External validation and comparison of published stroke risk scores in a Chinese cohort of patients with atrial fibrillation.

BACKGROUND

Accurate prediction of ischemic stroke or systemic embolism is central to atrial fibrillation (AF) management, yet the applicability of available risk scores to Chinese populations remains uncertain.

OBJECTIVE

The aim of this study was to externally validate and compare published stroke risk scores in Chinese patients with nonvalvular AF.

METHODS

We systematically retrieved and externally validated 17 published models in a retrospective cohort of 1283 adults hospitalized with nonvalvular AF at a Chinese tertiary-care center. Discrimination was evaluated using the 4-year Uno C-index and time-dependent areas under the curve. Models that provided absolute risks also underwent calibration, Brier score, and reclassification analyses. Sensitivity analyses accounted for death as a competing event and separately evaluated patients who were not receiving oral anticoagulant at baseline.

RESULTS

During 4 years of follow-up, 103 primary outcomes occurred, corresponding to 2.85 events per 100 person-years. Across models, 4-year Uno C-indices ranged from 0.572 to 0.767, compared with 0.690 for CHA2DS2-VASc. GARFIELD-AF 2017, GARFIELD-AF 2021, and ATRIA had higher Uno C-indices than CHA2DS2-VASc. However, no model had a significantly higher time-dependent areas under the curve at any time point. For 1-year calibration, CHA2DS2-VASc had predicted and observed risks of 3.39% and 3.54%, respectively, with an observed-to-expected ratio of 1.04, whereas several other models underestimated the risk. Brier scores and reclassification measures showed no consistent advantage. Competing-risk analyses lowered the cumulative incidence without changing comparative patterns. Among patients not receiving oral anticoagulant at baseline, no model outperformed CHA2DS2-VASc.

CONCLUSION

Candidate models did not consistently outperform CHA2DS2-VASc in this external validation; CHA2DS2-VASc remains practical, whereas other models may require population-specific validation and recalibration.