Atrial fibrillation (AF) severely impairs patients' quality of life, leads to high morbidity and mortality, and increases healthcare costs. Current classification methods for atrial fibrillation primarily rely on clinical symptoms and examination results, but they struggle to quantify the underlying pathophysiological burden, often leading to treatment strategies that do not match patients' actual risks.

To investigate the factors associated with persistent AF and to develop a classification model based on these factors.

Patients diagnosed with paroxysmal and persistent AF at the First Affiliated Hospital of Xinjiang Medical University between April 2012 and September 2023 were enrolled in this study. Clinical data, including demographic characteristics, biochemical parameters, renal function indices, and cardiac function-related metrics, were collected for analysis. Initially, univariate Logistic regression analysis was performed to screen for variables associated with the type of AF. The Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied for further feature selection to reduce model complexity and prevent overfitting. A multivariate Logistic regression model was then constructed to identify factors independently associated with persistent AF. Utilizing the bootstrap resampling method, the significant variables were incorporated into six machine learning algorithms—Random Forest (RF), Decision Tree (DT), Naive Bayes (NB), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and eXtreme Gradient Boosting (XGB)—to establish classification models. The discriminative performance of these models was evaluated using receiver operating characteristic (ROC) curves. Finally, the SHapley Additive exPlanations (SHAP) method was employed to evaluate the contribution of each variable to the classification models.

A total of 6 938 patients were enrolled, including 5 085 with paroxysmal AF and 1 853 with persistent AF. Univariate Logistic regression analysis initially identified 19 statistically significant independent variables. Following LASSO regression selection, 13 key variables were retained for multivariate Logistic regression analysis. The results indicated that the following were independently associated with persistent AF (all P<0.05): gender (OR=1.248, 95%CI=1.086-1.435), surgical history (OR=0.809, 95%CI=0.706-0.926), BMI (OR=1.028, 95%CI=1.012-1.045), mean platelet volume (MPV) (OR=1.121, 95%CI=1.059-1.186), serum magnesium (Mg_plus2 ) (OR=0.098, 95%CI=0.046-0.208), cardiac output (CO) (OR=1.115, 95%CI=1.009-1.233), left ventricular posterior wall thickness (LVPW) (OR=0.777, 95%CI=0.665-0.909), left atrial diameter (LAD) (OR=1.144, 95%CI=1.123-1.166), left ventricular ejection fraction (LVEF) (OR=0.955, 95%CI=0.938-0.972), right atrial diameter (RAD) (OR=1.031, 95%CI=1.005-1.057), triglycerides (TG) (OR=0.821, 95%CI=0.751-0.898), uric acid (UA) (OR=1.003, 95%CI=1.002-1.003), and left ventricular end-diastolic diameter (LVEDD) (OR=0.903, 95%CI=0.879-0.927). ROC curve analysis demonstrated that the XGB model achieved the best performance (mean AUC=0.823), followed by the SVM model (mean AUC=0.820) and the RF model (mean AUC=0.814). SHAP analysis of the XGB model revealed that LAD and RAD had the highest SHAP values, suggesting that atrial structural parameters exert the greatest influence on model classification.

Increased BMI, male sex, elevated MPV, higher UA, decreased TG levels, decreased Mg_plus2, reduced LVEF, decreased LVEDD, no surgical history, and enlarged RAD and LAD were all closely associated with the occurrence of persistent AF. These clinical parameters can be readily obtained through routine examinations, most of which are non-invasive, and may serve as important clinical indicators for identifying patients with persistent AF, thereby supporting early clinical risk stratification and the development of targeted intervention strategies.