Warfarin use and stroke, bleeding and mortality risk in patients with end stage renal disease and atrial fibrillation: a systematic review and meta-analysis

Background Patients with end stage renal disease (ESRD), including stage 5 chronic kidney disease (CKD), hemodialysis (HD) and peritoneal dialysis (PD), are at high risk for stroke-related morbidity, mortality and bleeding. The overall risk/benefit balance of warfarin treatment among patients with ESRD and AF remains unclear. Methods We systematically reviewed the associations of warfarin use and stroke outcome, bleeding outcome or mortality in patients with ESRD and AF. We conducted a comprehensive literature search in Feb 2016 using key words related to ESRD, AF and warfarin in PubMed, Embase and Cochrane Library without language restriction. We searched for randomized trials and observational studies that compared the use of warfarin with no treatment, aspirin or direct oral anticoagulants (DOACs), and reported quantitative risk estimates on these outcomes. Paired reviewers screened articles, collected data and performed qualitative assessment using the Cochrane Risk of Bias Assessment Tool for Non-randomized Studies of Interventions. We conducted meta-analyses using the random-effects model with the DerSimonian - Laird estimator and the Knapp-Hartung methods as appropriate. Results We identified 2709 references and included 20 observational cohort studies that examined stroke outcome, bleeding outcome and mortality associated with warfarin use in 56,146 patients with ESRD and AF. The pooled estimates from meta-analysis for the stroke outcome suggested that warfarin use was not associated with all-cause stroke (HR = 0.92, 95 % CI 0.74–1.16) or any stroke (HR = 1.01, 95 % CI 0.81–1.26), or ischemic stroke (HR = 0.80, 95 % CI 0.58–1.11) among patients with ESRD and AF. In contrast, warfarin use was associated with significantly increased risk of all-cause bleeding (HR = 1.21, 95 % CI 1.01–1.44), but not associated with major bleeding (HR = 1.18, 95 % CI 0.82–1.69) or gastrointestinal bleeding (HR = 1.19, 95 % CI 0.81–1.76) or any bleeding (HR = 1.21, 95 % CI 0.99–1.48). There was insufficient evidence to evaluate the association between warfarin use and mortality in this population (pooled risk estimate not calculated due to high heterogeneity). Results on DOACs were inconclusive due to limited relevant studies. Conclusions Given the absence of efficacy and an increased bleeding risk, these findings call into question the use of warfarin for AF treatment among patients with ESRD.


Background
The prevalence of atrial fibrillation (AF) in adults with end stage renal disease (ESRD) is 11.6 % [1], about 11-times higher than the prevalence of AF in the general adult population [2]. Among patients with ESRD and AF, the incidence of stroke is 5.2 per 100 person-years and the incidence of mortality is 26.9 per 100 person-years. These incidences are notably higher than the incidence of stroke (1.9 per 100 person-years) and the incidence of mortality (13.4 per 100 person-years) in patients with ESRD who do not have AF [1].
Anticoagulation therapy, such as warfarin, is commonly prescribed to prevent ischemic stroke and its efficacy is well demonstrated in a meta-analysis of randomized trials and observational studies in patients with chronic kidney disease (CKD) and AF [3]. Another meta-analysis suggested that using warfarin does not significantly increase adverse bleeding outcomes among patients with AF and mild to moderate CKD [4]. Direct oral anticoagulants (DOACs) including dabigatran, rivaroxaban and apixaban are available as alternatives to warfarin therapy for prevention of stroke and systemic thromboembolism in patients with AF without renal impairment. Some data on DOACs suggested a higher efficacy [3] and lower bleeding risk [4] of these agents compared with warfarin in patients with CKD and AF. However, the product labels stated that the use of dabigatran and rivaroxaban should be avoided in patients with severe renal impairment (i.e. creatinine clearance [CrCl] < 30 mL/min). Previous randomized controlled trials have excluded patients with advanced CKD on dialysis and thus there remains a lack of evidence to support the use of warfarin or DOACs in this population. Despite the wealth of evidence of anticoagulation therapy in patients with CKD, the benefits and risks of warfarin and DOACs in patients with ESRD and AF are unclear.
The guidelines for warfarin treatment in patients with ESRD and AF are not uniform. The current American Heart Association/American College of Cardiology/ Heart Rhythm Society (AHA/ACC/HRS) guideline recommends warfarin for oral anticoagulation in patients with ESRD and nonvalvular AF who have a CHA 2 DS 2 -VASc score of 2 or greater [5]. The Kidney Disease Improving Global Outcomes (KDIGO) guideline suggests that routine anticoagulation in patients with ESRD and AF for primary prevention of stroke is not indicated because of increased risk for bleeding and lack of systematic evidence for stroke prevention benefit, whereas recommendations for secondary prevention and careful monitoring of all patients receiving dialysis anticoagulation remain valid [6]. Recently published systematic reviews which examined the benefit and risk of warfarin in patients with ESRD and AF were limited to patients on hemodialysis (HD) or peritoneal dialysis (PD) [7,8], or used an inappropriate measure of association (risk ratio [RR] instead of hazard ratio [HR]) [9]. Because the risk of outcomes may not remain constant over the study period and loss to follow up are common in observational studies, HR is more appropriate for evaluating the effects of warfarin as most observational studies report time-to-event data. Moreover, these studies reported conflicting results regarding the association between warfarin use and stroke outcome: one review reported warfarin use was associated with higher risk of any stroke (RR 1.50, 95 % CI: 1.13-1.99) [9] while other reviews reported a lack of association between warfarin use and stroke [7,8,10].
Therefore, we expanded the population to stage 5 CKD, HD, and PD and conducted a systematic review and meta-analyses on the benefits and risks of warfarin use. We used appropriate analytic tools such as the Cochrane Risk of Bias Assessment Tool for Non-randomized Studies of Interventions [11] for qualitative assessment and the Knapp-Hartung methods [12] for quantitative assessment. The objective of this study was to review and summarize the associations between warfarin use and stroke outcomes, bleeding outcomes and all-cause mortality, as compared to no warfarin use, aspirin or DOACs, among patients with ESRD and AF.

