Machine learning models performance in prediction of paroxysmal atrial fibrillation recurrence

Abstract

Introduction: Machine learning algorithms have driven the development of models capable of predicting paroxysmal atrial fibrillation recurrence, outperforming conventional risk scores.

Objective: To synthesize the discriminative performance of machine learning algorithms for predicting paroxysmal atrial fibrillation recurrence.

Methods: A systematic review of articles published in LILACS, PubMed, OpenAlex, ScienceDirect, and Europe PMC was conducted. Clinical trials, cohort studies, and predictive model development/validation studies involving adults diagnosed with paroxysmal atrial fibrillation were included. The primary outcome was arrhythmia recurrence. Performance metrics were pooled using a random-effects meta-analysis, and the risk of bias was evaluated using the PROBAST tool.

Results: Seven studies were identified in the qualitative phase (N = 7,466 patients; post-ablation recurrence rate: 13.47%–37.5%), and five models provided data for the quantitative synthesis. Artificial intelligence-based algorithms demonstrated a high overall discriminative performance (AUC = 0.78; 95% CI: 0.71–0.85; p <0.0001), along with high inter-study heterogeneity (I² = 91.5%). Subgroup analysis reported higher stability in alternative architectures compared to Ensemble Trees. All included studies exhibited a high overall risk of bias, driven by deficiencies in the statistical analysis domain.

Conclusions: Machine learning models demonstrate high overall discriminative performance for predicting recurrence. Nonetheless, the high methodological risk of bias identified in primary literature requires a cautious interpretation of the overall performance.