Frailty and Clinical Outcomes of Direct Oral Anticoagulants versus Warfarin In a Cohort Study of Older Adults with Atrial Fibrillation

Abstract

BACKGROUND:

The role of differing levels of frailty in choice of oral anticoagulants for older adults with atrial fibrillation (AF) remains unclear.

OBJECTIVE:

To examine the outcomes of direct oral anticoagulants (DOAC) versus warfarin by frailty levels

DESIGN:

1:1 propensity score-matched analysis of Medicare data, 2010-2017

SETTING:

Community

PATIENTS:

Medicare beneficiaries with AF who initiated dabigatran, rivaroxaban, apixaban, or warfarin

MEASUREMENTS:

Composite endpoint of death, ischemic stroke, or major bleeding by frailty levels, defined by a claims-based frailty index

RESULTS:

In the dabigatran-warfarin cohort (n=158,712; median follow-up, 72 days), the event rate (per 1,000 person-years) was 63.5 for dabigatran-initiators and 65.6 for warfarin-initiators (hazard ratio [HR], 0.98 [95% CI: 0.92, 1.05]; rate difference [RD], −2.2 [−6.5, 2.1]). For non-frail, pre-frail, and frail groups, HRs were 0.81 (0.68, 0.97), 0.98 (0.90, 1.08), and 1.09 (0.96, 1.23), respectively. In the rivaroxaban-warfarin cohort (n=275,944; median follow-up, 82 days), the event rate (per 1,000 person-years) was 77.8 for rivaroxaban-initiators and 83.7 for warfarin-initiators (HR, 0.98 [0.94-1.02]; RD, −3.9 [−10.3, 2.4]). For non-frail, pre-frail, and frail groups, HRs were 0.88 (0.77, 0.99), 1.04 (0.98, 1.10), and 0.96 (0.89, 1.04), respectively. In the apixaban-warfarin cohort (n=218,738; median follow-up, 84 days), the event rate (per 1,000 person-years) was 60.1 for apixaban-initiators and 92.3 for warfarin-initiators (HR, 0.68 [0.65, 0.72]; RD, −32.2 [−36.1, −28.3]). For non-frail, pre-frail, and frail groups, HRs were 0.61 (0.52, 0.71), 0.66 (0.61, 0.70), and 0.73 (0.67, 0.80), respectively.

LIMITATIONS:

Residual confounding, lack of clinical frailty assessment

CONCLUSION:

For older adults with AF, apixaban was associated with lower rates of adverse events across all frailty levels. Dabigatran and rivaroxaban were associated with lower event rates only among non-frail patients.

FUNDING SOURCE:

National Institute on Aging

INTRODUCTION

The efficacy and safety of direct oral anticoagulants (DOACs) in patients with atrial fibrillation (AF) are well established from clinical trials.(1-4) DOACs are at least as efficacious as warfarin with generally lower rates of major bleeding, fewer drug-drug or drug-food interactions, and no need for therapeutic monitoring or frequent dose adjustments. Owing to these advantages, DOACs are preferred over warfarin.(5, 6) The safety profile of DOACs may be particularly desirable for older adults with frailty,(5) who are at high risk for falls(7) and drug-related adverse events.(8) Currently, decisions about anticoagulation therapy are mainly driven by risk assessment models for ischemic stroke(9, 10) and major bleeding,(10-12) without consideration of frailty. Due to poor representation of older adults with frailty and lack of frailty assessment in clinical trials, the role of frailty in the choice between a DOAC and warfarin is uncertain(13) and anticoagulant use remains suboptimal in frail patients with AF.(14, 15)

To fill this evidence gap, we analyzed Medicare data, which contain information on prescription drug use and clinical events in older adults with AF, including those with frailty. We measured frailty using a validated claims-based frailty index (CFI).(16-19) The objective of this study was to investigate how frailty affects the association of a DOAC (dabigatran, rivaroxaban, apixaban) vs warfarin with ischemic stroke, major bleeding, and death in older adults with AF. We postulated that frailty would be an effect modifier of the benefit and safety of a DOAC vs warfarin.

METHODS

Study Design and Population

This retrospective observational study using Medicare data was approved by the Institutional Review Board of the Brigham and Women’s Hospital, Boston, Massachusetts. We emulated 3 trials, each comparing a DOAC with warfarin. We created new-user cohorts of fee-for-service Medicare beneficiaries with AF who filled a prescription for one of the 3 DOACs or warfarin from the date of approval (dabigatran cohort: October 19, 2010; rivaroxaban cohort: November 4, 2011; apixaban cohort: December 28, 2012) through December 31, 2017. Eligible beneficiaries were those who were ≥65 years old; filled their prescription for a DOAC or warfarin with no exposure in the past 183 days; had ≥1 inpatient or outpatient diagnosis of AF in the past 183 days; and had continuous enrollment in Medicare Part A, B, and D, and no enrollment in Medicare Advantage Plan in the past 183 days. We excluded those who 1) had missingness on age, sex, race, or geographical regions; 2) received hospice care in the past 183 days; 3) resided in a nursing facility at the time of drug initiation; 4) received another DOAC in the past 183 days; 5) had other indications for anticoagulation (venous thromboembolism or joint replacement); or 6) had contraindications to either DOAC or warfarin therapy (valvular disease, intracranial or retroperitoneal bleeding, chronic kidney disease stage 5 or end stage renal disease or renal transplant, or stroke within 14 days). These conditions were defined in Supplement 1. Beneficiaries who satisfied these selection criteria represent “clinical care population”. We applied additional exclusion criteria to mimic clinical trials(1-4) (e.g., any nursing facility admission, chronic kidney disease stage 4, gastrointestinal bleeding, treatment for malignancy, chronic non-steroidal anti-inflammatory drug therapy, and limited life expectancy based on the combined comorbidity score(20) ≥5 or CFI(16-19) ≥0.35, which corresponds to the predicted 1-year mortality ≥20%). This “trial population” was used in a sensitivity analysis (see diagrams in Supplement 2, Figures 1-3).