Search strategy
We performed the systematic review and meta-analysis in adherence to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. First, we searched PubMed, Embase and the Cochrane Central register using synonyms and variations of the following search terms without language or date restrictions: "end stage renal disease", and "atrial fibrillation" and "anticoagulants or warfarin". We used a combination of controlled vocabulary (e.g. MeSH and Emtree), free-text words (i.e. words appearing in the title, abstract or keywords of a database entry), and truncated terms as appropriate for each database (Appendix). All databases were searched from their start date to February 10, 2016. In addition to the electronic database searches, we hand-searched the reference lists of review articles, relevant studies and clinical practice guidelines.

Study selection
We searched for published randomized controlled trials (RCTs) and quasi-RCTs (RCTs in which allocation to treatment was obtained by alternation, use of alternate medical records, date of birth or other predictable methods), and observational studies which examined the benefits and risks of warfarin in patients with ESRD and AF. We included studies of at least 10 patients with ESRD (HD, PD, stage 5 CKD i.e. GFR < 15 mL/min/ 1.73 m 2 ) and with pre-existing or newly diagnosed AF (all types). We also included studies with broader study populations (i.e. CKD) if they reported outcomes separately for participants with ESRD. We included studies which compared warfarin use with placebo, no treatment or other antithrombotic agents (e.g. aspirin, dabigatran, rivaroxban, apixaban). Studies needed to report quantitative data on the risk for any of the following outcomes: all-cause stroke (any stroke i.e. including ischemic stroke, hemorrhagic stroke, systematic thromboembolism and transient ischemic attacks; ischemic stroke), all-cause bleeding (any bleeding, major bleeding, gastrointestinal bleeding) or all-cause mortality. We only reviewed full articles because conference abstracts would not provide the details necessary for qualitative and quantitative assessments.

Data collection
Two reviewer authors (JT, SL) independently conducted abstract screening and selected relevant studies for data abstraction according to the inclusion/exclusion criteria above. We used a web-based systematic review software DistillerSR (Evidence Partners, Ottawa, Canada) to document the article screening process and to develop standardized data collection forms. For each study, we abstracted bibliographic information (first author, publication year); general information (location of study, sample size, number of treatment groups, number of participants); participant characteristics (age, gender, history of stroke and bleeding); interventions (treatment groups); outcomes (definition, analytic method, crude event data, adjusted risk estimates (HR) and their 95 % CIs); and study quality. We also contacted the corresponding authors of four included studies [13][14][15][16] to obtain missing outcome data. We used the Cochrane Risk of Bias Assessment Tool for Non-Randomized Studies of Interventions (ACROBAT-NRSI) [11] to assess the risk of bias because it was designed specifically for non-randomized studies that evaluate effectiveness of interventions. We rated the risk of bias on seven domains at the study level, and rated the overall risk of bias based on the domain with the highest risk of bias. Discrepancies in study selection and data collection were resolved by the two reviewers through discussions and consensus.

Data analysis
We assessed the clinical and methodologic heterogeneity in participant characteristics (i.e. ESRD status, age, gender, comorbidities, prevalent vs. incident warfarin users) and assessments of outcomes (i.e. outcome definitions and analytic methods). We used the Cochran Q test, which follows a Chi-square distribution with n-1 of freedom, with an alpha of < 0.10 to assess the presence of statistical heterogeneity between studies. We also calculated the I 2 statistic, which ranges between 0 and 100 %, to determine the proportion of between group variability that is attributable to heterogeneity rather than chance [17]. If there was evidence for considerable heterogeneity (i.e. I 2 ≥ 80 %), we displayed the risk estimates in a forest plot but did not calculate the overall risk estimates. Otherwise, we conducted meta-analyses using the random effects model with the DerSimonian-Laird estimator [18] and the Knapp-Hartung approach [12], where appropriate, using Stata 14.0 (StataCorp, College Station, TX).

