Safety of direct oral anticoagulants in older adults with atrial brillation: a systematic review and meta- analysis of (subgroup analyses from) randomized controlled trials

Katharina Doni University of Cologne: Universitat zu Koln Stefanie Bühn Witten/Herdecke University Alina Weise Witten/Herdecke University Nina-Kristin Mann Universität Witten/Herdecke: Universitat Witten/Herdecke Simone Hess Witten/Herdecke University Andreas Sönnichsen Medical University of Vienna: Medizinische Universitat Wien Dawid Pieper University Witten/Herdecke Petra Thürmann University Witten/Herdecke Tim Mathes (  Tim.Mathes@uni-wh.de ) University Witten/Herdecke https://orcid.org/0000-0002-5304-1717


Introduction
Balancing stroke prevention and risk of bleeding in patients with atrial brillation (AF) is challenging.
Vitamin K antagonists (VKA) have been the main treatment for stroke prevention in AF patients in the past. However, the huge inter-individual variability of the clinical response, the necessity of monitoring the INR (International Normalized Ratio) and the quite unmanageable spectrum of food and drug interactions are major disadvantages of VKAs [1].
Various patient related factors, in particular renal dysfunction, hepatic impairment and body weight can impact the pharmacokinetics of DOACs and consequently the risk for adverse events, such as major bleeding [12]. Studies under routine care conditions indicate that the real-world population differs from the one in RCTs with respect to these characteristics. These differences could have a signi cant impact on the bene t-risk ratio of DOACs [13]. Most conspicuous is that the population in real-world data-based studies is about ten years older than in RCTs [14][15][16][17][18]. As pharmacokinetics of DOACs are different in older adults when compared to younger patients, safety analysis of DOAC use in the elderly is of major interest [19].
A systematic review of RCTs and observational studies suggest superior effectiveness and similar safety of DOACs compared to VKAs and that apixaban probably has the best safety pro le in geriatric patients (≥ 75 years) [20]. Likewise, recent observational studies based on real-world data suggest that DOACs are not associated with an increased bleeding risk compared to VKAs but results appear to depend on the speci c DOAC and are heterogeneous across countries [14][15][16][17][18].
The previous evidence on the safety of DOACs in older adults mainly stems from observational evidence and therefore must be interpreted with caution because of the risk of confounding bias. Noticeably, confounding by indication for safety outcomes would mean that patients at higher risk for adverse events when using DOACs would have a lower chance to get a DOAC prescribed and consequently would mean a bias towards the null effect, i.e. would suggest no safety concerns [21].
Our objective was to assess the safety of long-term intake of DOACs in older adults with AF. Our analyses are based on data from RCTs or subgroup analyses from RCTs on older adults (≥ 65 years) to increase the applicability of the results to patients in routine care.

Methods
We registered the protocol for this review in PROSPERO: CRD42020187876. All changes to the protocol are explicitly reported in the methods section. This systematic review was performed according to the recommendations of the Cochrane Handbook for Systematic Reviews of Interventions [22] and follows the reporting recommendations of the updated Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement [23].

Eligibility criteria
Participants Eligible participants must be diagnosed with atrial brillation (AF) and above the age of 65 years. We operationalized the age criterion as follows: ≥80% of the randomized population aged ≥65 years.
Subgroup analysis reports on participants aged ≥65 years.

Intervention
The intervention group must be treated with any type of non-vitamin K antagonist oral anticoagulant.
These include: apixaban dabigatran edoxaban rivaroxaban We only included trials with long-term DOAC treatment, de ned as a treatment duration of at least 12 months. This criterion was added during study selection because different from our expectation, we recognized that in some, mainly early phase RCTs DOACs treatment was very short, which is not comparable to routine care. Any dose or regimen was eligible. Trials on DOACs not approved in the European Union before 2020 (e.g., ximelagatran, darexaban, or letaxaban) were excluded.
As comparator, we accepted any active control such as conventional anticoagulation treatment, and no treatment, or placebo treatment. Furthermore, additional antithrombotic treatment in combined regimens (i.e. antiplatelet therapy in addition to warfarin) had to be the same in all groups, so that the groups only differed regarding DOAC treatment.