Measurement of Frailty and Baseline Covariates

Frailty was measured using a CFI,(16-19) which estimates a deficit-accumulation frailty index (range: 0-1) using 93 variables defined by diagnosis, health services, and durable medical equipment codes in the year before the drug initiation (https://dataverse.harvard.edu/dataverse/cfi). The CFI has been validated against gait speed,(17) grip strength,(17) frailty phenotype,(18) deficit-accumulation frailty index,(18) and severe disability.(18) Using accepted cutpoints,(21-23) we defined non-frailty if CFI <0.15, pre-frailty if CFI 0.15-0.24, and frailty if CFI ≥0.25.

We measured comorbidities, prescription drug use, and health care utilization during the 183 days prior to the drug initiation. We computed the CHA₂DS₂-VASc score(9, 10) for the ischemic stroke risk and the modified HAS-BLED score (labile international normalized ratio was unavailable in Medicare data) for the bleeding risk.(10-12) The burden of comorbidity was quantified using a combined comorbidity score.(20)

Outcomes and Follow-Up

The primary outcome was a composite endpoint of death, ischemic stroke, or major bleeding. Secondary outcomes included individual components of the primary outcome, major gastrointestinal bleeding, and intracranial bleeding. The date of death was obtained through linkage to Social Security files. Clinical events were defined using validated claims-based algorithms (Supplement 1). Ischemic stroke was defined based on the diagnosis codes in the primary position of the inpatient discharge diagnoses (positive predictive value: 85%-90%).(24, 25) Major bleeding was defined based on 1) critical site bleeding (e.g., intraarticular, intracranial, intramuscular, intraocular, intraspinal, pericardial, and retroperitoneal) in the primary position of the inpatient discharge diagnoses or 2) bleeding-related diagnosis codes in the primary position of the inpatient discharge diagnoses (positive predictive value: 86-96%)(26) with transfusion procedure codes or revenue codes. This algorithm for major bleeding has been used in previous claims-based studies.(27, 28)

Follow-up began on the day after initiating the index drug and continued until the earliest of the following events: occurrence of study outcome, disenrollment from Medicare, end of study period (December 31, 2017), discontinuation of the index drug (defined as a treatment gap [grace period] longer than 5 days), switching to a different anticoagulant, admission to a nursing facility, enrollment in hospice, initiation of dialysis, or kidney transplantation. We assumed that individuals remained exposed to the index drug as long as the treatment gap is 5 days or less, and for an additional 5 days after discontinuation (exposure risk window). Because the duration of anticoagulant action differs between a DOAC (2-3 days)(29-31) and warfarin (up to 5 days),(32) we also analyzed data by allowing a 3-day grace period and an exposure risk window of 3 days.

Statistical Analysis

The analyses were conducted for each DOAC-warfarin cohort both in the entire population and stratified by the frailty category. We compared clinical characteristics between the DOAC-treated group and warfarin-treated group. To reduce imbalance in clinical characteristics between the groups, we estimated a propensity score (PS) using a logistic regression that modeled the probability of initiating a DOAC vs warfarin as a function of age, sex, race (white, black, and other), dual eligibility, chronic conditions, CHA₂DS₂-VASc score, HAS-BLED score, the combined comorbidity index, CFI, prescription drug classes, emergency department visits, home health days, hospitalizations, skilled nursing facility visits, calendar year, and geographical region (states) (see Supplement 2, Table 1 for the PS models). We performed 1:1 nearest neighbor matching with a caliper of 1% of the PS for the total population and repeated matching for each frailty category. The algorithm matches patient pairs with the smallest difference in the PS first. Standardized mean difference <0.10 was considered adequate. In the matched cohort, we summarized the rates of primary and secondary outcomes and estimated HRs and 95% confidence intervals (CI) using Cox proportional hazards model. Rate differences (RD) were also estimated. We tested whether there was a heterogeneity in HR and RD estimates across frailty categories.(33) Sensitivity analyses were performed to assess the robustness of the findings under different assumptions: 1) standard dose only excluding the renal dose of DOACs; 2) 3-day grace period and 3-day exposure risk window; 3) intention-to-treat analysis not censoring individuals at the time of the index drug discontinuation or switching and following them for 365 days; and 4) trial populations. We also examined how our results would change under different scenarios of unmeasured confounding and calculated the minimum strength of the confounder-exposure and confounder-outcome associations on the risk ratio scale to fully explain non-null associations (or E-values).(34)

Analyses were conducted in the Aetion Evidence Generation Platform (including R version 3.4.2), which has been validated by accurately repeating a range of published studies(35) and by replicating(36) or predicting clinical trial findings.(37)

Role of the Funding Source

This study was funded by National Institute on Aging. The funder had no role in the design, collection, analysis, or interpretation of the data, or the decision to submit the manuscript for publication.