Sensitivity analysis
We performed pre-specified sensitivity analyses where studies with prevalent warfarin users and studies with low methodological quality were excluded from metaanalyses. We assessed the presence of publication bias using funnel plots with the natural log of HR plotted on the y-axis and the standard error of natural log of HR plotted on the x-axis. We also tested the presence of small-study effects using the Egger's test [19]. We conducted meta-regressions with the Knapp-Hartung approach to evaluate the impact of study quality (moderate/high risk of bias vs critical risk of bias), patient population (HD only vs mixed ESRD population), and study design (studies including incident warfarin users only vs. studies including prevalent and incident warfarin users) on stroke outcome, bleeding outcome or mortality.

Description of included studies
We identified 2709 references from the electronic database search; no additional references were identified from hand searching. After removing 593 duplicate references and excluding 2022 references in titles and abstracts screening, we did a full-text review of 94 references. After excluding another 74 references, we included 20 articles for qualitative and quantitative assessment (Fig. 1).
All 20 investigations were observational cohort studies examining the outcomes of warfarin use in patients with ESRD and AF. Nineteen studies compared warfarin use to no warfarin use, while two studies also compared warfarin use to aspirin [20,21] and one study compared warfarin use to dabigatran and rivaroxaban [20]. A total of 56,146 patients with ESRD and AF, including 34,840 on HD, 315 on PD, 610 with stage 5 CKD, and 20,381 mixed ESRD population, were included in these studies. These studies included a median of 690 (interquartile range [IQR] 204-3012) patients with ESRD and AF, and had a median duration of 7.0 (IQR 2.9-9.4) years. Eight studies were based on administrative claims or national/ regional registry data [13,[22][23][24][25][26][27][28], and they generally were longer and larger studies. Twenty studies included participants with a mean age above 60 years old, including two studies that were limited to older adults above 65 years old [22,27]. Nine studies examined effects of warfarin use between incident warfarin users and nonusers [20,[22][23][24][25][26][27][28][29], whereas the other 11 studies compared prevalent warfarin users with nonusers (Table 1, Appendix Table 3).

Association of warfarin with stroke, bleeding and mortality
The meta-analyses of warfarin use included 15 studies that examined all-cause stroke (I 2 68.1 %), 11 studies that examined all-cause bleeding (I 2 48.3 %), and 12 studies that examined all-cause mortality (I 2 85.7 %) and reported HRs as outcome measures (Fig. 2a- Finally, there was high statistical heterogeneity among the 12 studies (I 2 = 85.7 %) that examined all-cause mortality, and thus we did not calculate an overall risk estimate (Fig. 2c). Most studies showed non-significant results except for 4 studies that found lower risk of mortality associated warfarin use [13,16,24,34].    In the meta-regression, we did not find significant impact of study characteristics on outcomes except for study population in the analysis of all-cause stroke outcome. Compared to studies that included mixed ESRD population, studies including only patients on HD reported higher association with all-cause stroke outcome (OR 5.83, 95 % CI 1.22-27.98; P = 0.03). In the funnel plots, we did not observe obvious asymmetry in the funnel plots for all three outcomes (Fig. 3a-c). Statistical tests for small-study effects were not statistically significant for the three outcomes (P = 0.21, 0.51 and 0.68 for all-cause stroke, all-cause bleeding and all-cause mortality respectively).

