Development and Validation of Dynamic Prediction Model for Thromboembolic and Bleeding Risk in Patients With Atrial Fibrillation Initiating Direct Oral Anticoagulants.

BACKGROUND

Current tools for assessing ischemic stroke and bleeding risks in patients with atrial fibrillation focus on warfarin users and initial therapy decisions; however, dynamic models for reassessing these risks in direct oral anticoagulant (DOAC) users are lacking. We aimed to develop a tool that enables the dynamic evaluation of thromboembolic and bleeding risks in patients with atrial fibrillation receiving DOAC therapy.

METHODS

Using a national claims database, we included 42 450 patients with atrial fibrillation who were newly prescribed DOACs and divided them into development (2018) and temporal-validation (2019) cohorts. We applied a landmarking approach and developed dynamic prediction models incorporating time-dependent variables. The Fine-Gray subdistribution hazard model, which accounts for death as a competing risk, was used to derive these models.

RESULTS

Dynamic prediction models incorporating time-dependent variables, such as DOAC adherence, were developed to estimate the 1-year thromboembolic and major bleeding risks over a 2-year period following DOAC initiation. The thromboembolic and major bleeding risk prediction models achieved a concordance index of 0.715 (95% CI, 0.701-0.729), and 0.697 (95% CI, 0.682-0.710), respectively. The thromboembolic model outperformed CHA2DS2-VASc (concordance index, 0.664 [95% CI, 0.652-0.678]) by 0.051 (95% CI, 0.038-0.062), and the major bleeding model outperformed HAS-BLED (concordance index, 0.627 [95% CI, 0.611-0.641]) by 0.070 (95% CI, 0.055-0.084). Calibration plots showed good alignment between the predicted and observed probabilities. Nomograms were constructed for each model to enhance its clinical utility.

CONCLUSIONS

These dynamic prediction models offer a novel approach for reassessing thromboembolic and major bleeding risks in patients with atrial fibrillation receiving DOACs, demonstrating superior performance compared with existing models. Enabling regular risk updates based on the patient's current status, these tools support personalized and dynamic risk assessments in clinical practice.