Outcomes
We prioritized all-cause mortality, all-cause hospitalization, and major or clinically relevant bleeding (MCRB) as primary outcomes (critical outcomes in GRADE). Secondary outcomes were any adverse event, discontinuation due to adverse events, renal failure, delirium, and falls (important outcomes). In addition, we extracted data on bleeding according to organ system classi cation.
We did not consider stroke or systemic embolism because we expected that the effectiveness of DOACs for reducing stroke is stable across age groups [24,25] and consequently the subgroup effect of age would not shift the bene t-risk ratio.

Types of studies
Only RCTs or subgroup analyses of RCTs on the relevant age group were eligible.

Publication status
We only included trials published in English or German or with data available in an English language trial registry.

Information sources
The identi cation of relevant literature comprised two stages.
First, we screened the titles/abstracts of the references of all systematic reviews included in an overview previously prepared by the research group of one member of our review team [24].
Second, we updated the electronic literature searches used in the aforementioned overview. For this purpose, MEDLINE, MEDLINE in Process, and Embase (all via Embase) were searched for studies published from 1st June 2014 onwards. We ran the last search on 9th November 2020.
In addition, we searched the reference lists of all included RCTs and systematic reviews on the same topic. Moreover, we searched ClinicalTrials.gov for ongoing and unpublished trials on 30 June 2020.

Search strategy
The search strategy was prepared by an experienced information specialist in collaboration with clinical experts. The full search is presented in supplement I. The search was limited to English and German. In addition, we limited the search to articles and reviews (i.e., excluded conference abstracts) and excluded case reports, in vitro studies and animal experiments. The search included a search lter for the elderly, a modi ed generic search lter (in addition to speci c terms such as bleeding or mortality) for adverse events and a validated search lter for RCTs [26][27][28]. The search strategy was reviewed by a second person using the PRESS-checklist and validated by checking if clearly eligible RCTs already known would have been identi ed [29].

Selection process
Two reviewers independently screened the titles and abstracts of all records identi ed by the literature search. Next, full-text articles of potentially relevant reports were retrieved and assessed for compliance with the eligibility criteria by two reviewers independently. Disagreements between reviewers were resolved by discussion until consensus.
Multiple reports of the same RCT were merged, so that each trial is the unit of analysis. The study selection process was summarized in an updated PRISMA ow diagram [42].

Data collection process
Descriptive data were extracted by one reviewer and checked for accuracy by a second reviewer. Two reviewers independently identi ed relevant outcome data by marking the section in the relevant source. Subsequently, one reviewer extracted the data, and a second reviewer checked its correctness. All disagreements were resolved in discussions until consensus.
In case of missing data or inconsistent data on primary outcomes in different sources, we contacted the corresponding author by e-mail.

Data items
Supplement II lists all items for which we extracted data.
We extracted data on outcomes for the last available follow-up, i.e. the longest observation period.
Supplemental to the outcome data, we extracted data on within study subgroup analyses. We only extracted data if the relevant subgroup analysis was pre-speci ed and a test of interaction was used to quantify the statistical certainty of the subgroup effect [30].

Study risk of bias assessment
We assessed the risk of bias with the revised Cochrane risk-of-bias tool for RCTs (RoB 2 tool) [31]. The RoB 2 tool provides a framework for assessing the risk of bias for one particular outcome that is for each outcome separately.

Effect measures
All considered outcomes were dichotomous. We extracted relative risk ratios from regression analyses (e.g., hazard ratios from a survival analysis) with 95% CIs. If these were not available (e.g., data from trial registries), we extracted raw data on events and number of participants for each group and calculated relative risks.

Statistical synthesis method
We pooled data only if RCTs were su ciently clinically and methodologically homogenous and the pvalue of the statistical test for heterogeneity was >0.05. To describe statistical heterogeneity, we calculated prediction intervals and I-square.
We pooled adverse event data separately for each comparator (VKAs, Aspirin only, Placebo) and dose because we assumed, they would have different risks, in particular for bleeding. We calculated systemic adverse events across AF patients (AF-only patients) and AF patients who had a percutaneous coronary intervention (AF-PCI patients), provided the patients were clinically comparable otherwise (e.g. renal function, comorbidity).
Mortality and hospitalization are composite outcomes, to be concrete measures that combine bene ts (e.g. stroke reduction) and harms (e.g. bleeding). Therefore, for mortality and hospitalization we combined different comparators because we were interested in the net bene t of DOACs compared to all possible treatments that are applied in routine care. Moreover, we pooled mortality and hospitalization separately for AF and AF-PCI patients because the bene ts of DOACs (e.g., stroke prevention) likely differ between AF and AF-PCI patients.
We derived the log standard errors, which are necessary for meta-analysis from the 95% con dence intervals (95% CIs). If more than one distinct subgroup for older adults was available (e.g. 65-74 years and ≥75 years), we pooled the results within one RCT using xed effect meta-analysis. To combine different RCTs, we performed inverse variance random effects meta-analyses using the Hartung-Knapp method and the Paule-Mandel heterogeneity variance estimator [32,33]. For outcomes for which only sparse data were available (event rate <5%, zero event studies, less than four RCTs in meta-analysis) we planned to use beta-binomial regression models for sensitivity analyses [34,35].
We used the R-Package Meta in R 9.4 for the meta-analyses [36]. In case of heterogeneity, we synthesized results across RCTs presenting range of effects of the point estimate of the relative risk ratio.