Discussion
We conducted a systematic review and meta-analyses of 20 observational studies examining the benefits and risks of warfarin use among patients with ESRD and AF. Metaanalyses provided no evidence to suggest associations between warfarin use and all-cause stroke (HR 0.92, 95 % CI 0.74-1.16) among these patients. By contrast, warfarin use was associated with a significantly increased risk of allcause bleeding (HR 1.21, 95 % CI 1.01-1.44). There were insufficient data with good quality to estimate the association between warfarin use and mortality.
We did not evaluate hemorrhagic stroke in the metaanalyses because only two studies reported hemorrhagic stroke as separate outcomes (Appendix Table 4 [22]. We attempted to evaluate warfarin use vs aspirin or DOACs which was not examined in previously published systematic reviews and meta-analyses, but there were not enough studies to draw conclusions regarding these comparisons. For the two studies that examined ischemic stroke outcomes comparing warfarin vs aspirin, one study showed significant increased risk from warfarin (unadjusted rate ratio (RR) 1.23, 95 % CI 1.01-1.52) [20] whereas another study showed a significant reduced risk (adjusted HR: 0.16, 95 % CI 0.04-0.66) [21]. Warfarin was associated with significantly increased risk of major bleeding compared to aspirin (adjusted RR 1.28, 95 % CI 1.19-1.39) [20]. On the other hand, dabigatran (RR 1.48, 95 % CI 1.21-1.81) and rivaroxaban (RR 1.38, 95 % CI 1.02-1.83) were associated with higher risk of hospitalization or death from bleeding when compared with warfarin [20]. In terms of stroke, the authors noted that there were too few events in the study to detect meaningful differences.
Compared to previously published systematic reviews [7-10], our review included several studies recently published [15,16,35] and expanded the study population to include PD [21] and stage 5 CKD [25], which were not included in these reviews [7][8][9]36]. Our pooled estimate of ischemic stroke and bleeding outcomes were consistent in direction and magnitude with those reported by Li et al., Liu et al. and Dahal et al. [7,8,10]. On the other hand, Lee et al. found that warfarin use was associated with increased risk of any stroke (RR 1.50, 95 % CI 1.13-1.99) [9], but our result was not significant (HR 1.01, 95 % CI 0.81-1.26) because we included eight more studies in our meta-analysis [13-15, 23, 26, 28, 31, 35] and abstracted a less extreme risk estimate from one of the studies [9,30]. We would like to point out that we abstracted the results from the intention-to-treat analysis from Shen et al. for the meta-analyses [28], whereas other systematic review abstracted the results from the as-treated analysis from Shen et al. [8]. Such difference did not change the general inferences about the lack of association between warfarin use and stroke and the increased risk of bleeding outcome.
Warfarin acts by inhibiting the synthesis of vitamin K-dependent clotting factors, and its anticoagulation effect is influenced by possible interactions between drugs or foods and warfarin [37]. The effectiveness of warfarin use for stroke prevention is crucially dependent on the quality of anticoagulation therapy, which can be monitored by international normalized ratio (INR) and time in therapeutic range (TTR). Only 5 of the included studies discussed the influence of INR or TTR on the outcome results [13,20,[30][31][32]. Patients with suboptimal warfarin management (e.g. warfarin users who did not receive INR monitoring [30] or patients with TTR < 60 % [13]) have the highest risk for stroke and thromboembolism. Increasing baseline INR level in warfarin users was positively associated with new stroke [30]. On the other hand, patients with CKD and AF treated with warfarin to maintain an INR between 2.0 and 3.0 had a significant reduction in thromboembolic stroke [32]. Higher TTR, an indicator for good warfarin management, had protective effect against bleeding risk [31]. These results highlighted the difficulty in achieving optimal warfarin management in patients with ESRD and AF, which could help explain the heterogeneous outcomes of warfarin use in this population.
Our review has several strengths including our conduct of a comprehensive search in multiple electronic databases with the application of rigorous qualitative and quantitative assessment. We performed our qualitative assessment using the recently developed Cochrane Risk of Bias Assessment Tool for Non-Randomized Studies of Interventions, which was designed specifically for non-randomized studies that compare the health effects of two or more intervention [11], and unlike the Newcastle-Ottawa scale, does not require modification for use in reviews of effectiveness of interventions [38]. We conducted our quantitative assessment using the Knapp-Hartung method based on small-sample adjustments [12], which provided more accurate confidence limits than the DerSimonian-Laird estimator and has been advocated as an alternative method for metaanalysis with a limited number of studies [39,40]. In addition, we were able to obtain missing outcome data from four study authors [13][14][15][16], and thus examined a greater number of studies than previous authors.
Our report also has several limitations. First, we observed high heterogeneity (I 2 = 85.7 %) in all-cause mortality across the 12 studies [13-16, 22, 24, 25, 28-30, 34], which limited our ability to estimate a pooled risk estimate. In the sensitivity meta-analysis of all-cause mortality using studies that only included incident warfarin users [22,24,25,28], the pooled risk estimate was not statistically significant (HR 0.99, 95 % CI 0.79-1.22). Second, we were not able to conduct meta-analysis on the association between warfarin use and hemorrhagic stroke or on the comparison between warfarin and aspirin or DOACs because there were insufficient studies available on this topic. Finally, as with all systematic review and meta-analyses, our results were limited by the quality of the available studies for inclusion. Although all included studies reported warfarin use in patients with ESRD and AF, we could not confirm that such use was indicated for AF treatment because the included studies did not report such information. We could not verify that the ischemic outcomes reported in the included studies were confirmed by imaging, since several studies were based on administrative claims or registry data [13,22,27,28].
We observed substantial clinical and methodological heterogeneity in the studies we examined with respect to participant characteristics, study conduct and outcome assessment. Study population seems to have significant impact on the association between warfarin use and all-cause stroke outcome as evidenced in the metaregression. Compared to 9 studies that included patients on HD only, 6 studies that included mixed ESRD population [13,14,22,23,26] or patients on PD only [41] reported lower association between warfarin use and allcause stroke. This may reflect heterogeneous treatment effects among subgroups of patients with ESRD and requires further investigation. Although meta-regression did not show significant impact on outcomes due to study quality or study design, these characteristics helped explain the heterogeneity observed in the included studies. A majority of the included studies had serious or critical risk of bias, particularly in the bias due to confounding [14-16, 21, 23, 24, 29, 32-35], bias in selection of participants [13,20,29,[31][32][33][34], and bias due to departures from intended interventions domains [13,15,16,20,21,23,24,31,32,34,42]. While all studies attempted to control for confounding bias by covariate adjustment or propensity score adjustment/matching except for one [15], there may be inherent confounding bias due to unobserved covariates, residual confounding or unsuccessful adjustment. Studies that included prevalent [13,15,16,21,29,[31][32][33][34][35]42], rather than new warfarin users, could introduce selection bias because the effect measure was weighted toward prevalent users who had survived the early events [43]. This would underestimate the events that occur early among prevalent users when the risk of treatment-related outcome varies with time [44]. Patients that started on warfarin could discontinue the therapy and thus switched over to the non-use group, leading to bias due to departures from intended interventions [28].

