Predicting incident atrial fibrillation using AI-ECG: enhancing risk stratification for stroke prevention

Abstract Background Atrial fibrillation (AF) is a prevalent cardiac arrhythmia associated with significant morbidity and mortality. Identifying individuals at risk of developing AF remains a clinical challenge. This study aims to develop machine learning models for predicting incident AF using ECG data from patients in sinus rhythm. Methods A total of 83,160 patients in sinus rhythm were enrolled, with 80% (n=66,528) allocated for training, 10% (n=8,316) for testing, and 10% (n=8,316) for internal validation. Patients presented with AF within 30 days of ECG acquisition were excluded, and those with newly diagnosed AF within one year were classified as having incident AF. Machine learning models, including CatBoost, XGBoost, LightGBM, Random Forest, and Extra Trees, were trained using tabular ECG data. Additionally, deep learning models were explored for image-based ECG data. Results The machine learning models demonstrated strong predictive performance for AF detection. The machine learning model, ExtraTree achieved the highest F1 score (91.76) and area under curve (AUC: 0.980), outperforming other models, including CatBoost (AUC: 0.936), XGBoost (AUC: 0.929), LightGBM (AUC: 0.871) and RandomForest (AUC: 0.979). Feature selection analysis suggests the morphology of horizontal P wave may aid in AF prediction. Conclusion AI-ECG effectively predicts incident AF, offering a valuable tool for risk stratification and early detection. Its integration into clinical practice may enhance rhythm management and stroke prevention by identifying high-risk individuals before AF onset.