Background

Atrial fibrillation (AF) is a common cardiac disorder with a prevalence ranging from 0.1% in patients < 55 years old to 9.0% in those ≥ 80 years of age. Patients with AF have a 5-fold increased risk of stroke when compared to those without AF. Long-term anticoagulation with (adjusted-dose) warfarin is highly efficacious in preventing stroke in patients with AF [1–3].

While treatment with warfarin increases the risk of major bleeding, evidence suggests that the benefits of warfarin outweigh the risks in most patients, including the elderly [4]. Despite warfarin's marked efficacy, not all eligible patients receive it for stroke prevention in AF [5].

Numerous researchers have attempted to elucidate why warfarin is not consistently used in patients with AF. Prescriber survey studies have found that not all physicians have a complete understanding of the benefits, risks and risk-to-benefit ratio of warfarin use for stroke prevention in AF [6–10]. Furthermore, observational studies have found that numerous patient characteristics, including (but not limited to) advanced age, gender, socioeconomic status, anticipated or known poor adherence to treatment or inconsistent follow-up, rural residency, perceived, fall risk, language difficulties, disabilities and other known or perceived risk factors for bleeding were independently associated with the under use of warfarin therapy [11–38]. Unfortunately, different studies use different populations, different sample sizes, and different types and numbers of covariates. Evaluating numerous covariates with a smaller population increases the risk that true independent predictors will not be found due to lack of power. As such, not all studies have shown consistent results making it is difficult to get a picture of true independent predictors of warfarin use [5–10].

To more completely identify which prescriber and patient characteristics are associated with warfarin use, and to better quantify the magnitude of the effects of these characteristics, we seek to conduct a systematic review and meta-analysis of the available medical literature evaluating the association between such characteristics and U.S. prescriber use of warfarin for stroke prevention in AF. Having such an analysis available will allow family practice prescribers and other healthcare decision-makers to target interventions or methods to improve utilization of warfarin when it is indicated for stroke prevention in AF.

Methods

Literature Search and Data Abstraction

Two independent investigators conducted systematic literature searches of the MEDLINE and EMBASE (earliest possible date through October 2010) computerized databases. The MeSH terms and keywords: warfarin, coumadin, coumarins, vitamin k antagonist, coumatetralyl, phenprocoumon, dicoumarol, tioclomarol, phenindione, clorindione, fluindione, diphenadione and indandione, along with atrial fibrillation were used. The utilized search strategy is included in See additional file 1: Search Strategy. Additionally, the American College of Chest Physicians (ACCP) guidelines [39, 40] were reviewed along with the references of each pertinent article identified to locate other relevant published works.

Studies were eligible for inclusion in the systematic review if they: (1) reported on a population of patients with atrial fibrillation, (2) reported on the relationship between prescriber and/or patient characteristics and the odds of receiving warfarin for stroke prevention, (3) conducted multivariate analysis to determine the relationship between prescriber and/or patient characteristics and the odds of receiving warfarin therapy, (4) enrolled patients treated in the United States only, (5) were published in English language, and (6) was published no earlier than 1996. The year for this later criterion was based upon the year of publication of the seminal study of warfarin risks and benefits by Hylek and colleagues published in the New England Journal of Medicine [41].

Through the use of a standardized data abstraction tool, two reviewers independently determined whether or not an article was to be included in the systematic review and collected data, with disagreement resolved through discussion. For each included study, data on the following was abstracted: author, year, study design, sample size, population and setting, exclusion criteria, whether studies were restricted to "ideal candidates" for warfarin or not (defined patient populations without warfarin contraindications), timing and duration of study period, (percent of patients receiving warfarin, manner of determining warfarin use, p-value for the univariate relationship between a prescriber or patient characteristic and warfarin use, and effect size and p-value for the multivariate relationship between a prescriber or patient characteristic and warfarin use.

Validity Assessment

Validity assessment was performed using the methodology utilized by the Agency for Healthcare Research and Quality (AHRQ) Evidence-based Practice Center program [42]. For the purposes of validity assessment in this systematic review, an evaluation was defined as an assessment of a prescriber or patient characteristic for its association with warfarin use (thus a single study likely included evaluations of multiple different characteristics). Each evaluation included in a study was separately assessed for the following individual criteria: hierarchy of study design, total, characteristic and warfarin use sample sizes, participant selection method, exposure measurement method (warfarin use), potential design biases, and appropriate analyses to control for confounding. Each evaluation in all identified studies were then be given an overall score of good, fair or poor as described in additional file 2: Three Summary Ratings of Quality of Individual Studies.