Conclusions
Despite the degree of heterogeneity across studies and the bias in selected studies, our study showed that warfarin use was not associated with a lower risk of ischemic stroke, consistent with recent studies [7][8][9], and was associated with a significant higher risk for bleeding [7][8][9][10]36] among patients undergoing HD. There was insufficient evidence with good quality to estimate the association between warfarin use and hemorrhagic stroke or mortality. Given the limitations of observational studies described above, large randomized controlled trials involving patients with ESRD and AF may be warranted to definitively evaluate the benefits and risks of warfarin. However, we recognize that such study may be too costly to be carried out, so high-quality observational studies are necessary to address the clinical decision dilemma regarding warfarin use in this population. #3 "kidney disease" or "renal disease" or "kidney failure" or "renal failure" or "kidney dysfunction" or "renal dysfunction" or "kidney impairment" or "renal impairment" or "kidney insufficiency" or "renal insufficiency" or ESRD or CKD #4 [Warfarin] explode all trees #14 anticoag* or anti-coag* or "vitamin K antagonist" or VKA or antithrombo* or anti-thrombo* or warfarin or "indirect thrombin inhibitor" or coumadine or dabigatran or "factor Xa inhibitor" or rivaroxaban or apixaban #15 #12 or #13 or #14 #16 #8 and #11 and #15     Bias due to confounding Moderate Analyses were adjusted for risk factors for stroke. These critically important domains were adjusted for using Cox regression analysis, and confirmed by propensity score adjusted analysis.

Appendix
Bias in selection of participants into the study

Serious
Study "excluded patients with <90 d of study enrollment" so there may be some selection bias. "Patient outcomes were followed from the date of analysis initiation", and drug exposure status was determined in the first 90 days of dialysis. Start of follow up and start of intervention do not coincide for all participants.
Bias in measurement of interventions Low Intervention status well defined and based on information collected at the time of intervention.

Low
The primary analysis was intention-to-treat whereby patients were not re-classified, two validation analyses with censoring and time-varying Cox model were used to account for departures from intended interventions. "Similar results were noted when patients were censored when they changed their warfarin, clopidogrel or aspirin prescription after study enrollment." Bias due to missing data Low Variables were identified from computerized medical results, so data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified from the diagnoses obtained from hospital discharge summaries or medical records. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.

Lai 2009 [32]
Bias due to confounding Critical For the stroke outcome, study did not adjust for congestive heart failure. For the bleeding outcome, study did not adjust for liver disease, alcohol use and bleeding history. Adjusted analysis was not applied in comparing incidence of stroke and major bleeding episodes. Confounding inherently not controllable.
Bias in selection of participants into the study Serious All participants who would have been eligible for the target trial were included ("no patients were excluded from the analysis"). Study included prevalent users, so a potential important amount of follow-up time is missing for prevalent users from analyses.

Bias in measurement of interventions Serious
Intervention status well defined but the intervention status was based on current use vs no use determined retrospectively. It may be determined in a way that could have been affected by knowledge of the outcome.

Serious
Bias due to departure from the intended intervention is expected, and is not adjusted for in the analyses.

Bias due to missing data No information
Medical charts of eligible patients treated in a medical center were reviewed. No information about loss to follow up or missing data.
Bias in measurement of outcomes Low Outcomes were identified from the medical charts. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.

Wizemann 2010 [33]
Bias due to confounding Serious Study adjusted for risk factors in Cox regression analysis, but not confirmed by additional analyses.
Bias in selection of participants into the study Critical Study included patients who had pre-existing (i.e. a history of) AF at enrollment and patients who subsequently hospitalized with the diagnosis of AF. Study included prevalent users, so a substantial amount of follow-up time is missing for prevalent users from analyses.

Bias in measurement of interventions Moderate
Intervention is well defined, but some aspects of the assignments of intervention status were determined retrospectively since self-reported medication use status may be subject to recall bias.

No information
Study did not describe the analytical method in detail.

Bias due to missing data No information
Study was based on an international, observational study of HD facilities.
No information about loss to follow up or missing data.

Table 4 Risk of bias assessment of included studies (Continued)
Bias in measurement of outcomes Moderate Outcomes were identified from hospitalization or death records. The outcome measure may be minimally influenced by knowledge of the intervention received by study participants, since study was retrospective in nature.

Serious
The outcome measurements and analyses were not clearly described. Study reported outcomes by age subgroups rather than the entire cohort. Study highlighted significant result in the highest risk age subgroup in the abstract.
Winkelmayer 2011 [22] Bias due to confounding Moderate Study adjusted for risk factors included in the CHADS 2 and HAS-BLED score in the time-fixed Cox regression analysis, and confirmed using propensity score-adjusted analyses which yielded similar results.
Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of intervention and start of intervention coincide for all subjects (30 days after AF hospital discharge).

Bias in measurement of interventions Low
Intervention is well defined, and based solely on information collected at the time of intervention in the regional hospital discharge abstract and drug claims databases.

