Residual risk prediction in anticoagulated patients with atrial fibrillation using machine learning: A report from the GLORIA-AF registry phase II/III.

Abstract

Although oral anticoagulation decreases the risk of thromboembolism in patients with atrial fibrillation (AF), a residual risk of thrombotic events still exists. This study aimed to construct machine learning (ML) models to predict the residual risk in these patients. Patients with newly diagnosed non-valvular AF were collected from the Global Registry on Long-Term Oral Anti-Thrombotic Treatment in Patients with Atrial Fibrillation (GLORIA-AF) registry. To predict the residual risk of the composite outcome of thrombotic events (defined as ischemic stroke, systemic embolism, transient ischemic attack and myocardial infarction), we constructed four prediction models using the logistic regression (LR), random forest, light gradient boosting machine and extreme gradient boosting machine ML algorithms. Performance was mainly evaluated by area under the receiver-operating characteristic curve (AUC), g-means and F1 scores. Feature importance was evaluated by SHapley Additive exPlanations. 15,829 AF patients (70.33 ± 9.94 years old, 55% male) taking oral anticoagulation were included in our study, and 641 (4.0%) had residual risk, sustaining thrombotic events. In the test set, LR had the best performance with higher AUC trend of 0.712. RF has highest g-means of 0.295 and F1 score of 0.249. This was superior when compared with the CHA2DS2-VA score (AUC 0.698) and 2MACE score (AUC 0.696). Age, history of TE or MI, OAC discontinuation, eGFR and sex were identified as the top five factors associated with residual risk. ML algorithms can improve the prediction of residual risk of anticoagulated AF patients compared to clinical risk factor-based scores.