Data Synthesis

Results of our systematic literature search were first summarized qualitatively using descriptive statistics. Qualitative synthesis consisted of detailed evidence tables and figures (stratified by CHADS2 criteria [43], characteristics listed in the black box warning of Coumadin's® prescribing information, contraindications or strong precautions listed in the prescribing information, and "other" characteristics [44]) demonstrating the number, overall conclusions, and assessed validity of evaluations stemming form multivariate analyses. We also assessed the change in warfarin use over time (defined as earliest year of patient inclusion) through utilization of linear regression analysis and report the Pearson's r value and its corresponding p-value.

We also undertook traditional meta-analysis for each prescriber and patient characteristic with at least 2 studies reporting data. We calculated weighted averages of effect size as pooled (adjusted) odds ratios (ORs) with associated 95% confidence intervals (CIs) using a DerSimonian and Laird random-.effects model. When studies reported results from overlapping populations (either in same of separate publications), we preferentially used the most recent data for meta-analysis, followed by the largest sized population if study period could not be used. The likelihood of statistical heterogeneity was assessed for in each analysis using the I2 statistic and Cochrane Q statistic p-values (either an I2 > 50% and a Cochrane Q statistic p < 0.10 were considered representative of important statistical heterogeneity). Egger's weighted regression statistic p-values were used to assess for the likelihood of publication bias. Traditional meta-analysis statistics were performed using StatsDirect statistical software, version 2.7.6 (StatsDirect Ltd, Cheshire, UK). A p-value less than 0.05 will be considered statistically significant for all analyses.

Grading the Strength of Evidence

We used the Grading of Recommendations Assessment, Development (GRADE) system to assess the strength of evidence [45]. This system uses four required domains - risk of bias, consistency, directness, and precision. Strength of evidence grade was determined for each association between prescriber and patient characteristics and warfarin use for stroke prevention in AF. The evidence pertaining to each prescriber and patient characteristic was classified into four broad categories: (1) "high", (2) "moderate", (3) "low" grade or "insufficient" as described in additional file 3: Definitions for Grading the Strength of Evidence.

Results

Study Identification and Characteristics

A summary of the results for our literature search are presented in additional file 4: Study Flow Diagram. A total of 1,060 non-duplicate citations were identified, of which 208 citations were retrieved for full-text review. Of these, 28 articles published between 1996 and 2010, representing 33 unique analyses met our inclusion criteria [11–38]. (Table 1). Of note, the studies by Stafford 1996, Brass 1998, Smith 1999 and Brophy 2004 each reported the results of 2 multivariate analyses on overlapping populations [16, 18, 33, 35]. Moreover, the studies by Schauer 2007 and Johnston 2003 were conducted in the same database and with overlapping, but not similar time frames [23, 32]. The studies by Antani 1996 and Beyth 1996 were conducted in the same population (although the analysis in the Beyth paper has slightly fewer patients due to incomplete data collection) [13, 14]. Finally, Lewis 2009 also conducted 2 multivariate analyses; however, the populations used were mutually exclusive and thus these analyses were treated as unique data points in our report [24]. Of the 28 studies identified, 4 were conducted in a prospective and 24 in a retrospective fashion. Sample sizes of studies ranged from 117 to 44,193 patients. While studies were published between 1996 and 2010, they evaluated patients treated for atrial fibrillation between 1980 and 2008. Five studies (17.9%) reported results of analyses restricted only to patient populations without warfarin contraindications ('ideal candidates') [18, 24, 25, 29, 30].

Table 1 Characteristics of Included Studies Evaluating the Association Between Covariates and Warfarin Use for Stroke Prevention in Atrial Fibrillation
Full size table

Results of Qualitative Synthesis

Warfarin use across included studies ranged from 9.1% to 79.8%, with a median of 49.1%. Linear regression analysis on the 23 studies providing data on warfarin use suggests that there is a statistically significant, moderately strong (per Cohen's Rule of Thumb) [46] correlation (r = 0.60) between warfarin utilization and progressing time. (Figure 1) This finding was not significantly changed when studies enrolling 'ideal candidates' only were excluded (data not shown).

