In this study, the data analyzed were acquired from the Hospital Information System of Beijing Chaoyang Hospital, Capital Medical University, from January 2019 to June 2021. We screened the medical records of 1,120 patients admitted to our medical center for receiving RFCA for AF. The inclusion criteria were as follows: age between 18 and 80 years; paroxysmal or persistent AF; and successfully performed RFCA for AF. The exclusion criteria were as follows: previous RFCA for AF; significant valvular heart disease; chronic heart failure; chronic obstructive pulmonary disease; autoimmune disease, including hyperthyroidism; missing follow-up data; or moderate to severe hepatic or renal dysfunction. We ultimately included 500 patients in this research (Figure 1). The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Beijing Chaoyang Hospital, Capital Medical University (2023-S-454).

Transthoracic echocardiography was performed to record the left-atrial diameter (LAD) in the four-cavity view, and left-ventricular end-diastolic diameter (LVEDD) and left-ventricular ejection fraction (LVEF) in the M mode. Transesophageal echocardiography was arranged to ensure the absence of intracardiac thrombus. Using a three-dimensional electro-anatomical mapping system, CARTO, continuous circumferential ablation that encircled the pulmonary vein ostia was performed. If AF persisted or recurred during paced rhythm after pulmonary vein isolation, additional substrate modification could be required. The endpoint of ablation was considered to be a bi-directional conduction block around linear ablation lesions. Patients were heparinized during the ablation procedure, maintaining an activation coagulation time of 250–350 s. After ablation, patients underwent electrocardiogram telemetry monitoring for 24 h.

After discharge, patients were scheduled for regular visits in an outpatient clinic at 3, 6, 12, 18, and 24 months. During each visit, patients underwent physical examination, were intensively questioned about arrhythmia-related symptoms, and underwent a 12-lead electrocardiogram. Meanwhile, a 24-h Holter recording was regularly performed every six months. The recurrence of AF was defined as any symptomatic or detected episode of AF, atrial flutter, or atrial tachycardia longer than 30 s beyond the blanking period of 3 months (14).

Five hundred eligible patients were randomly assigned to a training cohort and a validation cohort at a 6:4 ratio. General data, relevant information about AF, and medical history were extracted from the electronic medical records. Moreover, echocardiography data and laboratory results including hemoglobin, alanine aminotransferase, creatinine, NT-proBNP, and anti-M2-R were collected. The selection of putative predictors was based on previous studies and the investigators' clinical experience.

We used 10-fold cross-validation Lasso regression for variable selection among 27 variables. The variables included in the final model was determined based on the location of the corresponding lambda.1SE. Multivariate logistic regression was used to screen the independent factors for AF recurrence among the variables selected above. Then, a nomogram was established. The C-index was used to assess discrimination ability and the calibration curve was used to evaluate the calibration ability of the nomogram. Meanwhile, bootstrap resampling (bootstrap = 500) was used in the internal validation. To evaluate the practicality and clinical utility of the nomogram, we utilized decision curve analysis (DCA). A flowchart illustrating the model construction and validation process is shown in Figure 1.

We also calculated several other classic prediction scores for AF recurrence after RFCA therapy, including the APPLE score [age > 65 years, persistent AF, impaired eGFR (<60 ml/min/1.73 m²), LAD ≥ 43 mm, EF < 50%] (7), CAAP-AF score (coronary artery disease, atrial diameter, age, persistent or long-standing AF, number of antiarrhythmic drugs failed, and female sex) (8), and MB-LATER score [male sex, bundle branch block, LAD ≥ 47 mm, type of AF (paroxysmal, persistent, or long-standing persistent), and early recurrent AF] (9). Finally, the integrated discrimination improvement (IDI) and net reclassification improvement (NRI) indices were calculated to compare the performance of our nomogram with the above models.

