Abstract
Contemporary risk scores for ischemic or bleeding event prediction after drug-eluting stent (DES) implantation are limited to determining a single time duration for dual antiplatelet therapy (DAPT) and lack flexibility in providing dynamic risk stratification. This study developed artificial intelligence (AI) models to dynamically predict ischemic and bleeding risks at different time intervals for patients with DES implantation. Using a cohort of 81,594 adult patients from the Cerner HealthFacts® dataset, we designed eight prediction scenarios across four intervals (3, 6, 12, and 18 months). Our Light Gradient Boosting Machine (LGBM) model outperformed the clinical guideline-recommended DAPT score for 12–30 month predictions (AUROC 0.82 vs. 0.79 for ischemia; 0.77 vs. 0.72 for bleeding) and maintained robust performance in scenarios not covered by the DAPT score. This AI-empowered approach offers a novel clinical tool for refined prognostic judgment and optimal DAPT management.
Introduction
Following drug-eluting stent (DES) implantation, patients typically receive Dual Antiplatelet Therapy (DAPT)—a combination of aspirin and a P2Y12 inhibitor—to prevent ischemic events such as myocardial infarction or stent thrombosis. However, clinicians face a persistent "DAPT Dilemma": extended therapy reduces ischemic risk but significantly increases the risk of major bleeding. Conversely, early discontinuation may lead to thrombotic complications. Current clinical tools, such as the DAPT score, are static, providing a single risk estimate at a fixed time point (usually 12 months post-procedure). These tools fail to incorporate new clinical data that emerges during a patient's recovery. There is a critical need for dynamic systems that can adjust risk predictions over time to support personalized, interval-based clinical decisions.
Methods
We developed a framework to dynamically forecast adverse endpoints using longitudinal electronic health record (EHR) data. The study utilized the Cerner HealthFacts® dataset, initially screening over 68 million patients to select a final cohort of 81,594 adults who received DES and had at least one year of follow-up.
Five AI models were evaluated: Light Gradient Boosting Machine (LGBM), Logistic Regression (LR), Random Forest (RF), Gated Recurrent Unit (GRU), and REverse Time AttentIoN (RETAIN). The models utilized a non-feature selection-based approach, incorporating two years of historical clinical data (diagnoses, medications, procedures) updated at the moment of prediction. Model interpretability was achieved using SHAP (SHapley Additive exPlanations) values to identify key risk drivers at both group and individual levels.
Results
The LGBM model consistently outperformed other methods and the standard clinical DAPT score across various prediction windows.
In the 12–30 month prediction window, LGBM achieved an AUROC of 0.82 for ischemic risk and 0.77 for bleeding risk, compared to 0.79 and 0.72 for the DAPT score, respectively. Furthermore, the AI models demonstrated versatility in shorter intervals (3, 6, and 18 months) where the DAPT score is not applicable, maintaining AUROCs between 0.75 and 0.80. SHAP analysis revealed that DAPT status was the most influential feature for ischemic risk, while age, anemia, and cancer were primary drivers for bleeding risk.
Discussion and Conclusion
This study demonstrates that AI-driven dynamic prediction can significantly enhance risk stratification for patients post-DES implantation. By leveraging real-world EHR data to update assessments, clinicians can move beyond static scores toward precision cardiovascular medicine.
Despite these results, limitations remain. EHR data often lacks granularity regarding stent specifications (e.g., diameter, length) and patient adherence to prescribed medications. Additionally, variability in ICD coding across hospitals can introduce noise. Future research will focus on integrating unstructured clinical notes via Natural Language Processing (e.g., Med-BERT) and validating these models across more diverse global datasets. In conclusion, this AI-empowered approach serves as a valuable clinical decision support tool for optimizing DAPT duration and improving patient outcomes.
Summary updated