Figure 1
figure 1

Result of a Linear Regression Analysis Evaluating the Correlation Between Warfarin Use Over Progressing Time. Dotted lines represent timing of seminal warfarin publications (Hylek 1996 and 2001 Update of the AHA/ACC Atrial Fibrillation Guidelines) [42, 44]. As studies were plotted on the horizontal axis based upon the first year of patient inclusion, data on warfarin use prior to 1996 is depicted. r = Pearson's correlation coefficient.

Full size image

See additional file 5, 6, 7 and 8: Figures Depicting the Number, Validity and Statistical Conclusions of Studies Evaluating Associations Between Characteristics and Warfarin Use and additional file 9: Tables Depicting the Association Between Covariates and Warfarin Use for Stroke Prevention in Atrial Fibrillation) for figures and tables depicting the number, validity and statistical conclusions of identified studies evaluating associations between prescriber and patient characteristics and warfarin use. Characteristics had anywhere from 1 (multiple characteristics) to 16 (age) evaluation data points. Age, congestive heart failure, cerebrovascular accident, hypertension, male gender, renal impairment, prescriber specialty, geographic region and race were the most commonly evaluated in identified studies (all evaluated ≥ 5 times). Thirty-seven prescriber or patient characteristics were reported only once in included studies. (Table 2). Furthermore, 17 additional characteristics (malignancy, categorical age, admission source, prescriber specialty, insurance status, region of country treated in, race, rate of prior healthcare utilization, year of evaluation, and type of AF, aspirin or other antiplatelet use, perceived appropriateness of warfarin, perceived/actual risk of bleeding, seizures, increasing risk of stroke/embolic event) were identified in studies which were deemed inappropriate for pooling because of heterogeneity in characteristic definition or inconsistencies in data reporting.

Table 2 Association Between Prescriber and Patient Characteristics with Only One Data Point and Warfarin Use
Full size table

For the most part, evaluations of characteristics were rated as being of "good" or "fair" quality, with only 69 of 229 (69.9%) evaluations rated as "poor". Most characteristics were found to have conflicting data in regards to their effect on warfarin use. Only 6 characteristics with more than one data point were found to be consistently associated with [aspirin or other antiplatelet use (n = 4 evaluations), perceived appropriateness of warfarin (n = 2 evaluations), progressing time (n = 2 evaluations), dementia (n = 3 evaluations), AF frequency (n = 3 evaluations) and progressing time (n = 2 evaluations)] or without [coronary artery disease (n = 4 evaluations), seizures (n = 2)] a statistically significant effect on warfarin utilization. No characteristic was found to be significantly associated with both an increase and decease in warfarin use.

Results of Meta-Analysis

Meta-analysis was possible for 18 different prescriber and patient characteristics. The results of these analyses are summarized in Table 3. Upon meta-analysis, characteristics associated with a statistically significant increase in the odds of warfarin use included history of cerebrovascular accident (OR = 1.59), congestive heart failure (OR = 1.36), and male gender (OR = 1.12). Those associated with a statistically significant reduction in the odds of warfarin use included alcohol or drug abuse (OR = 0.62), perceived barriers to compliance (OR = 0.87), contraindication(s) to warfarin (OR = 0.81), dementia (OR = 0.32), falls (OR = 0.60), gastrointestinal hemorrhage (OR = 0.47), intracranial hemorrhage (OR = 0.39), hepatic impairment (OR = 0.59), and renal impairment (OR = 0.69). Age per 10-year increase (OR = 0.78) and advancing age as a dichotomized variable (OR = 0.57) were associated with a statistically significant reduction in warfarin use. Diabetes (OR = 1.11, p = 0.13), history of bleeding (OR = 0.47, p = 0.06) and hypertension (OR = 1.34, p = 0.06) all showed trends towards effect, but failed to reach the a priori cut-off for statistical significance. Coronary artery disease did not appear to affect warfarin prescribing (p = 0.59).