RESULTS

Characteristics of Study Populations

We created 3 new-user cohorts of dabigatran (n=81,863) vs warfarin (n=256,722); rivaroxaban (n=185,011) vs warfarin (n=228,028); apixaban (n=222,478) vs warfarin (n=206,031). Compared with warfarin initiators, dabigatran initiators (Supplement 2, Table 2) and rivaroxaban initiators (Supplement 2, Table 3) were younger (mean age, 76.3 vs 77.4 years in both cohorts) and healthier, as evidenced by lower mean CFI (0.19 vs 0.20 in both cohorts), lower mean CHA₂DS₂-VASc score (4.0 vs 4.3 in both cohorts), lower mean comorbidity score (dabigatran: 2.2 vs 2.6; rivaroxaban: 2.3 vs 2.7), and lower prevalence of comorbidities (e.g., heart failure, anemia, and acute and chronic kidney disease). Apixaban initiators (Supplement 2, Table 4) had similar age (77.3 vs 77.5 years), CFI (0.20 vs 0.20), mean CHA₂DS₂-VASc score (4.2 vs 4.2), mean comorbidity score (2.6 vs 2.7) to warfarin initiators. After PS matching, all characteristics were adequately balanced between DOAC and warfarin groups in the entire cohort (Table 1) and by frailty subgroups (Supplement 2, Tables 5-7). Frail patients were older and more likely to be female; had greater comorbidity score, CHA₂DS₂-VASc score, and HAS-BLED score; consumed more medications; and had higher health care utilization.

Table 1.

Propensity Score-Matched Clinical Care Population with Atrial Fibrillation Treated with Direct Oral Anticoagulants vs Warfarin^*

Characteristics	Dabigatran Cohort		Rivaroxaban Cohort		Apixaban Cohort
	Warfarin (n=79,365)	Dabigatran (n=79,365)	Warfarin (n=137,972)	Rivaroxaban (n=137,972)	Warfarin (n=109,369)	Apixaban (n=109,369)
Age, years	76.4 (7.1)	76.4 (7.1)	76.8 (7.3)	76.8 (7.2)	77.3 (7.4)	77.3 (7.4)
Female	49.9	49.9	49.9	49.9	50.6	50.4
White race	91.7	91.7	91.2	91.2	91.5	91.5
Combined comorbidity index	2.2 (2.1)	2.2 (2.2)	2.4 (2.3)	2.4 (2.3)	2.7 (2.5)	2.7 (2.5)
CFI	0.19 (0.06)	0.19 (0.07)	0.20 (0.07)	0.20 (0.07)	0.20 (0.07)	0.20 (0.07)
CHA2DS2-VASc score	4.1 (1.6)	4.1 (1.6)	4.1 (1.6)	4.1 (1.6)	4.2 (1.6)	4.2 (1.6)
HAS-BLED score	2.0 (0.7)	2.0 (0.7)	2.1 (0.7)	2.1 (0.7)	2.1 (0.7)	2.1 (0.7)
Medical history
Acute kidney disease	5.5	5.5	8.3	8.2	10.8	10.8
Acute myocardial infarction	3.3	3.2	4.4	4.3	5.4	5.4
Anemia	18.2	18.3	19.6	19.6	21.4	21.5
Cancer	13.4	13.5	13.8	13.9	14.0	13.8
Cardioversion	3.8	3.9	2.9	2.9	3.0	3.1
Chronic kidney disease	9.7	9.8	11.9	11.9	16.4	16.5
Chronic lung disease	19.6	19.6	21.3	21.2	22.1	22.1
Dementia	5.1	5.1	5.6	5.6	6.1	6.0
Diabetes	33.0	32.9	33.7	33.7	34.7	34.8
Falls	0.6	0.6	0.4	0.4	0.4	0.4
Fractures	0.7	0.7	1.3	1.3	1.1	1.1
GI bleeding (inpatient)	1.5	1.5	2.0	2.0	2.6	2.5
Heart failure (inpatient)	12.2	12.2	14.0	14.0	15.8	15.9
Hypertension	62.3	62.7	66.5	66.5	69.5	69.6
Ischemic heart disease	37.7	37.7	37.7	37.6	39.0	39.0
Stroke (inpatient)	4.6	4.6	4.8	4.7	5.6	5.5
Transient ischemic attack	5.1	5.1	4.4	4.5	4.6	4.6
Medications
ACE inhibitors	25.3	25.3	26.0	26.1	26.7	26.6
Angiotensin receptor blockers	7.1	7.1	5.4	5.3	5.0	4.9
Antiarrhythmic agents	20.4	20.5	17.3	17.3	17.0	17.0
Antiplatelet agents	14.6	14.8	14.0	13.9	14.5	14.4
Benzodiazepines	7.2	7.2	13.0	13.0	14.6	14.6
Beta-blockers	62.0	62.2	63.8	63.9	65.6	65.6
Calcium channel blockers	2.3	2.3	2.1	2.1	2.0	2.1
Dementia drugs	4.2	4.2	4.2	4.2	4.2	4.2
Diuretics	49.1	49.0	50.8	50.7	52.6	52.6
H2-blockers	5.8	5.8	6.1	6.1	6.4	6.4
Hypnotics	8.6	8.7	7.2	7.1	6.5	6.4
Insulin	5.8	5.8	6.6	6.7	7.5	7.5
Metformin	15.3	15.0	15.4	15.5	15.5	15.6
NSAIDs	13.6	13.6	12.5	12.5	11.5	11.5
Opioids	29.0	29.1	30.3	30.1	30.0	30.2
Proton pump inhibitors	26.9	26.9	28.3	28.1	29.1	29.1
SSRIs/SNRIs	16.9	16.8	17.6	17.5	18.1	18.1
Statins	57.8	57.9	58.3	58.3	59.7	59.8
Health care utilization in 183 days
Emergency department visits	0.36 (0.78)	0.35 (0.76)	0.40 (0.82)	0.40 (0.79)	0.42 (0.83)	0.42 (0.78)
Home health services	0.11 (0.46)	0.11 (0.44)	0.07 (0.36)	0.07 (0.35)	0.04 (0.28)	0.05 (0.28)
Hospitalizations	0.49 (0.77)	0.50 (0.77)	0.54 (0.81)	0.54 (0.79)	0.57 (0.84)	0.57 (0.82)
Skilled nursing facility stays	0.08 (0.52)	0.08 (0.50)	0.06 (0.43)	0.06 (0.44)	0.03 (0.30)	0.03 (0.31)
Geographic region
Northeast	19.9	19.8	19.3	19.3	19.8	19.8
Midwest	22.6	22.5	25.9	25.9	26.9	27.0
South	40.1	40.3	37.2	37.1	35.5	35.7
West	17.3	17.3	17.5	17.6	17.7	17.4
Other	0.1	0.1	0.1	0.1	0.1	0.1

