Risk Prediction of Stroke in Atrial Fibrillation Patients Using Machine Learning

Abstract

Atrial Fibrillation (AF) independently escalates the risk of stroke, leading to more severe neurological deficits and increased mortality among affected patients. Prior studies on nonvalvular atrial fibrillation (NVAF) and stroke risk often overlooked the use of longitudinal data, where some variables may evolve over time. In our analysis, we constructed three predictive models: Cox proportional hazard regression (CPH), random survival forest (RSF), and XGBoost Survival Embeddings (XGBSE) to estimate time-to-event probabilities for stroke, thromboembolic events, and death in patients, distinguishing between those receiving oral anticoagulants (OACs) and those not receiving them. Consequently, RSF emerged as the top-performing model for stroke outcomes, achieving Harrell's C-Index of 0.80 (CI95%: 0.79-0.81) and 0.71 (CI95%: 0.70-0.72) on the training and testing datasets, respectively. However, for the death outcome, XGBSE demonstrated superior performance, attaining the highest Harrell's C-Index on both datasets, with the value of 0.85 (CI95%: 0.83-0.87) and 0.85 (CI95%: 0.84-0.85), respectively. Moreover, the use of machine learning (ML) in time-to-event data analysis offers several advantages, even when the performance for the overall outcome overlaps.