Abstract 186: A New Risk Model to Predict Stroke in Atrial Fibrillation
Objective: Evaluating stroke risk in patients with atrial fibrillation (AF) is central to the anticoagulation decision. Our goal was to develop and internally validate a more accurate model to predict stroke in AF.
Methods: We analyzed all person-time off warfarin from our large ATRIA community cohort of patients with AF treated in Kaiser Permanente Northern California. Patient demographics, clinical features, and thromboembolism outcomes (TE, 92% ischemic stroke) were obtained from health plan databases. Outcomes were adjudicated by medical record review and assigned a Rankin severity score. Utilizing a split-sample approach to permit internal validation, we used Cox proportional hazards models with time-varying covariates and bootstrap sampling to minimize overconfidence in selecting among candidate variables for the new model. Points were assigned proportional to regression coefficients. The score’s 0-15 point range was collapsed into low risk (TE rate <1%/yr), moderate risk (TE rate 1- 2%/yr), and high risk (TE rate ≥2%/yr) categories. We compared our ATRIA score to the CHADS2 and CHA2DS2-VASc scores in our cohort by c-index measure of discrimination and by net reclassification improvement (NRI).
Results: 10,927 patients accumulated 32,609 person-years off warfarin during follow-up, with 685 TE events. Age was the dominant univariate risk factor among those without prior stroke but all patients with prior stroke were at high risk. Eight variables, age, prior stroke, age*prior stroke interaction, female sex, diabetes, congestive heart failure, hypertension, proteinuria, and eGFR<45 were included in the final ATRIA model (Table). For the entire cohort, the c-statistic for the ATRIA score was 0.73 (95% CI 0.71-0.75) versus 0.69 (95% CI 0.67-0.71) for the CHADS2score and 0.70 (95% CI 0.68-0.72) for the CHA2DS2-VASc score. The c-statistic for the ATRIA score improved to 0.76 (95% CI 0.74-0.79) when the model was tested on Rankin 3+ (severe stroke) events. The ATRIA score led to an NRI of 20% compared to CHADS2 and 9% compared to CHA2DS2-VASc, using cut-points for the latter scores optimized to fit ATRIA data. Using the scores’ original published cut-points, NRI increased to 26.4% for CHADS2 and 27.0% for CHA2DS2-VASc.
Conclusion: Our new stroke risk score with 8 easily available clinical features showed improved performance in categorizing stroke risk in AF patients as compared to existing stroke risk scores. Its performance was even better for predicting severe strokes. The ATRIA score merits external validation.
- © 2012 by American Heart Association, Inc.