Abbreviations: ACE, angiotensin converting enzyme; CFI, claims-based frailty index; GI, gastrointestinal; NSAID, non-steroidal anti-inflammatory drug; SNRI, serotonin-norepinephrine reuptake inhibitor; SSRI, selective serotonin reuptake inhibitor.

^*Data are presented in proportions or mean (standard deviation). Standardized mean difference <0.1 for all variables (see Supplementary Table 2 for the entire list of variables).

Among the 3 PS-matched cohorts (Table 1), the apixaban cohort had older mean age (dabigatran, 76.4 years; rivaroxaban, 76.8 years; apixaban, 77.3 years), greater mean CFI (dabigatran, 0.19; rivaroxaban, 0.20; apixaban, 0.20), greater mean CHA₂DS₂-VASc score (dabigatran, 4.1; rivaroxaban, 4.1; apixaban, 4.2), greater mean HAS-BLED score (dabigatran, 2.0; rivaroxaban, 2.1; apixaban, 2.1), mean comorbidity score (dabigatran, 2.2; rivaroxaban, 2.4; apixaban, 2.7), and higher prevalence of comorbidities (e.g., acute myocardial infarction, anemia, acute and chronic kidney disease, gastrointestinal bleeding, and heart failure) and medication use (e.g., benzodiazepines). The apixaban cohort showed highest use of emergency department visits and hospitalizations and lowest use of home health and skilled nursing facility services.

Clinical Outcomes Associated with Dabigatran vs Warfarin by Frailty Levels

Over a median follow-up (interquartile range) of 72 (33, 143) days, the rate of the composite endpoint of death, ischemic stroke, or major bleeding (per 1,000 person-years) was 63.5 for dabigatran initiators and 65.6 for warfarin initiators (HR [95% CI], 0.98 [0.92, 1.05]; RD [95% CI] per 1,000 person-years, −2.2 [−6.5, −2.1]) (Table 2 and Supplement 2, Table 8). There was evidence for heterogeneity by frailty on the HR scale (P_{heterogeneity}=0.027), not on the RD scale (P_{heterogeneity}=0.231). Dabigatran was associated with a lower rate of the composite endpoint than warfarin for non-frail subgroup (HR, 0.81 [0.68, 0.97]), but not for pre-frail or frail subgroup (Table 2 and Figure 1A).

Table 2.

Frailty and the Association of Direct Oral Anticoagulants vs Warfarin with Clinical Outcomes in Propensity Score-Matched Clinical Care Population with Atrial Fibrillation^*