Quantitative data were expressed as the mean ± standard deviation (SD) or median (Q1, Q3). Categorical variables were expressed as frequencies and percentages. Differences between the two groups were compared by the unpaired Student's t-test or the Mann-Whitney U-test for quantitative data and the chi-square test for categorical variables. Lasso regression was performed using the “glmnet” package for variable shrinkage and selection. The “glm” package was used for multivariate logistic regression analysis. The nomogram, DCA, and calibration curve were plotted by R package “rms.” The ROC curve was plotted by the “ggROC” package. The IDI and NRI indices of multiple models were compared separately using the “PredictABEL” and “nricens” packages. Two-tailed P-values lower than 0.05 was considered statistically significant. Statistical analysis was performed by R 4.2.1.

The clinical characteristics of the 500 patients, including 300 patients in the training cohort and 200 patients in the validation cohort, are presented in Table 1. The median age was 64.2 ± 7.4 years. A total of 337 patients (67.4%) had paroxysmal AF, and 163 patients (32.6%) had persistent AF. The median course of AF was 2.0 years, and the median CHA₂DS₂-VASc score was 2. The baseline data were matched between the two cohorts.

The new prediction model was developed in the training cohort with 78 cases of AF recurrence. Lasso regression with 10-fold cross-validation was performed for variable selection. Ten variables, including persistent AF, course of AF, alcohol consumption, coronary artery disease, eGFR, LAD, LVEDD, LVEF, NT-proBNP, and anti-M2-R, were selected based on the lambda.1SE with the criteria of “family = binomial”. The variable shrinkage and cross-validation results are shown in Figure 2. After multivariate logistic regression with the “backward” method, six variables were selected in this model. The regression equation of the model was: L = Persistent AF + Course of AF × 0.463/1.512 + LAD × 0.098/1.512−eGFR × 0.074/1.512 + NT-proBNP × 0.009/1.512 + Anti-M2-R × 0.008/1.512. The risk factors detected by multivariate logistic regression analysis for the recurrence of AF after first-time RFCA are shown in Table 2.

A nomogram was constructed using the six variables, namely persistent AF, course of AF, eGFR, LAD, NT-proBNP, and anti-M2-R to predict AF recurrence after first-time RFCA during a 2-year follow-up period (Figure 3). The score for each factor can be obtained by locating the corresponding values on the vertical line. To calculate the total score of the nomogram, the scores of each factor were added. The corresponding prediction probability is the incidence of AF recurrence.

The receiver operator characteristic (ROC) curve for this new prediction model regarding the AF recurrence is shown in Figure 4A. The C-index was 0.862 [95% confidence interval (CI), 0.815–0.909] in the developing cohort. The calibration plot and DCA curves for this prediction model are shown in Figures 4B,C, respectively.

We proved this prediction model in the validation cohort. The ROC curve is shown in Figure 5A, and the C-index was 0.831 (95% CI, 0.771–0.890). The calibration curve of this prediction model is shown in Figure 5B, suggesting a good coherence between the prediction results and the actual results for AF recurrence. In addition, the DCA showed that the nomogram had a high overall net benefit (Figure 5C).

To investigate the benefit of our new model, we calculated the APPLE score, CAAP-AF score, and MB-LATER score. The area under the curve (AUC) of the new model demonstrated significant differences compared with the APPLE score, CAAP-AF score, and MB-LATER score (Table 3, Figure 6A). We compared the nomogram with these models and evaluated the IDI and NRI indices. It can be seen that the IDI and NRI indices of the nomogram were noticeably higher than those of other models, indicating a positive improvement in efficacy and emphasizing its advantage in predicting the recurrence of AF for a 2-year period (Figure 6, Table 4).

This single-center study underscores the necessity for external validation to ensure the clinical generalization and applicability of the model. In addition, our study was limited by its retrospective nature. Some biomarkers with potential predictive power could not be systematically monitored during the perioperative period and therefore were not included in this model. Consequently, further prospective multi-center studies are much needed.