Left atrial anterior sphericity derived from electroanatomical mapping for prediction of atrial fibrillation progression: a machine learning approach

Background/Introduction Electroanatomical mapping (EAM) during atrial fibrillation (AF) ablation provides detailed characterization of the left atrial (LA) substrate, yet the prognostic utility of this data remains underexplored. Purpose We aimed to identify the EAM-derived predictors of AF progression and evaluate whether they provide independent prognostic value beyond a machine learning (ML) clinical risk model. Methods We analysed 1116 patients who underwent AF ablation (1/2017-10/2024). AF progression was defined as the transition from paroxysmal to persistent or persistent to permanent AF. Clinical risk model development: a ML survival model (XGBoost) using 30 clinical variables was developed on the full cohort via nested cross-validation, generating unbiased out-of-sample risk predictions for each patient. EAM feature selection: In a sub-cohort of 251 patients with EAM data, LA morphological and voltage substrate features were automatically quantified using a proprietary platform. Elastic Net Cox regression with 100-bootstrap stability selection identified the most robust EAM predictors. Independent value testing: selected EAM predictors were tested in a multivariable Cox model adjusted for the clinical risk score in the EAM cohort. Results The cohort (58.4±11.3 years, 66.0% male, median CHA2DS2-VASc 1) mainly featured paroxysmal AF (72.5%, median 2.0-year history) and primarily underwent radiofrequency ablation (75%). AF progression occurred in 121 patients (10.8%) over median follow-up of 742 days (29-2095 days; 3.55/100 person-years). The clinical ML model demonstrated moderate discrimination (Uno C-index: 0.65, 95% CV: 0.61-0.71), and outperformed traditional scores (HATCH, CHA2DS2-VASc) at all time horizons (1–5-year AUC: 0.61-0.74). LA volume, age and prior cardioversions were the most influential clinical predictors. The EAM sub-cohort had similar baseline clinical risk to the remaining cohort (6.8% vs. 6.2%, p=0.52), with progression observed in 19 patients (7.6%). EAM analysis demonstrated mean surface area of 77.0±20.5 cm², volume of 28.8±13.9 mL, sphericity of 58.4±14.3, median voltage of 2.1 mV (IQR: 1.49–2.72), and a median low voltage zone burden (<0.2 mV) of 1.54% (IQR: 0.24%–3.88%). Among 41 EAM parameters, LA anterior wall sphericity (anterior wall width/height ratio) was the most stable predictor (>60% bootstrap selection frequency). It was higher in patients who subsequently progressed (median 90.6% vs. 62.8%, p=0.035). Adjusted for the comprehensive ML clinical risk score, higher anterior sphericity remained an independent predictor of AF progression (HR 1.36, 95% CI: 1.01-1.84, p=0.04). Conclusions LA anterior sphericity provides independent prognostic value for AF progression beyond comprehensive clinical risk assessment. This EAM-derived morphological marker represents a novel substrate feature that may enhance risk stratification after AF ablation.Feature ImportanceAUC