Outcomes	Dabigatran Cohort (Rate per 1,000 PYs)				Rivaroxaban Cohort (Rate per 1,000 PYs)				Apixaban Cohort (Rate per 1,000 PYs)
	Warfarin (n=79,365)	Dabigatran (n=79,365)	HR (95% CI)	RD (95% CI)	Warfarin (n=137,972)	Rivaroxaban (n=137,972)	HR (95% CI)	RD (95% CI)	Warfarin (n=109,369)	Apixaban (n=109,369)	HR (95% CI)	RD (95% CI)
Composite event
Total	65.6	63.5	0.98 (0.92, 1.05)	−2.2 (−6.5, 2.1)	83.7	77.8	0.98 (0.94, 1.02)	−5.9 (−9.4, −2.4)	92.3	60.1	0.68 (0.65, 0.72)	−32.2 (−36.1, −28.3)
Non-frail	31.4	25.4	0.81 (0.68, 0.97)	−6.1 (−11.2, −0.9)	37.2	31.6	0.88 (0.77, 0.99)	−5.6 (−9.9, −1.3)	37.6	21.9	0.61 (0.52, 0.71)	−15.7 (−20.4, −11.0)
Pre-frail	64.1	62.1	0.98 (0.90, 1.08)	−2.1 (−7.8, 3.6)	76.7	76.5	1.04 (0.98, 1.10)	−0.2 (−4.8, 4.4)	86.0	54.2	0.66 (0.61, 0.70)	−31.8 (−36.7, −26.9)
Frail	160.2	170.0	1.09 (0.96, 1.23)	9.8 (−9.9, 29.6)	219.8	200.8	0.96 (0.89, 1.04)	−19.0 (−35.2, −2.9)	226.2	157.6	0.73 (0.67, 0.80)	−68.7 (−85.3, −52.0)
P heterogeneity			0.027	0.231			0.026	0.040			0.080	<0.001
Death
Total	30.8	31.2	1.03 (0.93, 1.13)	0.4 (−2.5, 3.4)	42.7	37.0	0.91 (0.86, 0.97)	−5.7 (−8.2, −3.2)	48.2	37.9	0.82 (0.77, 0.88)	−10.3 (−13.2, −7.3)
Non-frail	12.5	11.1	0.89 (0.67, 1.18)	−1.3 (−4.6, 2.0)	16.4	11.7	0.72 (0.60, 0.88)	−4.8 (−7.5, −2.0)	16.5	11.8	0.74 (0.58, 0.92)	−4.7 (−7.9, −1.5)
Pre-frail	28.5	28.1	1.00 (0.87, 1.14)	−0.4 (−4.2, 3.4)	37.4	32.6	0.90 (0.83, 0.98)	−4.8 (−7.9, −1.7)	43.0	30.9	0.74 (0.68, 0.82)	−12.2 (−15.7, −8.6)
Frail	84.3	98.7	1.19 (1.01, 1.40)	14.5 (−0.2, 29.1)	129.7	116.5	0.94 (0.85, 1.04)	−13.2 (−25.5, −0.9)	131.8	110.5	0.88 (0.79, 0.98)	−21.4 (−34.5, −8.2)
P heterogeneity			0.127	0.118			0.066	0.414			0.058	0.001
Ischemic stroke
Total	7.2	5.1	0.72 (0.58, 0.89)	−2.1 (−3.5, −0.8)	8.2	5.6	0.71 (0.61, 0.83)	−2.6 (−3.6, −1.5)	10.1	6.1	0.63 (0.54, 0.74)	−4.0 (−5.3, −2.7)
Non-frail	3.9	3.0	0.78 (0.46, 1.31)	−0.9 (−2.7, 0.9)	4.6	3.8	0.87 (0.61, 1.24)	−0.8 (−2.3, 0.7)	5.0	4.2	0.89 (0.60, 1.32)	−0.8 (−2.6, 1.1)
Pre-frail	6.7	5.7	0.86 (0.65, 1.15)	−1.1 (−2.8, 0.7)	8.6	6.5	0.78 (0.65, 0.95)	−2.1 (−3.5, −0.6)	10.4	6.9	0.68 (0.56, 0.83)	−3.6 (−5.3, −1.9)
Frail	8.4	8.1	0.99 (0.58, 1.69)	−0.2 (−4.6, 4.2)	14.0	9.4	0.68 (0.49, 0.95)	−4.6 (−8.4, −0.8)	17.4	11.2	0.67 (0.49, 0.92)	−6.2 (−10.8, −1.7)
P heterogeneity			0.819	0.945			0.607	0.149			0.457	0.024
Major bleeding
Total	30.7	29.0	0.97 (0.88, 1.07)	−1.7 (−4.6, 1.3)	36.1	37.6	1.09 (1.03, 1.17)	1.4 (−1.0, 3.8)	37.4	18.1	0.51 (0.46, 0.55)	−19.3 (−21.7, −17.0)
Non-frail	16.7	12.3	0.74 (0.58, 0.96)	−4.4 (−8.1, −0.7)	18.8	17.4	0.95 (0.80, 1.13)	−1.5 (−4.6, 1.7)	18.9	6.8	0.37 (0.29, 0.49)	−12.1 (−15.2, −9.0)
Pre-frail	31.5	29.9	0.97 (0.85, 1.10)	−1.6 (−5.6, 2.4)	33.9	39.5	1.22 (1.12, 1.33)	5.5 (2.4, 8.7)	35.6	18.2	0.53 (0.47, 0.60)	−17.4 (−20.5, −14.4)
Frail	71.9	67.7	0.98 (0.81, 1.17)	−4.2 (−17.0, 8.6)	82.2	80.3	1.03 (0.91, 1.16)	−1.9 (−11.8, 8.2)	83.4	40.7	0.51 (0.44, 0.60)	−42.7 (−52.2, −33.2)
P heterogeneity			0.158	0.591			0.010	0.006			0.071	<0.001
Major GI bleeding
Total	15.7	22.0	1.44 (1.27, 1.63)	6.3 (4.0, 8.7)	19.7	26.0	1.40 (1.29, 1.52)	6.2 (4.4, 8.1)	20.6	10.1	0.52 (0.46, 0.58)	−10.6 (−12.3, −8.8)
Non-frail	7.5	9.2	1.23 (0.88, 1.72)	1.7 (−1.1, 4.5)	8.9	12.0	1.42 (1.13, 1.79)	3.1 (0.8, 5.5)	9.1	3.5	0.40 (0.28, 0.58)	−5.6 (−7.8, −3.4)
Pre-frail	17.1	23.3	1.39 (1.18, 1.64)	6.2 (3.0, 9.4)	18.6	27.4	1.56 (1.39, 1.74)	8.8 (6.3, 11.3)	19.9	10.3	0.55 (0.47, 0.64)	−9.6 (−11.9, −7.3)
Frail	46.3	50.1	1.13 (0.90, 1.41)	3.8 (−6.8, 14.5)	48.3	55.8	1.22 (1.05, 1.43)	7.6 (−0.4, 15.6)	49.7	25.3	0.54 (0.44, 0.66)	−24.4 (−31.8, −17.0)
P heterogeneity			0.329	0.116			0.044	0.005			0.291	<0.001
Intracranial bleeding
Total	8.4	3.5	0.42 (0.33, 0.53)	−4.9 (−6.2, −3.6)	9.2	5.6	0.62 (0.53, 0.72)	−3.6 (−4.7, −2.5)	10.0	5.0	0.51 (0.43, 0.60)	−5.1 (−6.3, −3.8)
Non-frail	5.9	1.8	0.31 (0.17, 0.55)	−4.1 (−6.0, −2.2)	6.6	3.6	0.54 (0.38, 0.75)	−3.0 (−4.7, −1.4)	6.5	2.5	0.39 (0.25, 0.60)	−4.1 (−5.9, −2.2)
Pre-frail	8.0	3.6	0.45 (0.32, 0.61)	−4.4 (−6.2, −2.7)	8.5	5.6	0.66 (0.54, 0.81)	−2.9 (−4.3, −1.5)	9.2	5.3	0.58 (0.47, 0.73)	−3.9 (−5.5, −2.4)
Frail	10.5	6.9	0.68 (0.40, 1.15)	−3.6 (−8.1, 0.9)	17.8	10.1	0.58 (0.43, 0.79)	−7.7 (−11.9, −3.5)	18.0	8.3	0.47 (0.33, 0.66)	−9.7 (−14.1, −5.4)
P heterogeneity			0.138	0.936			0.550	0.100			0.234	0.046