Bias due to departures from intended interventions
Low Study used intention-to-treat in which patients were "only censored for end of database" in the primary analysis. Study also used as-treated analysis which "censored patients at treatment cross-over".
Bias due to missing data Low Study obtained data from the national patient registry and regional healthcare claims database which contains information on relevant ESRD patients. Data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified using previously validated claims-based algorithm. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Olesen 2012 [23] Bias due to confounding Serious Study adjusted for risk factors included in the CHA 2 DS 2 -VASc and HAS-BLED score in the time-dependent Cox regression analysis, but not confirmed by additional analyses.
Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of intervention and start of intervention coincide for all subjects. "The baseline assessment and follow-up period began 7 days after discharge."

Bias in measurement of interventions Low
Intervention is well defined, and based solely on information collected at the time of intervention in the national patient registry.

Serious
Study uses time-dependent analysis, so bias due to departures from intended interventions is expected. Study did not adjust for switches, co-intervention or problems with implementation fidelity in the analyses.
Bias due to missing data Low Study obtained data from the national patient registry which contains information on all residents. Data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified using claims-based algorithm. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.

Khalid 2013 [34]
Bias due to confounding Critical Analyses did not adjust for important confounders such as hypertension, cerebrovascular disease, diabetes.
Bias in selection of participants into the study Critical Participants re-started warfarin after experiencing bleeding outcomes, so their risk profile may be different. Study included prevalent users, so a potential important amount of follow-up time is missing for prevalent users from analyses.

Bias in measurement of interventions Moderate
Intervention was not well defined. Not clear when the start of intervention was.

Serious
Treatment could be discontinued due to experiencing bleeding outcomes. Study did not address bias due to departures from intended interventions.
Bias due to missing data No information Bias in measurement of outcomes Low Intervention not well defined, but were "adjudicated by two blind reviewers".

Table 4 Risk of bias assessment of included studies (Continued)
Bias in selection of the reported results

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Wakasugi 2014 [29] Bias due to confounding Critical For the stroke outcome, analysis was adjusted for CHADS 2 score and matched by propensity score. For the major bleeding and all-cause mortality outcomes, analyses were not adjusted.
Bias in selection of participants into the study Serious All participants who would have been eligible for the target trial were included. Study included prevalent users, so a potential important amount of follow-up time is missing for prevalent users from analyses.

Bias in measurement of interventions Low
Intervention is well defined, and based solely on information collected at the time of intervention since study was conducted prospectively.

Bias due to departures from intended interventions
Low "The primary analysis was intention-to-treat in which patients who started using warfarin after study enrollment were not reclassified", so the specified comparison relates to initiation of intervention regardless of whether it is continued. To account for possible longitudinal changes in drug prescription over time, study also repeated the analysis in which patients were censored when warfarin use changed.

Bias due to missing data No information
Medical charts of eligible patients treated in a medical center were reviewed. No information about loss to follow up or missing data.
Bias in measurement of outcomes Low Outcomes were identified from the medical charts. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.

Bonde 2014 [24]
Bias due to confounding Serious Analyses were adjusted for risk factors included in the CHA 2 DS 2 -VASc and HAS-BLED score with age as a continuous covariate. These critically important domains were adjusted for using Cox regression analysis, but not confirmed by additional analyses.
Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of follow up and start of intervention coincide for all subjects (follow-up began 7 days after discharge).

Bias in measurement of interventions Moderate
Intervention is well defined, but assignment of intervention status by "dividing the number of tablets dispensed with the estimated daily dosage" was determined retrospectively.

Bias due to departures from intended interventions
Serious "Treatment is often discontinued in terminal patients" so switches and discontinuation of treatment were apparent. Such departures from intended intervention may be a result of outcomes of interest. Study used time-dependent, as-treated Cox regression analysis, but did not adjust for such departures appropriately.
Bias due to missing data Low Accurate data on all patients actively treated for end-stage CKD with RRT recorded in national registry, so data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified from ICD-8 or ICD-10 codes and the same method of outcome assessment was applied across intervention groups.

Serious
Study stratified and analyzed results based on CHA 2 DS 2 -VASc score subgroups. Main result reported the high risk subgroup, which appears to be reported on the basis of significant result.

Carrero 2014 [25]
Bias due to confounding Moderate Study accounted for the most important confounders and confirmed results by propensity score-matched analysis. Residual confounding is inherent in observational studies.
Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of intervention and start of intervention coincide for all subjects (starts with warfarin prescription at discharge).

Bias in measurement of interventions Low
Intervention is well defined, and based solely on information collected at the time of intervention.

Bias due to departures from intended interventions
Low "Warfarin use vs no warfarin use was considered as a time-fixed binary variable throughout the follow-up period." Specified comparison relates to initiation of intervention regardless of whether it is continued as in intention-to-treat (ITT) analysis in target trial.
Bias due to missing data Low "The use of unique personal identification number for all Swedish citizens and continuously updated national registries on death date, cause of death and emigration allow a virtually complete follow-up" with no loss to follow up.

Table 4 Risk of bias assessment of included studies (Continued)
Bias in measurement of outcomes Low Outcomes were identified from the National Inpatient Registry and the same method of outcome assessment was applied across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Chen 2014 [26] Bias due to confounding Moderate Analyses were adjusted for risk factors for stroke. These critically important domains were adjusted for using propensity score matched analysis. Ischemic stroke/TIA history was excluded from the study population. Residual confounding is inherent in observational studies.
Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of intervention and start of intervention coincide for all subjects.
Bias in measurement of interventions Low Intervention status well defined and based on information collected at the time of intervention.

