Predictors of long-term atrial fibrillation recurrence after catheter ablation: non-linear analytical approach for individualized prognostic stratification

Abstract Background Recurrences are common in patients with atrial fibrillation (AF) during long-term follow-up after catheter ablation (CA) for pulmonary vein isolation (PVI). Recently, machine learning (ML) models identifying non-linear relationships among various patient parameters have been applied prognostic stratification different cardiac diseases. Aim This is a retrospective study aimed to determine whether ML-based can identify individual clinical baseline characteristics and CT-quantified volumetric of epicardial fat tissue (EFT) aid prognosing outcome PVI paroxysmal AF. Methods A cohort 92 (median age 60.2 [51.9–64.0]; 74% male) AF (a single persistent episode was accepted) undergoing targeting analysed. All underwent CT imaging were fitted implantable loop recorder (ILR) prior CA. For PVI, radiofrequency CA electro-anatomical mapping used 79 patients, cryoballoon 13 patients. recurrence, defined as burden >0.1% the blanking period (90 days), continuously assessed by ILR. Feature selection on 23 performed using random forest (XGBoostRegressor). Mean absolute Shapley values (|mSHAP| – Additive expLanations) quantify relative discriminative power analysed parameters. Results During 3-years, recurrence detected 58 (63%) 29 (50%) them repeat ablation. Five most important predictors 3-year upper volume, BMI, burden, pericardial volume (lower segment) (Fig. 1). Upper EFT twice males than females (0.44 0.21 |mSHAP| respectively). no low (41 years) (13.2 ml) drivers predicting positive 2A). In contrast, post CA, an above-average 55.5 ml high BMI had significant net contribution his failed 2B). Conclusion Non-linear ML analysis our limited suggests: i) association ILR determined follow-up; ii) potential role such analyses more granular highly individualized prediction planned However, these results need further testing, validation prospective trials. Funding Acknowledgement Type funding sources: None.