAK2 trial participants

^‖Smoking status was not available for AFASAK1, CAFA, or BAFTA participants. In EAFT, data collection did not distinguish between former and never smokers

^#CHA₂DS₂-VASc scores are for those with complete cases. It excludes 2395 participants enrolled in trials where peripheral vascular disease data are not available, and 543 participants enrolled in trials where history of myocardial infarction was not available.

Since missing data were missing for entire trials, they were assumed to be missing completely at random.

Table 2:

Trial outcomes, rates, and follow-up time for the full cohort and stratified by arm

First clinical outcome	Full Cohort		Control		Vitamin K Antagonist
	Events (%)	Time to event, median days (IQR)	Events (%)	Time to event, median days (IQR)	Events (%)	Time to event, median days (IQR)
Ischemic stroke or systemic embolism	530 (7%)	334 (120, 580)	403 (9%)	330 (119, 563)	127 (4%)	349 (154, 630)
Systemic bleed	175 (2%)	413 (163, 731)	78 (2%)	438 (191, 753)	97 (3%)	403 (125, 720)
Intracranial hemorrhage	29 (0.4%)	300 (201, 620)	12 (0.3%)	535 (242, 918)	17 (0.5%)	265 (158, 499)
Death, all-cause	630 (8%)	457 (216, 772)	344 (8%)	468 (232, 783)	286 (8%)	451 (205, 750)
Study end without a clinical event	6569 (83%)	731 (415, 1025)	3689 (82%)	711 (399, 1005)	2880 (85%)	753 (445, 1049)

IQR – Interquartile range

Comparison of absolute risk reduction estimates

Relative to the Competing Risk Model, the CHA₂DS₂-VASc model overestimated the absolute risk reduction (ARR) of vitamin K antagonists (Figure 1, Figure S2, Table S6). As the time horizon increased, the CHA₂DS₂-VASc estimate of median benefit increased linearly. In contrast, the Competing Risk Model estimated a non-linear absolute risk reduction over time—while benefit increased over time, it decelerated, i.e., absolute risk reduction grew by less each year. As a result, the CHA₂DS₂-VASc model increasingly overestimated vitamin K antagonists benefit as the time-horizon increased. After 1 year, the CHA₂DS₂-VASc model and the Competing Risk Model produced clinically similar estimates of absolute risk reduction (median ARR 2.3% by CHA₂DS₂-VASc estimate vs. 2.4% by Competing Risk Model, p<0.001). After 3 years, the CHA₂DS₂-VASc-based ARR was clinically and statistically larger than the ARR from the Competing Risk Model (median ARR 6.9% vs. 5.2%, p<0.001). This difference increased when absolute risk reduction was estimated over a 5-year horizon (median ARR 11.2% vs. 6.3%, p<0.001).

Figure 1: Estimated median absolute risk reduction of vitamin K antagonists by CHA₂DS₂-VASc model compared with Competing Risk Model.

CHA₂DS₂-VASc - congestive heart failure, hypertension, age, diabetes, stroke, and vascular disease score. ARR - absolute risk reduction. For each patient in the cohort, we estimated the absolute risk reduction attributable to vitamin K antagonists annually for 5 years using the CHA₂DS₂-VASc model (red line) and again using a Fine-Gray model (blue line), a survival model that accounts for the competing risk of death. We plotted the median benefit at each time point. We graphed the median ARR because the CHA₂DS₂-VASc model produces discrete estimates of benefit (i.e., not normally distributed). Data are presented as a table in Table S6. Component on- and off-treatment cumulative incidence rates are also displayed in Figure S2.

We assess whether the observed discrepancy in ARR could be because the guidelines estimate of off-treatment stroke risk were produced in an external cohort and therefore were miscalibrated (Supplemental Methods 3). The sensitivity (Table S7, Table S8, and Figure S3) showed that recalibration to the AFI database did not meaningfully change the results presented in Figure 1.

Life expectancy and misestimation of benefit

As life expectancy decreased, the CHA₂DS₂-VASc model increasingly overestimated the stroke and systemic embolism risk reduction attributable to vitamin K antagonists in absolute and relative terms (Figure 2, Table S9). The figure plots the absolute and relative difference between the 3-year ARR estimated by the CHA₂DS₂-VASc model and the Competing Risk Model by life expectancy decile at trial enrollment. In the decile with the highest life expectancy (16 to 47 years), on average, over 3 years, the CHA₂DS₂-VASc model underestimated benefit by 1.3% (95% CI 1.2% to 1.3%) in absolute terms and 44% (95% CI 41% to 46%) in relative terms. By comparison, in the decile with the lowest life expectancy (1 to 4 years), on average over 3 years, the CHA₂DS₂-VASc model overestimated benefit by 4.7% (95% CI 4.5% to 5.0%) in absolute terms and 78% (95% CI 74% to 82%) in relative terms.

Figure 2: Misestimation of Stroke Risk Reduction of vitamin K antagonists by CHA₂DS₂-VASc score at 3 years.