Abbreviations: CI, confidence interval; GI, gastrointestinal; HR, hazard ratio; PY, person-years; RD, rate difference.

^*Individuals were classified as non-frail (claims-based frailty index <0.15), pre-frail (0.15-0.24), and frail (≥0.25). The composite endpoint (primary endpoint) was defined as any of death, ischemic stroke, or major bleeding (which includes major GI bleeding and intracranial bleeding).

Figure 1. Frailty and Cumulative Incidence Plots of a Composite Endpoint of Death, Ischemic Stroke, or Major Bleeding in Older Adults with Atrial Fibrillation Newly Treated with Direct Oral Anticoagulants vs Warfarin.

Abbreviations: CI, confidence interval; HR, hazard ratio; RD, rate difference. In the propensity score-matched cohort of 158,730 Medicare beneficiaries newly treated with dabigatran vs warfarin (panel A), dabigatran was associated with a modest reduction in the composite endpoint of death, ischemic stroke, or major bleeding only in non-frail subgroup. In the propensity score-matched cohort of 275,944 Medicare beneficiaries newly treated with rivaroxaban vs warfarin (panel B), rivaroxaban was associated with a modest reduction in the composite endpoint in non-frail subgroup. In the propensity score-matched cohort of 218,738 Medicare beneficiaries newly treated with apixaban vs warfarin (panel C), apixaban was associated with a reduction in the composite endpoint in non-frail, pre-frail, and frail subgroups. Frailty status was defined using a validated claims-based frailty index: non-frail if <0.15, pre-frail if 0.15-0.24, and frail if ≥0.25.

For secondary endpoints, dabigatran initiators showed lower rates of ischemic stroke (HR [95% CI], 0.72 [0.58, 0.89]; RD per 1,000 person-years [95% CI], −2.1 [−3.5, −0.8]) and intracranial bleeding (HR, 0.42 [0.33, 0.53]; RD, 4.9 [−6.2, −3.6]); similar rates of death (HR, 1.03 [0.93, 1.13]; RD, −0.4 [−2.5, 3.4]) and major bleeding (HR, 0.97 [0.88, 1.07]; RD, −1.7 [−4.6, 1.3]); and a higher rate of major gastrointestinal bleeding (HR, 1.44 [1.27, 1.63]; RD, 6.3 [4.0, 8.7]) (Table 2 and Supplement 2, Table 8). There was no statistically significant evidence for heterogeneity by frailty on either HR or RD scale for these secondary endpoints (Table 2).