Subgroup analyses for exploring heterogeneity
We expected that our primary analyses would be mainly based on data from subgroup-analyses, and we had therefore not planned to perform subgroup analyses. However, in some meta-analyses there was statistically signi cant heterogeneity, and therefore we performed post-hoc subgroup analyses on study level according to agent.

Sensitivity analyses
We planned to perform a sensitivity analysis excluding RCTs at high risk of bias in the randomisation domain.

Reporting bias assessment
We planned to assess publication bias by visual inspection of funnel plots for asymmetry, if at least 10 trials for each outcome were available.
We expected adverse events and mortality to be assessed in all RCTs. We considered RCTs/publications speci cally on older adults in which mortality, overall adverse events, or discontinuation due to adverse events were not reported (and for which we got no information in response to author requests) susceptive for reporting bias. Bias in selection of the reported results within one trial is a domain of the RoB2 tool (see above). In the RoB2 assessment, we compared the list of outcomes reported in the protocols or methods section with the outcomes reported in the published paper.

Certainty of evidence assessment
We rated the certainty of the body of evidence using the GRADE approach (Grading of Recommendations, Assessment, Development and Evaluation). In the GRADE system evidence from RCTs starts as "highcertainty" and the following criteria are applied for downgrading the certainty of evidence by one or two levels [37]:

Risk of bias
Imprecision Inconsistency Indirectness

Publication bias
The rating of these criteria leads to four levels of the certainty of evidence for each of the prioritized outcomes [38]: High-certainty evidence: the review authors have a lot of con dence that the true effect is similar to the estimated effect.
Moderate-certainty evidence: the review authors believe that the true effect is probably close to the estimated effect.
Low-certainty evidence: the review authors believe that the true effect might be markedly different from the estimated effect.
Very low-certainty evidence: the review authors believe that the true effect is probably markedly different from the estimated effect.
One reviewer judged the certainty of the evidence and a second reviewer veri ed the assessment.
Disagreements were resolved by discussion until consensus.
The certainty of evidence and results are presented in 'Summary of Findings' (SoF) tables [39]. The SoF tables were prepared using GRADEpro GDT [40]. For estimating the absolute effect, we used absolute risks for the control group based on publications thought to be representative for routine care in western countries [15,16,18]. If we could not nd a suitable publication for one outcome, we used the risk of the comparator group of included RCTs.
To report the ndings in consideration of the certainty of evidence, we used the standardized informative statements suggested by the GRADE working group [41].
The certainty of evidence is expressed with the following statements: High-certainty: reduces/increases outcome Moderate-certainty: "likely/probably" reduces/increases outcome Low-certainty: "may" reduce/increase outcome Very low-certainty: the evidence is uncertain Figure 2 shows the study selection according to the PRISMA statement [23]. The initial screening of publications included in the previously published overview [24] identi ed 87 potentially relevant RCTs (based on 111 trial reports) of which we screened full-text versions. The electronic search provided a total of 845 citations after duplicate removal. Titles/abstracts of these were screened and 40 potentially eligible study reports were identi ed. The screening of full-text publications yielded ten RCTs (reported in eighteen publications) which met all eligibility criteria [2-9, 42, 43]. The search in ClinicalTrials.gov and the screening of reference lists of included RCTs and relevant systematic reviews did not lead to additional inclusions. A list of excluded studies and the primary reason for exclusion are provided in Supplement III.

