OBJECTIVES:  Anticoagulation is recommended for atrial fibrillation (AF) patients with moderate-high stroke risk and low risk of bleeding.  Previous data have shown that bleeding rates increase in parallel with stroke risk because stroke and bleeding risk factors overlap.  The objective of this study was to assess whether adding ATRIA specific bleeding risk variables to the CHADS2 stroke risk variables will improve the ability to predict major bleeds.   METHODS:  Medical claims for AF patients (ICD-9 code 427.31) with continuous eligibility 12 months prior to an AF diagnosis in the MarketScan database between January 2003 to December 2007 were analyzed.  Data before an index AF diagnosis were used to assess the presence of CHADS2 (congestive heart failure, hypertension, age ≥75, diabetes, prior stroke/transient ischemic attack) and ATRIA (anemia, renal disease, age ≥75, prior bleeding, hypertension) risk factors. The Pearson correlation coefficient between the CHADS2 and ATRIA risk scores and Cox proportional hazards regressions comparing a reduced model (CHADS2 covariates only) and a full model (CHADS2 + ATRIA covariates) to predict major bleeds after an AF diagnosis were estimated.  Likelihood ratio tests were used to test for differences between the models.   RESULTS:  A total of 64,946 AF patients were included in the analysis (47% ≥75 years of age, 45% females).   ATRIA and CHADS2 scores were correlated (r=0.63, p<0.0001).  CHADS2 risk factors predicted major bleeding (LR=444.66, P<0.001). The addition of ATRIA risk factors enhanced predictive power (LR=519.65, P<0.001) and resulted in a significant improvement in model fit (P<0.001). A range of sensitivity analyses revealed similar findings.   CONCLUSIONS:  Adding ATRIA bleeding risk factors to the CHADS2 significantly improved the model fit. These results suggest that information provided by the ATRIA bleeding risk index improves bleeding risk assessment and may potentially help optimize anticoagulation decisions. Further research is needed to assess improvements in model discrimination.