The number of patients and patient evaluations in our study was generally greater for major bleeding than minor or any bleeding. None of the CHADS2 covariates had a high strength of association with warfarin-related bleeding risk. The strength of evidence of most covariates was low to very low, with the exception of advanced age and major bleeding. The moderate strength of association between age and major bleeding risk was the strongest among all covariates, whereas the very low strength of association between diabetes and minor bleeding was the weakest. The subgroup analysis in our study revealed that advanced age was also the strongest independent positive predictor of bleeding risk among all CHADS2 covariates in patients specifically with AF. As in the full analysis, the strength of associations between the other covariates and bleeding risk was very low.
Diener et al38
and Lip et al56
published an evaluation comparing the predictive value of 5 different contemporary bleeding risk stratification schemas using combined SPORTIF III and V warfarin data (N=3665). Of the schemas evaluated, the HAS-BLED (Hypertension, Abnormal Renal/Liver Function, Stroke, Bleeding History or Predisposition, Labile INR, Elderly, Drugs/Alcohol Concomitantly) score appeared to be the best predictor of bleeding in patients taking warfarin (C statistic, 0.66; 95% CI, 0.61-0.70); however, all schemas provided only moderate predictive ability (C statistic, 0.52-0.66; a C statistic of 0.50 implies that a schema is no better at predicting a bleeding event than random guessing). The fact that no one schema was vastly superior to another despite each using different risk factors for bleeding is not surprising given the current results of our systematic literature review. The preponderance of the medical literature suggests that many risk factors have either conflicting or scant data supporting their association with bleeding of any severity, clarifying why developing a highly predictive schema has proved difficult. Of the 5 schemas, none used heart failure as a risk factor, 2 (40%) used HTN, 3 (60%) used stroke, and 1 used diabetes by itself (20%). In contrast, age was used as a risk factor in all 5 schemas, albeit in different forms (continuous and categorical formats with different cutoffs for advancing age). The use of advancing age to predict bleeding is very much supported by the findings of our systematic review, which suggest that age is a relatively reliable risk factor for bleeding.
In the evaluation of CHADS2 covariates, it is important to recognize that one covariate may be serving as a surrogate for a different but related risk factor. Therefore, we caution that results must not be taken at face value. For example, the association seen with increasing age and bleeding risk may also be attributed to a decline in renal function, multiple comorbid conditions, or concomitant use of many other drugs. Furthermore, why these risk factors are associated with bleeding risk is often unknown.
Unlike the weak association with bleeding risk, the use of CHADS2
covariates in predicting stroke risk is relatively well established.56,57
Previous studies found that the risk of stroke in patients with AF increased as the number of CHADS2
risk factors increased, with an increase of up to 1.5 times for each 1-point increase in the CHADS2
score when no anticoagulant therapy was given. Warfarin therapy significantly reduced stroke risk, by approximately two-thirds, from 4.5% to 1.4%, when compared with placebo or no treatment, and by two-fifths when compared with aspirin.2-5,56
The Anticoagulation and Risk Factors in Atrial Fibrillation cohort study showed that the highest net benefit of warfarin was among patients with moderate to high risk for stroke because the absolute increase in risk for intracranial hemorrhage due to warfarin remains fairly stable across CHADS2
Several limitations must be considered when interpreting the results of our systematic review. Included studies defined bleeding end points in various ways. Although we limited ourselves to descriptive analysis only (no pooling of results through meta-analysis) and stratified our analyses by bleeding severity (major, minor, or any), important residual heterogeneity may still have been present.
Our systematic review included mostly observational studies, which have inherent selection and information biases that may result in an erroneous estimate of association. Selection bias may have occurred if patients were included in the selected studies because risk factors were presumed to be associated with bleeding. Furthermore, information bias could have occurred if patients were misclassified as having or not having a potential risk factor (ie, HTN, diabetes) or as having or not having a bleeding event. Observational studies can only reveal potential associations between covariates and the risk of bleeding; they cannot prove causality.
Negative reporting bias is a potential limitation of any systematic review. Authors commonly omit insignificant results from their publications, resulting in conclusions biased toward a covariate being a risk factor for bleeding. Of the 41 studies included in our systematic review, 27 (66%) reported insignificant results, suggesting that this is of some potential concern in this systematic review. However, this bias would tend to further weaken the associations we have found. In addition, many of the studies included in this systematic review were underpowered (type 2 error) because the presence and/or occurrence of covariate or bleeding events was low in the patient samples. For example, in the study by Landefeld and Goldman,20
the failure to find CVD as a risk factor for major bleeding may be explained by the low baseline presence of CVD (only 4 of 375 patients) in the overall population. We devised an analysis protocol to accommodate for this, a description of which follows.
Given these cautions, limitations, and caveats, we used a strength-of-evidence rating scale that incorporated many features, allowing decision makers to draw conclusions on the basis of the preponderance of the evidence and reducing the influence of any 1 potentially biased study on the process. First, we evaluated the number of evaluations and patients included because many evaluations allow for the assessment of consistency, and larger evaluations are less likely to be underpowered. We then evaluated both the number and corresponding percentage of the total number of evaluations for a covariate when there was a significant positive association or a number of evaluations with a positive direction of effect regardless of significance. This gave us insight into the effect of a covariate on bleeding and the consistency of the effect across studies. Finally, we evaluated the quality of the evaluations because higher-quality evidence increases confidence in the results.
In light of their expected efficacy, safety, and ease of use, the newer oral anticoagulants (eg, Factor Xa inhibitors, direct thrombin inhibitors) are likely to replace warfarin therapy in many patients. Because the risk of bleeding associated with various patient characteristics is not yet known specifically with these agents, new data will be required. Fortunately, detailed analysis of the large-scale phase 3 clinical trials of these agents should provide much stronger evidence than is available for warfarin.