Study selection
We contacted nine authors by e-mail for additional information. Four authors responded, and one provided additional numerical data [44]. In addition, we received results of an analysis of subgroup effects from an individual patient data (IPD) meta-analysis of ve of the included RCTs, in response to an author request [2][3][4][5][6]45]. Study characteristics Page 10/33 The eligible RCTs/subgroup analyses of RCTs (in the following all only called RCTs), included 61,948 participants in total. Table 1 shows the characteristics of the included RCTs (for detailed characteristics   see supplemental table IV).
The median or mean age was 70 years or older. All RCTs included more men than women. In all RCTs, a signi cant proportion of the study population had an increased risk of bleeding and suffered from reduced renal function. Average BMI/weight was above the normal in most studies but does not reach severe obesity in any RCT.
All trials were funded by the pharmaceutical industry. Table 2 contains the risk of bias assessment for each individual RCT. Results are presented on study level (not outcome level) because in none of the RCTs the risk of bias differed for different outcomes (e.g., bleeding and falls). For ve RCTs we assessed the overall risk of bias to be low [44,46,48,50,51] and for ve RCTs we had some concerns regarding the overall risk of bias [9,42,43,47,52].

Reporting bias
We could not prepare funnel-plots because none of the meta-analyses included at least ten studies. Three publications focused on the elderly but did not report mortality or any adverse event [48, 51, 52].

Effects of DOACs on the elderly
The results of the meta-analyses and of each individual RCT included in the meta-analyses are shown in the forest-plots (Figs. 2-4 and supplemental gures I and II). Results of the syntheses with certainty of evidence ratings are presented in the Summary of Findings table (Table 3). The RCTs that were not included in the meta-analyses, because they did not match any pre-speci ed comparison, or because of clinical heterogeneity are presented in supplement V.

Major or clinically relevant bleeding
In the meta-analyses, there was statistically signi cant heterogeneity and therefore the results were not pooled across all included RCTs [9,42,46,48,[50][51][52]. This was true for both, the meta-analysis on lowdose and on high-dose DOACs. To explore this heterogeneity, we performed post-hoc subgroup analyses. We decided to stratify the analyses according to agent because previous systematic reviews and large real-world studies had suggested that dabigatran and rivaroxaban tend to have a higher bleeding risk than apixaban and edoxaban [13,15,16,18,20,54].
For low doses, the separate analyses according to agents did not resolve heterogeneity [9,46,48,50,52]. A common quantitative measure would therefore be misleading and consequently no meta-analysis was performed, and we only compiled a narrative synthesis. According to this, low-dose DOACs likely reduce bleeding compared to VKAs (HR ranged from 0.47 to 1.01). Likewise, in end-stage renal disease patients, low-dose rivaroxaban decreased major bleeding risk numerically compared to VKAs (RR 0.58 95%CI 0.25 to 1.34) [44]. In the ELDERCARE trial low-dose edoxaban increased major bleeding numerically compared to placebo (HR 1.87 95%CI 0.90 to 3.89) [43]. In the AVERROES trial, apixaban increased major bleeding risk numerically compared to aspirin, but 95%CIs overlapped appreciable bene t and harm (1.21 95%CI 0.69 to 2.12) [47].
For high-dose DOACs, the distinct meta-analyses according to agent resolved heterogeneity and we pooled the data. The analyses showed that high-dose edoxaban decreases MCRB risk (HR 0.82 95%CI 0.73 to 0.93) [42,48], but that high-dose dabigatran or rivaroxaban increase MCRB risk compared to VKAs (HR 1.15 95%CI 1.02 to 1.30) [50][51][52]. Table 4 shows the results of the within study subgroup analyses for age. The subgroup analyses indicate that the positive effect on mortality in favour of DOACs might decrease with age. The MCRB risk appears to increase with age, whereby the effect direction in favour of DOACs might reverse in very old people (about 85-90 years).

Subgroup considerations
Subgroup-analyses for major bleeding according to all AF patients versus AF-PCI patients [8, 42, 52], do not change the results (data not shown). An explorative analysis of bleeding risk according to body part suggested that DOACs increase the risk of gastrointestinal bleeding but reduce the risk of intracranial bleeding numerically (data not shown).

