Abstract
AIMS: Atrial fibrillation (AF) is a common arrhythmia associated with increased risks of stroke and heart failure. Early differentiation between paroxysmal and persistent AF at first diagnosis is critical for guiding treatment decisions. This study aimed to develop an interpretable machine learning model based on structured electronic health records (EHR) to distinguish AF subtypes and identify key contributing factors. METHODS AND RESULTS: In this multicenter, retrospective cohort study, data were collected from three tertiary hospitals in China between January 2013 and January 2023. A total of 11,986 patients with suspected AF were screened, of whom 4155 patients with first-diagnosed AF were included (paroxysmal: 2565 [61.3%]; persistent: 1620 [38.7%]). Structured EHR variables, including clinical demographics, serological indicators, and echocardiographic parameters, were extracted for analysis. Variable selection was performed using Spearman correlation and least absolute shrinkage and selection operator regression. Three machine learning algorithms were trained and externally validated. The CatBoost model achieved the best performance, with an area under the receiver operating characteristic curve of 0.876 (95% CI: 0.871–0.880) and accuracy of 0.808 (95% CI: 0.803–0.816). Sensitivity and specificity ranged from 0.802 to 0.811. Shapley additive explanations (SHAP) were used to interpret model outputs and identify variables most associated with AF subtype classification. CONCLUSION: This multicenter study demonstrates that interpretable machine learning models based on structured EHR data can accurately distinguish paroxysmal from persistent AF at first diagnosis. The proposed model may facilitate early subtype-specific risk stratification and personalized treatment, potentially improve outcomes, and reduce disparities in AF care across different medical conditions. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s13040-026-00525-5.