Table 3 Results of Meta-Analysis Evaluating the Association Between Prescriber and Patient Characteristics and Warfarin Use
Full size table

Statistical heterogeneity (I2 > 50% or a Cochrane Q p < 0.10) was found to be present in over half of the characteristics evaluated. However in most cases, the heterogeneity appeared to be due to variance in the magnitude and not the direction of effect.

Review of Egger's weighted regression statistic p-values suggested a lower likelihood of publication bias (p > 0.17 for all evaluable); however, many (n = 10) analyses contained too few studies to allow for proper assessment. Twenty-one (75%) of included studies reported at least one nonsignificant result suggesting a lower likelihood of a negative reporting bias (failure to report nonsignificant results).

Strength of Evidence Grading

Results of strength of evidence grading are found in Table 4. Four prescriber or patient characteristic associations with warfarin use were found to have a "high", 20 a "moderate", and 3 a "low" strength of evidence rating. Five characteristics with more then one evaluation reported in identified studies were deemed to have data 'insufficient" to make rate the strength of evidence. Based upon the limited amount of data, the strength of evidence for all prescriber and patient characteristics evaluated only once in identified studies (n = 370) was automatically deemed "insufficient" and have not included in the strength of evidence table.

Table 4 Strength of Evidence Supporting the Systematic review and Meta-Analysis' Conclusions
Full size table

Discussion

It has been suggested that the complex nature of warfarin prescribing has resulted in under prescribing of warfarin for stroke prevention in AF patients [5]. Our systematic review and meta-analysis confirm this assertion of under prescribing of real-word warfarin (median ~ 49%); however, it also suggested that warfarin utilization has been increasing somewhat over time. This increase in prescribing may be a result of greater awareness of the benefit-to-risk ratio of warfarin in this setting stemming from updated treatment guidelines and national organization campaigns such as the American Cancer Society, American Diabetes Association

and American Heart Association's "The Guideline Advantage" initiatives [47]. In addition, our analysis suggests that a prescriber's decision to administer warfarin for stroke prevention based upon the interaction of multiple prescriber and patient characteristics. Upon meta-analysis, independent positive predictors of warfarin use included history of cerebrovascular accident, congestive heart failure, and male gender. Independent negative predictors included alcohol or drug abuse, perceived barriers to compliance, contraindication(s) to warfarin, dementia, falls, gastrointestinal hemorrhage, intracranial hemorrhage, hepatic impairment, and renal impairment. Upon qualitative analysis of characteristics with data not suitable for pooling, 5 additional positive predictors and 4 additional negative predictors of warfarin use were identified. The strength of evidence supporting these conclusions was for the most part deemed "high" to "moderate". Interestingly, many of the characteristics identified as negative independent predictors of warfarin use are integral parts of commonly used anticoagulation and bleeding risk prediction schemas [48].

Age was the most evaluated patient characteristic in identified studies. Interestingly, both the age per 10 year increase and dichotomous advancing age covariates demonstrated statistically significant reductions in warfarin use with increasing or advanced age. However, qualitative review of results of studies which evaluated age as a categorical variable reveals that the youngest of AF patients do not receive warfarin as commonly, likely as a result of perceived lack of stroke risk [2, 43]. As patients age, they begin to be prescribed warfarin more; however, this increased utilization continues only until patients reach a more advanced age (~ 75-80 years old) at which point warfarin use is reduced once again. These results suggest the relationship between age and warfarin use is complex and nonlinear in nature based heavily upon balancing of stroke and bleeding risks. Consequently, including age into a multivariate model as a continuous or even dichotomous variable will likely provide a result that is in error. Future studies should be careful to avoid oversimplifying the association between age and warfarin use.

Surprisingly, hypertension and diabetes both characteristics comprising the CHADS2 score) were not found to be independent positive predictors of warfarin prescribing, although trends were observed in our analysis (p = 0.06 and 0.13, respectively). The lack of statistically significant findings, even after pooling, is possibly a result of type 2 error. For example, hypertension was assessed by only 5 studies in this meta-analysis. Hypertension is also common in AF patients, with ~75% patients enrolled in recent randomized trials reporting it as a comorbid condition [49, 50]. The infrequent reporting and small number of patients without the characteristic in a given study, make it particularly difficult to demonstrate significant effects.