Low
The primary analysis was intention-to-treat, thus the specified comparison relates to initiation of intervention regardless of whether it is continued.
Bias due to missing data Low Variables were identified from the universal national health insurance program, so data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified from the ICD diagnosis codes. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Friberg 2014 [13] Bias due to confounding Moderate Analyses were adjusted for risk factors for stroke and bleeding. These critically important domains were adjusted for using propensity score adjusted analysis. Residual confounding is inherent in observational studies.
Bias in selection of participants into the study Serious Study included prevalent users (taking warfarin at baseline i.e. 5 months before, and up to 1 month after the index AF diagnosis), so an important amount of follow-up time is missing for prevalent users from analyses. Start of follow up and start of intervention do not coincide for all subjects (time at risk for survival analyses started on Day +14 after index AF diagnosis).

Bias in measurement of interventions Low
Intervention status well defined and based on information collected at the time of intervention.

Low
The primary analysis was intention-to-treat, thus the specified comparison relates to initiation of intervention regardless of whether it is continued.
Bias due to missing data Low Variables were identified from the national patient registry, so data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified from the ICD diagnosis codes. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Shah 2014 [27] Bias due to confounding Moderate Study adjusted for risk factors included in the CHADS 2 and HAS-BLED score in the time-fixed Cox regression analysis, and confirmed using propensity score-adjusted analyses which yielded similar results.
Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of follow up and start of intervention coincide for all subjects (30 days after AF hospital discharge).

Bias in measurement of interventions Low
Intervention is well defined, and based solely on information collected at the time of intervention in the regional hospital discharge abstract and drug claims databases.

Bias due to departures from intended interventions
Low Study uses time-fixed, intention-to-treat analysis, so the specified comparison relates to initiation of intervention regardless of whether it is continued.
Bias due to missing data Low Study obtained data from the regional hospital and drug claims databases which contains information on all residents in the region. Data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified using claims-based algorithm. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups. Bias due to confounding Moderate Analyses were adjusted for risk factors for stroke and bleeding. These critically important domains were adjusted for using multivariate Cox regression analysis followed by propensity score matched analysis. Residual confounding is inherent in observational studies.
Bias in selection of participants into the study Serious Study included prevalent users (taking anticoagulant at recruitment) and incident users (starting anticoagulant within 2 weeks following recruitment), so a potential important amount of follow-up time is missing for prevalent users from analyses.
Bias in measurement of interventions Low Intervention status well defined and based on information collected at the time of intervention.

Low
The primary analysis was intention-to-treat, thus the specified comparison relates to initiation of intervention regardless of whether it is continued.
Bias due to missing data No information Patients from 10 hemodialysis center were prospectively followed-up for 2 years. No information about loss to follow up or missing data.
Bias in measurement of outcomes Low Outcomes were identified from medical charts using clinical diagnosis criteria. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Chan KE 2015 [20] Bias due to confounding Moderate Study accounted for the most important confounders and "mitigated potential bias from unmeasured factors by performing a matched analysis which supported the main findings of the study". Residual confounding is inherent in observational studies.
Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of follow up and start of intervention coincide for all subjects (followed from the time the initiated anticoagulant treatment de novo).

Bias in measurement of interventions Low
Intervention status well defined and based on information collected at the time of intervention.

Bias due to departures from intended interventions
Serious Patient was censored with discontinuation of treatment medications. Study used Poisson regression which does not take into account of the potentially varying rate (hazard) ratio for the effect of intervention.
Bias due to missing data Low "All subjects registered in the database are followed longitudinally, where data parameters are actively collected," so data were reasonably complete.
Bias in measurement of outcomes Low The methods of outcome assessment were comparable across intervention groups. Bleeding outcomes were adjudicated retrospectively, but "the adjudicator was blinded to patient, treatment group and outcome".

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Chan PH 2015 [21] Bias due to confounding Serious Study did not list variables adjusted in the multivariate Cox regression analysis. Results were not confirmed by additional analyses.
Bias in selection of participants into the study Critical Study included patients who had pre-existing AF at enrollment. Study included prevalent users, so a substantial amount of follow-up time is missing for prevalent users from analyses. Study excluded patients with incomplete follow-up data, so there may be some selection bias.

Bias in measurement of interventions Low
Intervention is well defined, and based on medical records and discharge summaries at the time of medication prescription.

Bias due to departures from intended interventions
Serious Study compared current users vs non-users. Bias due to departure from the intended intervention is expected, but not adjusted for in the analysis.
Bias due to missing data Low Study data were retrieved from the medical records and discharge summaries from the territory-wide information network of all public hospitals in the region. Study excluded patients with incomplete clinical and/or follow-up data, so data were reasonably complete.
Bias in measurement of outcomes Moderate Outcomes were identified from medical records. The outcome measure may be minimally influenced by knowledge of the intervention received by study participants, since study was retrospective in nature.

