Beyond static risk scores: longitudinal ai-based modeling for five-year stroke and mortality risk in atrial fibrillation
T Arbel, Y Aperstein, L Rokach, R ArbelAbstract
Background
Accurate risk assessment is central to the management of atrial fibrillation (AF) and to anticoagulation decisions. The ESC guidelines recommend CHA2DS2-VASc for thromboembolic risk stratification; however, this static score provides only moderate discrimination and does not capture evolving clinical trajectories. Real-world electronic health records (EHR) are irregular, heterogeneous, and temporally fragmented. Transformer architectures, originally developed for sequence modeling in natural language processing, can model long-range dependencies and are therefore promising for longitudinal EHR analysis, but require explicit handling of irregular clinical timing.
Purpose
To evaluate whether a time-aware transformer model that explicitly encodes irregular time gaps between clinical events improves the prediction of 5-year stroke or death in AF compared with CHA2DS2-VASc.
Methods
We developed a time-aware transformer-based AI model specifically designed to analyze longitudinal, heterogeneous data, enabling dynamic patterns in patient trajectories to inform risk prediction. We included all AF patients with ≥1 year of pre-AF history from Clalit Health Services, which covers two-thirds of this patient population in Israel. The model incorporated demographics, comorbidities, renal function, laboratory values, vital signs, medication exposure, hospitalization patterns, and evolving clinical events. Discrimination accuracy for the five-year composite outcome of stroke and mortality was assessed using the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) with 95% confidence intervals (CI), compared with CHA2DS2-VASc >= 2. Model error was assessed with the Brier score.
Results
The cohort includes 84,816 patients. The AI model achieved higher discrimination for five-year outcomes (AUC-ROC 0.85; 95% CI 0.84–0.86; Brier score 0.16), outperforming CHA2DS2-VASc ≥2 (AUC-ROC 0.58; 95% CI 0.57–0.59; Brier score 0.41). Incremental prognostic value was driven primarily by prior cerebrovascular disease, frailty and comorbidity burden, and renal, metabolic, and liver dysfunction.
Conclusions
In a large real-world AF cohort, a dedicated time-aware Transformer-based AI model substantially improved discrimination of 5-year stroke and mortality outcomes beyond guideline-based scoring, supporting dynamic AI-driven refinement of AF management.Graphic abstract