Clinical Outcomes Associated with Rivaroxaban vs Warfarin by Frailty Levels

Over a median follow-up (IQR) of 82 (33, 156) days, the rate of the composite endpoint (per 1,000 person-years) was 77.8 for rivaroxaban initiators and 83.7 for warfarin initiators (HR [95% CI], 0.98 [0.94, 1.02]; RD [95% CI] per 1,000 person-years, −5.9 [−9.4, −2.4]) (Table 2 and Supplement 2, Table 9). On the HR scale, rivaroxaban was associated with a lower rate of the composite endpoint than warfarin for non-frail subgroup (HR, 0.88 [0.77, 0.99]) but not for pre-frail or frail subgroup (P_{heterogeneity}=0.026) (Table 2 and Figure 1B). On the RD scale, rivaroxaban was associated with fewer composite events among non-frail (RD, −5.6 [−9.9, −1.3]) and frail (RD, −19.0 [−35.2, −2.9]) subgroups (P_{heterogeneity}=0.040).

For secondary endpoints, rivaroxaban initiators had lower rates of death (HR [95% CI], 0.91 [0.86, 0.97]; RD [95% CI], −5.7 [−8.2, −3.2]), ischemic stroke (HR, 0.71 [0.61, 0.83]; RD, −2.6 [−3.6, −1.5]), and intracranial bleeding (HR, 0.62 [0.53, 0.72]; RD, −3.6 [−4.7, −2.5]); and higher rates of major bleeding (HR, 1.09 [1.03-1.17]; RD, 1.4 [−1.0, 3.8]) and major gastrointestinal bleeding (HR, 1.40 [1.29, 1.52]; RD, 6.2 [4.4, 8.1]) (Table 2 and Supplement 2, Table 9). There was statistically significant evidence for heterogeneity by frailty for major bleeding and major gastrointestinal bleeding on both the HR and RD scales (Table 2). The positive association between rivaroxaban and bleeding events seemed stronger for pre-frail subgroup. No statistically significant heterogeneity was found for death, ischemic stroke, and intracranial bleeding.

Clinical Outcomes Associated with Apixaban vs Warfarin by Frailty Levels

Over a median follow-up (interquartile range) of 84 (33, 157) days, the rate of the composite endpoint (per 1,000 person-years) was 60.1 for apixaban initiators and 92.3 for warfarin initiators (HR [95% CI], 0.68 [0.65, 0.72]; RD [95% CI] per 1,000 person-years, −32.2 [−36.1, −28.3]) (Table 2 and Supplement 2, Table 10). This beneficial association for apixaban seemed consistent across frailty subgroups on the HR scale (P_{heterogeneity}=0.080) but greater for frail subgroup on the RD scale (P_{heterogeneity}<0.001) (RD for non-frail, pre-frail, and frail groups: −15.7 [−20.4, −11.0], −31.8 [−36.7, −26.9], and −68.7 [−85.3, −52.0], respectively) (Table 2 and Figure 1C).

For secondary endpoints, apixaban initiators had lower rates of death (HR [95% CI], 0.82 [0.77, 0.88]; RD [95% CI], −10.3 [−13.2, −7.3]), ischemic stroke (HR, 0.63 [0.54, 0.74]; RD, −4.0 [−5.3, −2.7]), major bleeding (HR, 0.51 [0.46, 0.55]; RD, −19.3 [−21.7, −17.0]), major gastrointestinal bleeding (HR, 0.52 [0.46, 0.58]; RD, −10.6 [−12.3, −8.8]), and intracranial bleeding (HR, 0.51 [0.43, 0.60]); RD, −5.1 [−6.3, −3.8]) (Table 2 and Supplement 2, Table 10). While the associations were consistent across frailty subgroups on the HR scale, RD estimates were greater as frailty increased (P_{heterogeneity}<0.05 for all secondary endpoints) (Table 2).

Sensitivity Analyses

Sensitivity analyses showed that the HR estimates under different design and analytic scenarios were qualitatively consistent with the primary analysis results (Supplement 2, Tables 11-13 and Supplement 2, Figures 4-6). The HR estimates for the composite endpoints varied minimally under different scenarios. There were greater variations in the HR estimates for some individual endpoints. In general, compared with the primary analysis estimates, HR estimates from sensitivity analyses were similar or larger when we compared standard DOAC doses with warfarin; similar when we used a 3-day grace period and exposure risk window; and attenuated when we applied the intention-to-treat approach. Applying more restrictive exclusion criteria of a clinical trial tended to produce somewhat stronger associations. The associations of dabigatran vs warfarin and of rivaroxaban vs warfarin were sensitive to the presence of moderate or strong unmeasured confounding, whereas the association of apixaban vs warfarin was robust to most scenarios of unmeasured confounding (Supplement 2, Table 14). E-values for non-null associations were 1.77 for dabigatran vs warfarin in non-frail subgroup (HR, 0.81); 1.53 for rivaroxaban vs warfarin in non-frail subgroup (HR, 0.88); and 2.08-2.66 for apixaban vs warfarin (HR ranging from 2.08 to 0.66).

DISCUSSION

Real-world database studies of anticoagulation therapy in AF using Medicare(27, 28, 38-40) and commercial insurance claims(40-45) have shown results that are generally consistent with clinical trials.(1-3) However, lack of information on frailty in real-world databases and clinical trials does not allow evaluation of the benefits and harms of DOACs vs warfarin across frailty levels. While frailty has long been a key clinical concept that should be taken into consideration before initiating high-risk medical or surgical treatments, there is little empirical evidence from clinical trials to guide how the treatment should be modified based on frailty information.(13, 46) Our study exemplifies how CFI can be used to evaluate heterogeneity in the emulated treatment effectiveness by frailty in a Medicare claims-based comparative effectiveness and safety study.