(A) Absolute misestimation, (B) Relative misestimation. CHA₂DS₂-VASc - congestive heart failure, hypertension, age, diabetes, stroke, and vascular disease score. ARR – absolute risk reduction. Absolute misestimation is defined as: ARR_CHA2DS2-VASc – ARR_{Competing Risk Model}. Relative misestimation is defined as: ARR_CHA2DS2-VASc / ARR_{Competing Risk Model} – 1. The dots represent the mean overestimation, and the error bars represent the 95% confidence interval of the mean. The misestimation of ARR was calculated for each patient as the difference between the ARR computed by the CHA₂DS₂-VASc score and the ARR computed by the Fine-Gray model, a survival model that accounts for the competing risk of death. Positive numbers represent the overestimation of the CHA₂DS₂-VASc model. Tabular results can be found in Table S9.

DISCUSSION

Using patient-level data from 12 randomized trials, we demonstrated that while vitamin K antagonists effectively reduce ischemic stroke and systemic embolism risk, failing to use a competing risks framework resulted in a meaningful overestimation of treatment benefit. This finding was most pronounced when risk reduction was estimated over a multi-year horizon. Further, we showed that as life expectancy decreased, treatment benefit was increasingly overestimated. While those with the highest life expectancy may benefit more than guideline estimates would suggest, benefit for those with the lowest life expectancy was strikingly overestimated.

The study results directly apply to the AHA/ACC/HRS^1,2 and European Society of Cardiology³ atrial fibrillation guidelines in which the cornerstone of anticoagulant decision-making is estimating the absolute risk reduction. Guidelines ask clinicians to estimate the off-treatment stroke risk using the CHA₂DS₂-VASc score and to use that baseline risk to infer the probable absolute risk reduction. Anticoagulants are recommended above a CHA₂DS₂-VASc score threshold—i.e., when the absolute risk reduction exceeds a threshold. Thus, if clinical guidelines continue to recommend treatment using an absolute risk reduction threshold, these results suggest guidelines should re-estimate benefit using a competing risk framework. At the very least, guidelines should acknowledge that current methods overestimate benefits for those with limited life expectancy and when estimating benefits over a multi-year horizon.

Until guidelines formally account for it, clinicians should consider using life expectancy to guide the discussion about the benefit of anticoagulants when treating older adults. When treating older adults with atrial fibrillation, physicians should start with the approach recommended in current guidelines by calculating the CHA₂DS₂-VASc score. Then, based on the results of this study, physicians should estimate life expectancy and use it to individualize treatment recommendations. This may be particularly relevant when treating patients with both a limited life expectancy and borderline CHA₂DS₂-VASc scores. In this study, we estimated life expectancy using basic medical comorbidity data available in the trial database. Modern, more accurate tools like ePrognosis go beyond comorbidities, using physical function, cognition, and self-reported health to estimate life expectancy.^39,40 Life expectancy estimates from such tools are routinely used to inform the risks and benefits of interventions in older adults (e.g., cancer screening). These study results also lend credence to physicians who already factor advanced age, frailty, and function—all significant determinants of life expectancy—into in their anticoagulant decision-making.^41,42 Prior work indicates that many older adults with atrial fibrillation have geriatric syndromes (e.g., dependency in activities of daily living) known to be associated with a reduced life expectancy.^43,44 Finally, these results should prompt patients and physicians to periodically reassess the expected benefit of anticoagulants as comorbidities, life expectancy, and preferences change as patients with atrial fibrillation age.

The current study’s findings should influence anticoagulant decision aids for patients with atrial fibrillation. To advance anticoagulant shared decision-making, investigators and professional societies have developed conversation aids that display a patient’s risk of stroke with and without anticoagulants. For example, the American College of Cardiology’s CardioSmart tool and the Mayo Clinic Anticoagulation Choice Decision Aid both display a pictogram of absolute risk with and without treatment to communicate treatment effects.^4,5 These implementation tools are built to reflect clinical guidelines and do so faithfully. However, both should note that benefit estimates are overstated for those with limited life expectancy and when benefit is estimated over a multi-year horizon. Further, decision aids should limit the time horizon to 1-year when displaying event rates since the effect of competing risks is smaller over shorter time horizons.

