Machine‑Learning Models for Predicting Atrial Fibrillation Recurrence and Stroke Risk: A Systematic Review

Abstract 2294792, presented at Western Atrial Fibrillation Symposium 2026

Traditional risk scores such as CHADS‑VASc inadequately predict AF recurrence and stroke. Machine‑learning (ML) models that incorporate clinical, imaging and electrocardiographic features may offer better individualized risk stratification.PubMed, Scopus and arXiv were searched for studies (January 2014–August 2025) developing or validating ML models to predict (a) AF recurrence after ablation or cardioversion and (b) stroke in AF patients. Model types, input variables, performance (area under the curve [AUC]) and calibration were extracted. Risk of bias was assessed using the PROBAST tool.Twenty‑six studies were included. Models employed random forests, support‑vector machines, gradient boosting or deep neural networks. Predictors commonly included demographics, AF type, comorbidities, left‑atrial volume and ECG features. For AF recurrence, AUCs ranged from 0.70 to 0.88, improving discrimination over traditional scores by 5–15 percentage points. Stroke‑prediction models that integrated ECG and imaging features achieved AUCs up to 0.87, compared with ~0.65 for CHADS‑VASc. Only eight models underwent external validation; calibration performance was often underreported.ML‑based models show promise for personalized prediction of AF recurrence and stroke, outperforming existing risk scores. Greater transparency, external validation and integration into clinical workflows are needed before widespread adoption.