We found that apixaban was associated with a 32% relative reduction in the hazard of the composite endpoint of death, ischemic stroke, or major bleeding compared with warfarin in the overall population and consistent 27-39% relative reductions across frailty subgroups. Although dabigatran and rivaroxaban were not associated with a lower hazard of the composite endpoint than warfarin in the overall population, we found a 12-19% relative reduction in the hazard for non-frail subgroup and little or no reductions for pre-frail or frail subgroups. The strength of association on the RD scale was proportional to the absolute event rates across frailty categories. In the dabigatran cohort, HRs of 0.81 (non-frail) and 1.09 (frail) corresponded to 6.1 fewer and 9.8 excess composite endpoints per 1,000 person-years, respectively. In the rivaroxaban cohort, HRs of 0.88 (non-frail) and 0.96 (frail) were comparable to 5.6 and 19.0 fewer composite endpoints, respectively. In the apixaban cohort, HRs of 0.61 (non-frail) and 0.73 (frail) were consistent with 15.7 and 68.7 fewer composite endpoints, respectively. This beneficial association for apixaban vs warfarin in the frail subgroup appeared to be mainly driven by a large reduction in major bleeding, which occurred 4 to 5 times more than ischemic stroke in our study. Because older adults with frailty are predisposed to decline in renal function, higher renal clearance of dabigatran (80-85%) and rivaroxaban (66%) vs apixaban (27%) can result in higher plasma drug concentration, thereby increasing bleeding risk.(47) Our findings are consistent with the results from a previous commercial claims database study of frail older adults with AF (major bleeding at 1 year of follow-up, dabigatran vs warfarin: HR [95% CI], 0.92 [0.62, 1.37]; rivaroxaban vs warfarin: HR 1.06 [0.81, 1.39]; apixaban vs warfarin: HR, 0.61 [0.39, 0.93]).(48) Our study provides more precise estimates by frailty using Medicare data.

While our study suggests that frailty may inform choice between a DOAC and warfarin for older adults with AF, there are important caveats in applying our results to clinical practice. We emulated 3 target trials, each comparing a DOAC with warfarin. Thus, the 3 populations had notable differences, and direct comparison across the trials among DOACs was not possible. The warfarin group in the apixaban cohort had the highest rate of the composite endpoint (92.3 per 1,000 person-years) and the warfarin group in the dabigatran cohort had the lowest rate (65.6 per 1,000 person-years). Because PS matching created a target population that resembles the DOAC initiators within each cohort, our emulated treatment effectiveness estimate represents the effectiveness of the DOAC vs warfarin among the patients treated with the DOAC (“treatment effect among the treated”).

We defined frailty using a validated CFI because a clinical frailty assessment was unavailable in the Medicare database. Since our CFI was developed using a deficit-accumulation frailty index as the reference standard,(16-18) the CFI value is interpretable as a frailty index value calculated from a clinical assessment (non-frail: <0.15, pre-frail: 0.15-0.24, mildly frail: 0.25-0.34, moderately frail: 0.35-0.44, and severely frail: ≥0.45).(21-23) However, due to lack of time and resources, frailty assessment is not routinely performed in clinical practice.(46) Busy clinicians can adopt the Clinical Frailty Scale,(49) a clinical judgment-based scale that can be administered within 3 minutes. A Clinical Frailty Scale category of 5 or higher corresponds to a frailty index ≥0.25. Alternatively, an online frailty index calculator is available (https://www.bidmc.org/research/research-by-department/medicine/gerontology/calculator) to estimate a frailty index based on a comprehensive geriatric assessment.

Strengths and Limitations

The main strengths of our study include novel application of a CFI to evaluate heterogeneity of treatment effectiveness by frailty in a claims-based pharmacoepidemiologic study; emulation of a clinical trial design(50) by applying inclusion and exclusion criteria of a trial, new-user, active comparator design, and PS matching to achieve balance in a large number of covariates; sensitivity analyses to test robustness of our findings to various study design assumptions; and generalizability of our results to older Medicare beneficiaries in routine practice.

Our study also has limitations. In a claims-based study, we were unable to measure clinical variables (e.g., International Normalized Ratio, creatinine, and aspirin use) that influence the effectiveness and safety of oral anticoagulants. Despite PS matching, unmeasured confounding remains possible. The different associations by CFI might reflect a different degree of unmeasured or residual confounding by frailty subgroups. Our sensitivity analysis suggests that our results from the dabigatran and rivaroxaban cohorts were sensitive to moderate unmeasured confounding, while the results from the apixaban cohort were robust to strong unmeasured confounding. Moreover, the short median duration of follow-up (dabigatran: 72 days; rivaroxaban: 82 days; apixaban: 84 days), measured from prescription fill records, did not allow assessment of long-term associations. Given these limitations, the quality of evidence from our observational study can be rated moderate for the apixaban results and low for the dabigatran and rivaroxaban results.

Conclusions

Our study provides evidence to guide choice of a DOAC vs warfarin for older adults with AF. Only apixaban was consistently associated with lower rates of the composite endpoint of death, ischemic stroke, and major bleeding than warfarin across all frailty levels. Our novel approach to evaluate heterogeneity of the emulated effectiveness of drug therapy using CFI can be extended to claims-based pharmacoepidemiologic studies to generate evidence that is unavailable from clinical trials.