Machine-learning versus traditional scoring methods for prediction of atrial fibrillation recurrence after catheter...

Atrial fibrillation (AF) recurrence after catheter ablation remains common, with recurrence rates up to 45%. Traditional risk scores demonstrate only modest predictive performance. Machine learning (ML) approaches have shown promise in improving risk stratification by leveraging high-dimensional clinical and imaging data, but direct comparisons with traditional scores remain limited.

To compare the discriminative performance of ML models versus traditional risk scores in predicting AF recurrence following catheter ablation.

MEDLINE, EMBASE, Web of Science, and Cochrane databases were searched to 11 June 2026 for studies reporting head-to-head comparisons of ML and traditional risk models predicting AF recurrence after ablation. The primary outcome was model discrimination, measured by the C-statistic. Random-effects meta-analysis was used to pool discrimination estimates, and meta-regression was performed to compare ML and traditional approaches. Subgroup analysis was conducted to explore the effect of ML model architecture. Methodological quality was assessed using the PROBAST tool.

Eleven studies (n = 17,150) were included. ML models demonstrated higher pooled discrimination compared with traditional risk scores (AUC 0.80, 95% CI 0.72-0.86 vs. 0.58, 95% CI 0.53-0.62; QM 18.96, p < 0.001). Subgroup analysis suggested that multimodal fusion models achieved the highest pooled discrimination, followed by neural-network models. Tree-based models did not significantly outperform traditional risk scores in the primary analysis.

ML models showed higher discrimination compared to traditional risk scores in predicting AF recurrence after ablation. Methodological limitations, lack of calibration reporting, and absence of external validation limit immediate clinical adoption.