Abstract
Background: Atrial fibrosis is a key contributor to atrial cardiomyopathy and can be assessed invasively using mean left atrial voltage (MLAV) from electroanatomical mapping. However, the invasive nature of this procedure limits its clinical applicability. Machine learning (ML), particularly regression tree-based models, may offer a non-invasive approach for predicting MLAV using clinical and echocardiographic data, improving non-invasive atrial fibrosis characterisation beyond current dichotomous classifications. Methods: We prospectively included and followed 113 patients with paroxysmal or persistent atrial fibrillation (AF) undergoing pulmonary vein isolation (PVI) with ultra-high-density voltage mapping (uHDvM), from whom MLAV was estimated. Standardised two-dimensional transthoracic echocardiography was performed before ablation, and clinical and echocardiographic variables were analysed. A regression tree model was constructed using the Classification and Regression Trees-CART-algorithm to identify key predictors of MLAV. Results: The regression tree model exhibited moderate predictive accuracy (R(2) = 0.63; 95% CI: 0.55-0.71; root mean squared error = 0.90; 95% CI: 0.82-0.98), with indexed minimum LA volume and passive emptying fraction emerging as the most influential variables. No significant differences in AF recurrence-free survival were found among MLAV tertiles or model-based generated groups (log-rank p = 0.319 and p = 0.126, respectively). Conclusions: We present a novel ML-based regression tree model for non-invasive prediction of MLAV, identifying minimum LA volume and passive emptying fraction as the most significant predictors. This model offers an accessible, non-invasive tool for refining atrial cardiomyopathy characterisation by reflecting the fibrotic substrate as a continuum, a crucial advancement over existing dichotomous approaches to guide tailored therapeutic strategies.