Moderate
The statistical analyses were not clearly described. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups. Bias in selection of participants into the study Low All participants who would have been eligible for the target trial were included. Start of intervention and start of intervention coincide for all subjects (30 days after AF hospital discharge).

Bias in measurement of interventions Low
Intervention is well defined, and based solely on information collected at the time of intervention in the national patient registry.

Bias due to departures from intended interventions
Low Study used intention-to-treat and as-treated analyses, and adjusted for selection bias using inverse probability of treatment and censoring-weighted (IPTW) analysis.
Bias due to missing data Low Study obtained data from the national patient registry which contain information on all ESRD patients. Data were reasonably complete.
Bias in measurement of outcomes Low Outcomes were identified using previously validated claims-based algorithm. The methods of outcome assessment were comparable across intervention groups.

Moderate
The outcome measurements and analyses were clearly defined. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Wang 2015 [14] Bias due to confounding Serious Study reported adjusted multivariate analyses for outcomes (all predictors with P < 0.10 listed). Confounding inherently not controllable. Results not confirmed by additional analyses.
Bias in selection of participants into the study Critical Study included prevalent users, so a substantial amount of follow-up time is missing for prevalent users from analyses.

Bias in measurement of interventions Low
Intervention is well defined, and based on medical records and discharge summaries at the time of medication prescription.

Bias due to departures from intended interventions
Serious Study compared current users vs non-users. Bias due to departure from the intended intervention is expected, but not adjusted for in the analysis.
Bias due to missing data No information Study data were retrieved from the medical records in a hospital. No information about loss to follow up or missing data.
Bias in measurement of outcomes Moderate Outcomes were identified from medical records. The outcome measure may be minimally influenced by knowledge of the intervention received by study participants, since study was retrospective in nature.

Serious
Study analyzed multiple outcomes and only reported results for predictors with significant results (P < 0.10) rather than a priori model. There is a high risk of selective reporting from among multiple analyses.
Yodogawa 2015 [35] Bias due to confounding Serious Study adjusted for CHADS 2 in Cox regression analysis, but not confirmed by additional analyses.
Bias in selection of participants into the study Critical Study included patients who had pre-existing AF at enrollment. Study included prevalent users, so a substantial amount of follow-up time is missing for prevalent users from analyses. Study excluded patients with <6 months life expectancy, so the association may be attenuated since the high risk patients were excluded from analysis.

Bias in measurement of interventions Moderate
Intervention is well defined, but some aspects of the assignments of intervention status were determined retrospectively since self-reported medication use status may be subject to recall bias.

No information
Study did not describe the analytical method in detail.
Bias due to missing data Low Study was based on a retrospective, observational study on medical records of patients treated in a hospital. Study excluded patients without follow-up data, so data were reasonably complete.
Bias in measurement of outcomes Moderate Outcomes were identified from medical records. The outcome measure may be minimally influenced by knowledge of the intervention received by study participants, since study was retrospective in nature.

Moderate
The outcome measurements and analyses were not clearly described. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Findlay 2016 [15] Bias due to confounding Critical Study used Kaplan-Meier survival curve to analyze outcomes between warfarin and no warfarin group, and did not use adjusted analysis to control for confounding.

Bias in measurement of interventions No information
Study did not provide information about measurement of interventions.

Bias due to departures from intended interventions
Critical Study did not describe or adjust for any bias due to departures from intended interventions.
Bias due to missing data No information Study did not provide information about missing data.
Bias in measurement of outcomes Moderate Outcomes were identified from the electronic patient record, and was reviewed by 2 independent clinicians. The outcome measure may be minimally influenced by knowledge of the intervention received by study participants, since study was retrospective in nature.

Moderate
The outcome measurements and analyses were not clearly described. There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.
Tanaka 2016 [16] Bias due to confounding Critical Study used Kaplan-Meier survival curve to analyze outcomes between warfarin and no warfarin group, and did not use adjusted analysis to control for confounding.
Bias in selection of participants into the study Critical Study included patients who had pre-existing AF at enrollment. Study included prevalent users, so a substantial amount of follow-up time is missing for prevalent users from analyses.

Bias in measurement of interventions Low
Intervention is well defined, and based on records collected during the period of dialysis initiation. Since study was prospective, assume no recall bias.

Bias due to departures from intended interventions
Critical Study did not describe or adjust for any bias due to departures from intended interventions.
Bias due to missing data Low Study excluded patients lost to follow-up, so data were reasonably complete.
Bias in measurement of outcomes Moderate Outcomes were identified from survey slips that were sent to the dialysis facilities. The outcome measure may be minimally influenced by knowledge of the intervention received by study participants.

Moderate
The outcome measurements and analyses were not clearly described.There is no clear indication of selection of the reported analysis from among multiple analyses or multiple subgroups.  Fig. 5 a Sensitivity meta-analysis of on stroke outcome by warfarin use, excluding studies with critical risk of bias. b Sensitivity meta-analysis of on bleeding outcome by warfarin use, excluding studies with critical risk of bias. c Sensitivity meta-analysis of on mortality outcome by warfarin use, excluding studies with critical risk of bias