Of note, both gender and race appeared to be associated with warfarin prescribing even after adjustment for confounding through multivariate analysis. Our analysis found that men had a 12% increased (pooled) odds of receiving warfarin for stroke prevention compared to women, and that African-American patients were anywhere from 24% to 69% less likely to receive warfarin. These data may suggest the presence of gender and racial inequalities in AF care in the US. Furthermore, warfarin appeared to be prescribed to patients more commonly in the Northeast and less commonly in the South. These later findings are consistent with prior data suggesting the provision of healthcare in the Northeast is often more in line with select quality metrics and clinical practice guidelines [51].

It should be noted that our systematic review and meta-analysis only attempted to identify predictors of warfarin initiation or use during a defined period of time and not predictors of warfarin persistence. In a recent study by Fang and colleagues, over 25% of patients newly started on warfarin for AF were found to discontinue therapy in the first year [52]. Patients of a younger age and those with fewer stroke risk factors and poorer international normalized ratio (INR) control were found to be less likely to remain on warfarin.

Interestingly, not all available research suggests that warfarin is underused for stroke prevention in AF [53]. While not included in our systematic review because it failed to meet inclusion criteria, Weisbord and colleagues [53] surveyed primary care providers and concluded that few patients with AF and no contraindications to oral anticoagulation were not receiving warfarin. Most importantly, the results of this study compel us to consider a number of prescriber and/or patient characteristics that might not have been captured by studies included in our systematic review, such as previous adverse events on warfarin, polypharmacy, remote living and patient unwillingness to take warfarin.

There are a number of limitations of our systematic review and meta-analysis that should be noted. First, observational studies - such as those included in our review - are prone to bias. Perhaps most concerning in this case, is the potential of misclassification bias, or the improper coding of an AF diagnosis, warfarin prescription, or one the prescriber and/or patient characteristics. All included studies utilized databases detailing patient diagnoses and warfarin utilization; however, many of these databases were likely never designed for research purposes. Moreover, these databases typically included only a fraction of prescriber and/or patient characteristics that could potentially be a predictor of warfarin prescribing. Thus, some prescriber and/or patient characteristics were identified as independent predictors more often than others, not necessarily because of a true or more potent relationship, but because it was more commonly collected in analyzed databases. Also of import, individual studies were often conducted in local or regional databases, so their results may not reflect a US AF population as a whole. Next, there was unexplained statistical heterogeneity present in many of our meta-analyses of prescriber and patient characteristics. Importantly, however, disagreement between studies seemed to be a result in differences in estimation of the magnitude and not direction of effect. While we were unable to determine definitive causes for this heterogeneity, it is likely a result of differences in patient populations evaluated (inclusion and exclusion criteria, ideal vs. not ideal populations) in studies, differences in definitions of prescriber and patient characteristics (which were often not provided, i.e., whether patients had valvular and non-valvular AF), and changes in warfarin use over time. We attempted to limit the influence of the latter by restricting our analysis to studies published during or after 1996. Furthermore, as with any systematic review and meta-analysis, we can not rule out the possibility of publication bias (bias due to incomplete data). While Egger's weighted regression statistic p-values for all characteristics meta-analyzed suggested a lower likelihood of publication bias, many characteristics could not be adequately evaluated because of limited data points or their inappropriateness for pooling. As this was a systematic review and meta-analysis of observational studies, it is also possible that publication bias may occur as a result of authors failing to report nonsignificant finds from their multivariate analyses. However, we are less concerned about this type of negative reporting bias since a large majority of all identified studies reported at least one nonsignificant finding.

Conclusion

Implications for Practice

Warfarin use has increased somewhat over time. Evidence suggests that the decision to prescribe warfarin for stroke prevention in patients with atrial fibrillation is based upon multiple prescriber and patient characteristics. These findings can be used by family practice prescribers and other healthcare decision-makers to target interventions or methods to improve utilization of warfarin when it is indicated for stroke prevention.

Implications for Research

Future studies evaluating predictors of warfarin prescribing in AF patients should focus on prescriber and/or patient characteristics that have never been previously or are infrequently evaluated. In addition, targeted interventions addressing modifiable prescriber and/or patient characteristics identified in this review should undergo evaluation to test their effectiveness at increasing proper warfarin prescription.