We used patient-level data from the ENVISAGE-TAVI AF trial (NCT02943785), a multinational, multicenter, prospective, randomized, open-label trial with blinded endpoint adjudication, designed to compare edoxaban with vitamin K antagonist (VKA) therapy in patients with AF following successful TAVR. The trial enrolled patients who were 18 years of age or older, with either pre-existent paroxysmal/persistent AF or new-onset AF after successful TAVR for severe AS. Patients with comorbidities associated with a high bleeding risk (i.e., history of intracranial hemorrhage, esophageal varices, and unresolved periprocedural complications, etc.) were excluded. A full description of inclusion and exclusion criteria is given in Supplementary Table 1 . A total of 1,426 patients were enrolled across international centers and randomized between 12 hours and 5 days after the TAVR procedure. ^, All patients were followed for a median of approximately 18 months. Patients were allocated to receive either once-daily oral edoxaban (60 mg or 30 mg based on dose-reduction criteria per local labeling) or a standard VKA regimen targeting an international normalized ratio of 2.0 to 3.0 (or 1.6-2.6 for patients aged ≥70 years in Japan). Concomitant antiplatelet therapy was permitted at the discretion of the treating physician. Randomization was stratified by anticipated edoxaban dose reduction. The maximum treatment duration was up to 36 months.

We examined 4 clinical outcomes of interest: (1) clinically relevant bleeding (CRB), (2) ischemic stroke (IS), (3) major gastrointestinal bleeding (MGIB), and (4) net adverse clinical events (NACE). CRB was classified according to the International Society on Thrombosis and Hemostasis definition and included both major (clinically overt bleeding associated with a reduced hemoglobin level, blood transfusion, symptomatic bleeding at a critical site, or death) and clinically relevant nonmajor bleeding (any sign or symptom of hemorrhage that does not fit the criteria for major bleeding but does meet at least 1 of the following criteria: (1) requires medical intervention by a healthcare professional, (2) leads to hospitalization or increased level of care, or (3) prompts a face-to-face evaluation). ^, NACE was defined as the composite of death from any cause, myocardial infarction, IS, systemic thromboembolic event, valve thrombosis, or major bleeding. IS was defined in accordance with the Valve Academic Research Consortium-2 consensus as an acute episode of focal cerebral, spinal, or retinal dysfunction caused by infarction of the central nervous system tissue.

This analysis utilized data from the safety cohort, which included patients who received ≥1 dose of the study drug. The dataset included 180 variables, selected by experts for their potential clinical relevance. After basic data transformations and cleaning, including the removal of homogeneous features and highly correlated variables, 155 variables remained in the final dataset. Since this dataset was derived from a randomized controlled trial, baseline characteristics were well balanced across the edoxaban and VKA arms. Therefore, we did not separate model development by treatment group. Instead, treatment allocation was included as a candidate predictor. Ten different ML algorithms (logistic regression, lasso regression, ridge regression, elastic net, linear discriminant analysis, Naïve Bayes, decision tree, random forest, XGBoost, and multilayer perceptron) were evaluated using nested cross-validation on a training dataset comprising 75% of the patients from the full dataset. The nested cross-validation involved 2 loops to prevent overfitting and bias in model selection. In the inner loop, hyperparameter tuning was performed using Bayesian optimization. In the outer loop, the models were scored using the F1 score to determine the best model for each endpoint. The F1 score was selected as the appropriate performance metric for imbalanced classification tasks. Finally, trained models were evaluated on a holdout test set comprising 25% of the patients not included in the training set. We also compared the performance of the best ML model for each endpoint to a logistic regression model that used a traditional risk score (CHA₂DS₂-VA and HAS-BLED) as its only input variable. The CHA₂DS₂-VA and HAS-BLED scores are widely used clinical tools to guide anticoagulation management in patients with AF. The CHA₂DS₂-VA score estimates the risk of stroke or systemic embolism by assigning points for congestive heart failure, hypertension, age ≥75 years (2 points), diabetes mellitus, prior stroke or transient ischemic attack (2 points), vascular disease, and age 65 to 74 years. In contrast, the HAS-BLED score evaluates the risk of major bleeding, with points assigned for hypertension, abnormal renal or liver function, prior stroke, history of bleeding, labile INR, age >65 years, and concomitant drug or alcohol use. Both scores share several overlapping factors (i.e., hypertension, stroke history, and age), but they differ in their primary purpose. Unlike CHA₂DS₂-VA and HAS-BLED, which are based on a small set of predefined clinical risk factors, our ML models incorporated over 150 routinely collected clinical, laboratory, and procedural variables. This broader input allows ML to explore nonlinear and higher-order interactions that are not captured by traditional scores and to identify predictors specific to the TAVR population.

To identify the most impactful variables for the outcomes of interest, SHAP (Shapley Additive exPlanations) was employed. SHAP values utilize a game-theoretical approach to quantify the importance of a variable in a prediction by assessing its contribution across various combinations of variables. The features were ranked using the mean absolute SHAP value from the holdout test set. A detailed explanation of the methods used for the derivation and optimization of the ML models was previously described.

Baseline characteristics of the overall safety cohort are listed in Table 1 . The study population included 1,377 patients, with a mean age of 82.1 ± 5.4 years. Approximately half (52.2%) were men, and the majority (83.2%) identified as White. The mean body mass index was 27.7 ± 5.5 kg/m², with a mean creatinine clearance of 58.2 ± 24.1 mL/min. Cardiovascular and metabolic comorbidities were highly prevalent, including hypertension (91.4%), hypercholesterolemia (70.0%), diabetes (36.7%), and advanced heart failure (New York Heart Association class III or IV) in nearly half (44.5%) of the cohort. Baseline thromboembolic and bleeding risks were elevated, reflected by a median CHA₂DS₂-VA score of 4 (mean 4.5 ± 1.3) and a mean HAS-BLED score of 1.6 ± 0.8.

Out of 1,377 patients, CRB was recorded in 375 patients (27.2%). The Naïve Bayes model ( Supplementary Table 2 ) provided the best predictive performance, with an F1 score of 0.39, closely matching HAS-BLED (F1 score = 0.41) and CHA₂DS₂-VA (F1 score = 0.38). Among the ML predictors, the absence of history of previous valve replacement before index procedure (no valve replacement) had the highest SHAP value, indicating its strong influence ( Supplementary Figure 1 ).