Secondary outcomes
Apixaban likely reduces overall hospitalisations (HR 0.84 95%CI 0.76 to 0.93) [46,47]. In the ELDERCARE trial the difference in hospitalizations was negligible (RR 1.02 95%CI 0.67 to 1.58). We did not nd any RCT that reported on hospitalisations in AF-PCI patients. There is no evidence from RCTs in general on overall adverse events, discontinuation due to adverse events, renal failure, falls or delirium in elderly patients with AF treated with DOACs.

Sensitivity analyses
We performed no sensitivity analysis according to risk of bias because none of the RCTs was assessed to be at high risk of bias or risk of bias in the randomisation domain.
Sensitivity analyses of meta-analyses including < 4 RCTs and few numbers of events were not possible because the beta-binomial model is a one-stage model, which requires data that allow to reconstruct a contingency table, but for almost all RCTs only aggregated data (e.g., HRs) were available.

Summary and interpretation in consideration of other evidence
Our systematic review shows that DOACs probably reduce mortality in elderly AF-only patients to a larger extent than VKAs. The ndings were consistent across different agents and different doses agree with previous results of RCTs on all age groups, which suggests that the global effectiveness of DOACs in AF is not signi cantly in uenced by age and a positive bene t-risk ratio of DOACs in comparison to VKAs does also exist in the older population with AF-only [13,25,45]. In the population with AF, the lower risk for bleeding in the low-dose treatment groups is apparently not counterbalanced by a higher risk for lethal thromboembolic events. In the high-dose treatment groups a signi cantly higher bleeding risk does exist for dabigatran and rivaroxaban, but not for edoxaban, which might be explained by the different extent of renal elimination.
We found no RCT that reported on mortality for AF-PCI patients, however meta-analyses in the entire population, i.e. not only elderly, showed that mortality in the DOACs group was not statistically signi cantly higher than in the VKA group [42,55]. Apparently, in this population, the approach of combining DOACs with only one antiplatelet agent (instead of dual antiplatelet therapy = DAPT) in comparison to the e cacy and safety with VKA plus DAPT results in a lower risk for bleeding, but a higher risk for thromboembolic or coronary events.
Studies based on real-world data showed heterogeneous results for mortality when using DOACs compared to VKAs in AF patients [14][15][16][17]. The studies neither distinguished between AF-only and AF-PCI patients nor patients with different heart disease severity in general. Remarkably, in these observational studies, morbidity due to cardiovascular disease was high and more similar to the PCI population than to the AF-only population in our review. The differences in morbidity, in particular the probable differences in proportion of PCI-patients, might be one explanation for the heterogeneous results for the effectiveness of DOACs in Studies based on real-world data and also for the tendency of a weaker impact of DOACs on mortality under routine conditions compared to the RCTs on AF-only patients [14,15,55].
Another explanation for the heterogeneous ndings could be the type of VKA used. In large real-world studies performed in the USA and Denmark, taking DOACs was associated with fewer deaths compared to VKAs [16,17]. In contrast, similar studies performed in Germany showed higher mortality compared to VKAs [15]. The reason for this difference could be the different VKA prescribing practices; in the USA, warfarin is mainly, whereas but phenprocoumon is prescribed in Germany. Pharmacological studies showed that for long-term use phenprocoumon is preferable compared to warfarin because phenprocoumon patients more often have an INR in the therapeutic range [56]. Conspicuously, the patterns of mortality and major bleeding risk appear to agree, concrete studies showing higher major bleeding risk tend to show less favourable results for mortality, indicating that at least a part of the differences in mortality might be explained by death as a result of major bleeding.
We found that low-dose DOACs probably decrease MCRB compared to VKA in AF-only and similarly in AF-PCI patients but could not quantify this reduction reliably because of statistical heterogeneity. The heterogeneity could neither be fully explained by subgroup analyses of patient type (AF-only vs AF-PCI) nor by subgroup analysis on drug type. All but one RCT showed reduced MCRB and in one RCT the bleeding rates were comparable between DOAC (dabigatran) and warfarin. In the two RCTs on low-dose dabigatran the PCI trial showed lower major bleeding risk using dabigatran compared to warfarin, which is probably explained by the concomitant therapy with only one antiplatelet agent in the DOAC group and DAPT in the warfarin group. Consistent with real-world studies, apixaban and edoxaban showed the lowest risk for MCRB [15,16].
Preliminary evidence, showed numerically more MCRB when taking low-dose DOACs compared to aspirin and placebo in AF-only patients [43,47]. Furthermore, the RCT of low-dose rivaroxaban in AF-only patients with end-stage renal disease provides a hint that the results for MCRB in this population might be similar to the results in the elderly in general, meaning that the current evidence from RCTs does not indicate that in patients with end-stage renal disease DOACs should not be taken in general [44].
For high-dose DOACs the MCRB risk seems to depend on the agent. This seems to be true both for AFonly and AF-PCI patients. Rivaroxaban and dabigatran increased the bleeding risk. In contrast, high-dose edoxaban reduced the MCRB risk. Again, the real-world studies found heterogeneous results for these drugs [14][15][16][17][18]. However, it must be considered that the impact of different doses of DOACs was not analysed in these observational studies. Considering that the quality of evidence for our ndings is high and considering results of previous analyses on the in uence of dosing, it appears plausible that the different doses are an additional important explanation for the heterogeneous ndings on safety of DOACs in the real-world [13].

