Abstract 11999: External Validation of a New Risk Score to Predict Stroke in Atrial Fibrillation

Abstract

Background and Objective: Accurate stroke risk scores are needed to optimize anticoagulation use in atrial fibrillation (AF). In its derivation cohort, the ATRIA score was superior to CHADS₂ and CHA₂DS₂-VASc. Here we test the ATRIA score in a separate community-based cohort with incident AF.

Methods: The ATRIA risk model (Figure) was derived and internally validated using the ATRIA1 AF cohort, n=13,559, assembled in 1996-97 from Kaiser Permanente Northern California. We externally validated the model in the ATRIA2-CVRN incident AF cohort, n=34,959, from Kaiser Permanente Northern and Southern California assembled 2006-2009. Databases provided patient features and warfarin use. Ischemic stroke and other thromboembolic events (both=TE) were identified through 2009 and confirmed by chart review. We assessed score performance in patient-time off warfarin by c-index and net reclassification improvement (NRI).

Results: There were 517 TEs in 27,644 person-years off warfarin, rate=1.87%/yr. C-indices were: ATRIA: 0.694 (95% CI 0.671-0.718); CHADS₂: 0.663 (95% CI 0.640-0.686); and CHA₂DS₂-VASc: 0.676 (95% CI 0.653-0.698). Using published point score thresholds for low/moderate/high risk categories the c-indices were 0.656 (95% CI 0.635-0.677) for ATRIA, 0.646 (95% CI 0.624-0.667) for CHADS₂ and 0.581 (95% CI 0.573-0.590) for CHA₂DS₂-VASc. Using rate thresholds of 1% and 2%/yr, the NRI from the ATRIA score was 22% versus CHADS₂ and 24% versus CHA₂DS₂-VASc (both p<0.001). When restricted to severe (Rankin 3+) events, the ATRIA point score c-index improved markedly to 0.749 (0.722-0.777) versus 0.691 (0.661-0.720) for CHADS₂ and 0.713 (0.686-0.740) for CHA₂DS₂-VASc.

Conclusions: These results externally validate the improved performance of the ATRIA score and confirm its very good accuracy for predicting disabling stroke. They also highlight the problem of generalizing categorical risk thresholds to lower risk community-based cohorts.