AI-ECG Risk Stratification for Atrial Fibrillation

Figures

• 

  Figure 1. Distribution of Artificial Intelligence–Enabled Electrocardiography Risk Categories and Association With Atrial Fibrillation (A) Distribution of AI-ECG risk categories. (B) ORs for prevalent AF across AI-ECG risk categories, using the Low-risk group as the reference category. AF = atrial fibrillation.

• 

  Figure 2. Area Under the Curve Performance for Internal Validation Internal validation results for support vector machine (SVM), adaptive boosting (AdaBoost), and artificial neural network (ANN) across model 1, model 2, and model 3.

• 

  Figure 3. Area Under the Curve Performance for External Validation External validation results for support vector machine (SVM), adaptive boosting (AdaBoost), and artificial neural network (ANN) across model 1, model 2, and model 3.

• 

  Figure 4. SHapley Additive exPlanations Summary Plot for Model 1 Mean absolute SHapley Additive exPlanations (SHAP) values are shown for all predictors in Model 1 across the 3 classifiers (SVM, AdaBoost, and ANN). AdaBoost = adaptive boosting; ANN = artificial neural network; BMI = body mass index; BUN = blood urea nitrogen; Cr = serum creatinine; EF = ejection fraction; eGFR = estimated glomerular filtration rate; Hb = hemoglobin; HR = heart rate; IVS = interventricular septum; LAD = left anterior descending coronary artery; LAVi= left atrial volume index; NT-proBNP = N-terminal pro–B-type natriuretic peptide; PW = posterior wall; RBC = red blood cell; SVM = support vector machine; TIA = transient ischemic attack.

• 

  Figure 5. SHAP Summary Plot for Model 2 Mean absolute SHapley Additive exPlanations (SHAP) values are shown for all predictors in model 2 across the three classifiers (SVM, AdaBoost, and ANN). Abbreviations as in Figure 4.

• 

  Central Illustration. AI-ECG for Atrial Fibrillation Risk Stratification AI-ECG enables risk stratification for atrial fibrillation in a multicenter real-world cohort. Higher AI-ECG risk categories are associated with an increased prevalence of AF, with progressively higher odds ratios compared with the low-risk group. Receiver operating characteristic analysis demonstrates the performance of AI-ECG relative to conventional clinical risk scores. SHAP (SHapley Additive exPlanations) analysis illustrates the contribution of both clinical variables and ECG features to model predictions, supporting the interpretability of AI-ECG. These findings suggest that AI-ECG provides a simple and explainable tool that complements established clinical risk scores for AF risk assessment. AdaBoost = adaptive boosting; AF = atrial fibrillation; AI-ECG = artificial intelligence–enabled electrocardiography; ANN = artificial neural network; BMI = body mass index; BUN = blood urea nitrogen; Cr = serum creatinine; EF = ejection fraction; eGFR = estimated glomerular filtration rate; Hb = hemoglobin; HR = heart rate; IVS = interventricular septum; LAD = left anterior descending coronary artery; LAVi= left atrial volume index;…