Finally, these findings should inform the methods used to test new atrial fibrillation therapies and the methods used to develop and validate future stroke risk models. When new atrial fibrillation therapies are tested, RCTs should be analyzed using an approach that accounts for competing risks. This is because the sizable background mortality rate in people with atrial fibrillation can result in misestimation of benefits when competing events are discounted. For example, RCTs testing left atrial appendage closure (LAAC) did not use a competing risks framework.⁴⁵ Subsequent observational research showed substantial competing risks in patients undergoing LAAC, which Wang et al. argue should factor into clinical decision-making.⁴⁶ Additionally, while the CHA₂DS₂-VASc score continues to be endorsed by U.S. and European guidelines, investigators are actively developing a new generation of stroke risk prediction models. Contemporary models like the ABC (Age, Biomarkers, Clinical history) stroke risk score, CARS (Calculator of Absolute Stroke Risk), and the ATRIA (Anticoagulation and Risk Factors in Atrial Fibrillation) stroke model all outperform the CHA₂DS₂-VASc score.^47–49 However, none used an analysis framework that both accounts for the competing risk of death and changing risk over a multi-year horizon. The ABC model establishes a non-linear stroke risk by showing 3-year risk is not simply three times the 1-year risk, findings that were redemonstrated in this paper. When developing the ABC model, Hijazi et al. also conducted a sensitivity analysis comparing their model to a competing risk model.⁴⁷ They found a tight correlation when risk was estimated over a 1-year horizon, results mirrored in this study. We expanded on their work by showing that this correlation was weaker when using a longer time horizon (e.g., 3 years). More importantly, we identified substantial heterogeneity—as life expectancy decreased, overestimation increased.

There are limitations to this study inherent to the data available and the study design. First, this study relied on data from RCTs conducted between 1989 and 2007 and thus may be only partially representative of contemporary patients with atrial fibrillation. Specific differences include the risk of stroke and death from non-atrial fibrillation causes and the added safety of direct-acting anticoagulants.⁵⁰ This limitation is balanced by the fact that the AFI cohort is one of the largest patient-level atrial fibrillation cohorts where anticoagulant treatment was randomized against placebo and antiplatelets. The results were unaffected by selection bias that hampers contemporary risk models developed in observational cohorts. More importantly, while dated, these trials are the foundation upon which current guidelines recommend anticoagulants. Second, this study could not address the relationship between life expectancy and the potential misestimation of anticoagulant harm. Specifically, the AF Investigators database does not include inputs used in contemporary hemorrhage prediction tools (e.g., ATRIA bleed, HAS-BLED). While it is important to consider the effect a competing risk framework may have on estimating the risk of bleeding, current guidelines do not incorporate bleeding risk into treatment recommendations. For example, for patients with scores above the CHA₂DS₂-VASc treatment threshold, guideline recommendations do not change whether the bleeding risk is high or low. Third, because nationality was unavailable for study participants, we relied on U.S. life tables to calculate life expectancy. Finally, while the RCTs reflect a diverse experience with studies from the U.S. and Europe, most did not collect data on race, ethnicity, or socioeconomic status, which are important to communicate the generalizability of the study findings.

In summary, we showed that while vitamin K antagonists were clearly effective, treatment benefit was overstated when using the guideline-endorsed approach because guidelines do not account for the competing risk of death and assume a constant growth in treatment benefit over time. Overestimation was most pronounced in patients with the lowest life expectancy and when benefit was estimated over a multi-year horizon. These findings should inform guidelines and decision aids, clinicians treating patients with limited life expectancy, and investigators developing stroke risk prediction models.

What is Known

• When using anticoagulants for stroke prevention in atrial fibrillation, clinicians must balance benefits (i.e., stroke prevention) against harms (i.e., excess bleeding).

• Guidelines recommend using the CHA2DS2-VASc score to estimate benefit, which does not account for life expectancy and assumes a constant benefit over time.

What the Study Adds

• This study finds that anticoagulants were very effective, but the guideline approach overstated anticoagulant benefits.

• We find that the overestimation of benefit tracked with life expectancy--as life expectancy decreased, treatment benefit was increasingly overestimated.

• Guidelines, decision aids, and clinicians should individualize treatment recommendations by acknowledging that treatment benefits may be overstated for those with limited life expectancy.

Non-standard Abbreviations and Acronyms

ABC

Age, Biomarkers, Clinical history stroke risk score

ACC

American College of Cardiology

AF

atrial fibrillation

AFASAK

Atrial Fibrillation, Aspirin, and Anticoagulation Study

AFI

Atrial Fibrillation Investigators

AHA

American Heart Association

ARR

absolute risk reduction

ATRIA

Anticoagulation and Risk Factors in Atrial Fibrillation

BAATAF

Boston Area Anticoagulation Trial for Atrial Fibrillation

BAFTA

Birmingham Atrial Fibrillation Treatment of the Aged Study

CAFA

Canadian Atrial Fibrillation Anticoagulation Trial

CARS

Calculator of Absolute Stroke Risk

CHA₂DS₂-VASc score

congestive heart failure, hypertension, age, diabetes, stroke, and vascular disease score

HAS-BLED

hypertension, abnormal renal/liver function, stroke, bleeding history or predisposition, labile international normalized ratio, age ≥65 years, drugs/alcohol concomitantly

HRS

Heart Rhythm Society

INR

international normalized ratio

IQR

interquartile range

LAAC

left atrial appendage closure

NASPEAF

National Study for Prevention of Embolism in Atrial Fibrillation

PAATAF

Primary Prevention of Arterial Thromboembolism in Atrial Fibrillation Trial

RCT

randomized controlled trial

SPAF

Stroke Prevention in Atrial Fibrillation Trial

SPINAF

Stroke Prevention in Non-rheumatic Atrial Fibrillation

VKA

vitamin K antagonist