Estimating Stroke and Bleeding Risk in Atrial Fibrillation

Abstract & Commentary

By Michael H. Crawford, MD, Editor

Source: Friberg L, et al. Evaluation of risk stratification schemes for ischemic stroke and bleeding in 182,678 patients with atrial fibrillation: The Swedish Atrial Fibrillation cohort study. Eur Heart J 2012;33:1500-1510.

Various risk schemes have been promulgated for assessing the risk of stroke and bleeding in patients with atrial fibrillation (AF). These investigators used the Swedish Atrial Fibrillation cohort study to investigate the comparative utility of four schemes that have been validated in AF cohorts: CHADS2, CHA2DS2-VASc, HAS-BLED, and HEMORR2HAGES. All 182,678 hospitalized patients with a diagnosis of AF in Sweden between 2005 and 2008 were included in this prospective registry. The median follow-up was 1.4 years. Excluded were 7167 patients who either died during the initial hospitalization or had valvular AF, leaving 170,291 who met study criteria. The National Swedish Drug Registry was used to ascertain who used oral anticoagulants (OA). The 90,490 patients (53%) who never received OA were used for most of the analyses in this report. The stroke risk clearly increased with age: > 75 years (HR 5.49) and 65-74 years (HR 3.07), and was more common in women (HR 1.21). Other significant risk factors on multivariate analysis included prior cerebral events, coronary revascularization, vascular disease, hypertension, and diabetes. The discriminant ability of the stroke risk schemes (c-statistic) was 0.66 for CHADS2 and 0.67 for CHA2DS2-VASc. However, the CHA2DS2-VASc is superior for identifying very low-risk patients: A CHADS2 score of 1 has a stroke/100 years at risk of 3.0 vs 0.6 for a CHA2DS2-VASc score of 1. Significant multivariate risk factors for major bleeding in patients not on OA included age, male sex, prior stroke, prior bleeding event, heart failure, hypertension, renal failure, liver disease, anemia, coagulopathy, alcohol abuse, and cancer < 3 years ago. The discriminant ability of HAS-BLED and HEMORR2HAGES for major bleeding in patients not on OA or aspirin were 0.61 vs 0.69, with similar risks at scores between 0 and 5. When scores are greater than 5 the risk data varies, but all would be considered at high risk. The authors concluded that CHA2DS2-VASc performs better than CHADS2, especially in low-risk patients. While the two bleeding risk schema perform similarly, HAS-BLED has the advantage of simplicity.


Europe is leading the world in large registry studies, which is one advantage of single-payer systems. This one sheds light on some important practical issues with anticoagulation for AF. First, CHA2DS2-VASc does a better job at defining the low-risk population. A CHADS2 score of 0 is associated with a stroke risk of < 1% in AF patients not on anticoagulants or aspirin, but a score of 1 equals a 3% risk. With CHA2DS2-VASc, the stroke risk associated with each score is: 0 = < 1%, 1 = < 1%, 2 = 2%, and 3 = 3%. Second, the HAS-BLED score performs almost as well as HEMORR2HAGES, but is much simpler. With both schemas, the risk of major bleeding rises progressively with the score. At a score of 5, the two schemas give a bleeding risk on OA of 5.7 and 6.0 bleeds per year. Third, controversial risk predictors were clarified. Heart failure is in both stroke risk scores and was not a multivariate predictor of stroke, but did predict major bleeding (HR 1.15) even though it is not in HAS-BLED. Thyroid disease was not predictive of stroke or bleeding and is not in these schemas. Renal disease patients, who are not usually included in clinical trials, were predictive of bleeding and included in HAS-BLED.

One thing to consider is that this study is based on an administrative database — although a fairly robust one — and has all the limitations and biases expected. In addition, this hospital-based population included patients who probably are older and have more comorbidities than patients cared for solely as outpatients. Since no INR data were available, this variable in the bleeding schemas (labile INR) was omitted. Also, the stroke risk in those not on OA or aspirin must suffer from selection bias since these patients are likely to be lower-risk patients. So, the risk may be higher in a less selected group. Despite these significant limitations, it is unlikely that a randomized, controlled trial of these risk prediction schema will ever be done. Consequently, this large, well-done registry may be the best data we have for a long time.