Applicability of ndings
Comparing our study population to the patient population from real-world studies con rmed that our population mirrors the patients in routine care quite well. Therefore, none of the RCTs was down-graded due to limited applicability in the certainty of evidence assessment. Notwithstanding, the AF-only patients still tend to be less morbid and comorbid than patients in real-world studies [14-17, 57, 58]. Moreover, all but two studies that compared DOACs to VKAs used warfarin whereas in many countries other VKAs are mainly prescribed, which limits applicability of the results to these countries [59].

Quality of the evidence
The risk of bias of the body of evidence was low. The main limitation of the certainty of evidence for mortality was statistical imprecision. In addition, for low-dose DOACs, the certainty of evidence on MCRB is limited by unexplained heterogeneity.
The evidence in this systematic review is incomplete regarding several safety outcomes including overall adverse events, adverse events leading to discontinuation, and adverse events particularly relevant for the elderly such as delirium or falls.

Limitations
One limitation of this systematic review is the literature search. We decided to identify the evidence using previous systematic reviews to speed up the review process. We anticipated that this is a reasonable shortcut considering the very huge number of systematic reviews on DOACs and therefore low risk of missing relevant literature when relying on previous systematic literature searches. In addition, some might argue that the ndings are limited because a large part of them stems from subgroup analyses from RCTs on elderly. However, most of the RCTs were very large and additionally strati ed the randomisation for age and adjusted the analyses for important prognostic factors. Therefore, it seems improbable that this approach has introduced bias.

Conclusion Implications for research
There is an important research gap on overall adverse events and particularly outcomes that are relevant for older adults such as falls, fractures or renal impairment in AF patients in general [60]. In addition, for AF-PCI patients high quality data on mortality is lacking. Studies on these outcomes are necessary for su cient balancing of the bene ts and harms of DOAC use in elderly patients, especially given the low absolute mortality and MCRB risk. Moreover, patient characteristics which might explain the heterogeneity in the real-world, such as very high age, weight, renal function, severe and multi-morbidity, should be further explored because better information on these potential predictors could contribute to an improved individualization of anticoagulation therapy.

Implications for practice
No conclusive judgement on the safety of DOACs in older adults is possible because of the lack of RCTs assessing overall adverse events and outcomes relevant in the elderly (e.g., fractures, delirium) [60]. Our data and external evidence from real-world studies suggest that the bleeding risk depends on agent, dose and age. Moreover, the impact of DOACs on mortality and hospitalization probably depends on patient type (AF-only vs AF-PCI). Similarly, to previous systematic reviews on all age groups, we found that lowdose DOACs probably decrease mortality in older AF-only patients. Moreover, Apixaban and edoxaban are associated with fewer MCRB compared to VKAs [13]. For dabigatran and rivaroxaban, the risk of MCRB varies depending on dose. Moreover, subgroup analyses indicate that in the very old (≥85) the bleeding risk of DOACs in general, but especially for dabigatran and rivaroxaban might be even higher than for VKAs. The uncertainty due to heterogeneous results and the limited impact of DOACs on absolute mortality, clarify once again that the individual anticoagulation treatment choice should cautiously balance the individual patient's bene t-risk pro le, especially in the very old.    Risk with DOACs *The risk in the intervention group (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: Confidence interval; HR: Hazard Ratio GRADE Working Group grades of evidence High certainty: We are very confident that the true effect lies close to that of the estimate of the effect Moderate certainty: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different Low certainty: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect Very low certainty: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect   Major bleeding high dose edoxaban (n=16.668) Figure 4