In patients discharged with non-valvular atrial fibrillation who were not receiving treatment with vitamin K antagonists or heparins, we estimated event rates for thromboembolism and death for the various CHADS2 and CHA2DS2-VASc scores and for the specific covariate combinations forming the scores of 1 or 2. We estimated the risk of thromboembolism by using Cox proportional hazard regression models. In the Cox models, we analysed the risk associated with all possible risk factor combinations for CHADS2 score=1 (four combinations) and CHADS2 score=2 (seven combinations); we used CHADS2 score=0 as the reference. In the same manner, other Cox models analysed the risk associated with all possible risk factor combinations for CHA2DS2-VASc score=1 (six combinations) and CHA2DS2-VASc score=2 (17 combinations), with CHA2DS2-VASc score=0 used as the reference. We did all analyses for one, five, and 10 years of follow-up. In additional Cox regression models, we included concomitant treatment with antiplatelet drugs (that is, primary acetylsalicylic acid, clopidogrel, and dipyridamole), to adjust for this potential confounder. We also did sensitivity analyses by not including pulmonary embolism as an outcome.
We used C statistics estimated from Cox regression models to assess the predictive capability of CHADS2 and CHA2DS2-VASc for thromboembolism, using the method described by Liu and colleagues.24 C statistics give a measure of how well the risk prediction scheme identifies patients who will have a future event. For estimating C statistics, we analysed CHADS2 and CHA2DS2-VASc as risk scores (0-6 and 0-9) and as risk groups (low, intermediate, and high). We also evaluated the scores both as categorical and as continuous covariates. We constructed survival curves, based on Kaplan-Meier estimates of the probability of remaining free of thromboembolism with a score of 0 and 1, for the two risk stratification schemes. We considered a two sided P value <0.05 to be statistically significant. In all Cox models, the model assumptions (that is, proportional hazards, linearity of continuous covariates, and lack of interactions) were found to be valid. We used SAS statistical software version 9.1 and Stata statistical software version 11.0 for the analyses.
We used C statistics estimated from Cox regression models to assess the predictive capability of CHADS2 and CHA2DS2-VASc for thromboembolism, using the method described by Liu and colleagues.24 C statistics give a measure of how well the risk prediction scheme identifies patients who will have a future event. For estimating C statistics, we analysed CHADS2 and CHA2DS2-VASc as risk scores (0-6 and 0-9) and as risk groups (low, intermediate, and high). We also evaluated the scores both as categorical and as continuous covariates. We constructed survival curves, based on Kaplan-Meier estimates of the probability of remaining free of thromboembolism with a score of 0 and 1, for the two risk stratification schemes. We considered a two sided P value <0.05 to be statistically significant. In all Cox models, the model assumptions (that is, proportional hazards, linearity of continuous covariates, and lack of interactions) were found to be valid. We used SAS statistical software version 9.1 and Stata statistical software version 11.0 for the analyses.
