This study involved 1,551 patients who underwent PVI in the Cardiac Pacing and Electrophysiology Department of the First Affiliated Hospital of Xinjiang Medical University. The definition of AF recurrence is that, 3 months after ablation surgery, any episode of AF, atrial flutter, or atrial tachycardia lasting for more than 30 s is recorded through electrocardiogram monitoring. All patients in the non-recurrence group received at least 6 months of follow-up, during which no documented recurrence of arrhythmia was observed. The patient data included clinical examination data of the first preoperative examination and CTA imaging data of the left atrium and pulmonary vein. Of note, all patients at this center undergo routine imaging before AF ablation procedure, not leading to a substantial selection bias. Detailed inclusion/exclusion criteria and enrollment process are illustrated in Figure 2. Among all eligible cases, 160 patients had a recurrence of AF after catheter ablation for PVI. In contrast, 1,391 patients did not experience recurrence. To ensure data integrity and mitigate potential confounding biases, we applied predefined exclusion criteria. A total of 645 cases were excluded due to incomplete or missing key clinical variables. To evaluate whether this exclusion introduced systematic differences, we compared the available baseline characteristics between the primary cohort (n = 1,551) and those retained in the prematching cohort (n = 906). The comparison results showed no statistically significant differences between the two groups on most available variables, indicating that the exclusions were primarily driven by data completeness rather than specific clinical phenotypes, This mitigates, to some extent, the selection bias introduced by the exclusion process. Detailed comparison results are provided in Supplementary Table S1. Following this data quality control step, 746 patients without recurrence were retained for subsequent analysis. To enhance the comparability between the recurrence and non-recurrence groups, a 1:1 propensity score matching (PSM) was performed. Matching was based on relevant preoperative clinical characteristics. This process yielded a final well-balanced study population of 302 patients, comprising 151 with AF recurrence and 151 without.

In this study, propensity scores were estimated using a multivariate logistic regression model, and the nearest neighbor method was adopted for matching. To minimize potential confounding biases and enhance the comparability between groups, a 1:1 matching was performed between patients with recurrent AF requiring repeat ablation and those without recurrence. This score is defined as the conditional probability of an individual being assigned to a specific group under the condition of a given set of baseline covariates. Then, based on this score, matching is performed between the recurrence group and the non-recurrence group to simulate the effect of a randomized controlled trials, and a cohort with balanced distribution of key baseline characteristics and stronger comparability is constructed. To prevent information leakage from correlated matched pairs during cross-validation, the propensity-matched pairs were kept intact and assigned as a single unit to each fold. Matching was performed using the MatchIt package in R and the following covariates were included for stratification: gender, age, body mass index (BMI), vital signs (body temperature and respiratory rate), smoking status, alcohol use, history of diabetes, hypertension, and surgical history. The balance of baseline characteristics between the matched cohorts was assessed using the standardized mean difference (SMD), with an SMD < 0.1 indicating adequate balance. Finally, the predictive performance of the model was assessed in the matched cohort (20, 21).

The CT image data used in the study were collected according to the routine contrast-enhanced imaging protocol. The scan extended from the level of the main pulmonary artery to approximately 2 cm below the diaphragmatic surface of the heart. Before imaging, all patients underwent uniform breathing training. A contrast agent (Ioversol, 350 mg/mL) was administered intravenously via the antecubital vein at a dose of 80–100 mL and an injection rate of 4–5 mL/s. This was followed by a 30 mL saline flush at a flow rate of 4 mL/s.

Medical radiomics research highly relies on accurate labeling data. However, the traditional manual labeling method has inherent limitations such as low efficiency, high cost and poor consistency. In order to break through this bottleneck, a semi-automatic annotation method was developed. Its core value lies in the following: the initial contour is generated with the assistance of algorithms, which significantly improves the annotation efficiency and reduces the labor cost. At the same time, this method can effectively standardize the labeling process, enhance the consistency and repeatability of the data, and provide key technical support for the construction of high-quality data sets and the acceleration of radiomics model development and clinical transformation. To address the challenges of manual annotation, we implemented the SwinUNETR model for semi-automatic segmentation (22). Its integration of a shifted window transformer mechanism allows it to effectively capture global contextual information, leading to improved accuracy and robustness in segmenting complex structures from 3D CT scans. This model was trained on a manually annotated dataset comprising 50 patients. The average number of images per case is 218, for a total of 10,900 images. To generate a high-confidence ground truth for model training, the image annotations were performed through a collaborative consensus process by two radiologists with 6 years of experience, using 3D Slicer (version 5.2.2). A standardized protocol was strictly followed: the delineation plane extended from the lower margin of the right pulmonary trunk to the cardiac apex, with a consistent fat attenuation threshold defined as −190 to −30 HU. During the annotation sessions, the radiologists worked jointly to delineate each case, discussing and resolving any ambiguities regarding EAT boundaries in real time to reach a consensus contour. To further ensure the quality and consistency of the final dataset, all consensus segmentations underwent a final review by a senior cardiothoracic imaging specialist. This rigorous, consensus-driven approach was adopted to prioritize the creation of a single, reliable “gold standard” segmentation per patient, which is crucial for the supervised training of the deep learning model. MONAI frameworks were utilized for model training, with the training set and validation set divided at a ratio of 4:1. The manually annotated dataset comprising 50 patient scans was divided in a stratified manner, with 40 scans allocated for training and 10 scans held out as an independent validation set. This validation set was strictly reserved for the final evaluation of the trained model's segmentation performance and was not used during any stage of model training or hyperparameter tuning. Before model training, spatial transformations, intensity transformations, and rotational transformations were applied to the original 3D images for image augmentation. During model training, the image size was 512 × 512 × 512, the batch size was set to 2, the patch resolution was 96 × 96 × 96, the initial learning rate was 0.0001, and 10,000 iterations were performed using the AdamW optimizer. The pre-trained model was used to initialize the backbone network's weights. The segmentation performance of the final model was evaluated using the dice similarity coefficient (DSC) to assess volumetric overlap and the hausdorff distance (HD) to quantify boundary accuracy.

Radiomics facilitates the translation of medical images into a high-dimensional, quantitative feature space. By analyzing the intensity, shape, and texture patterns within a region of interest (ROI), it provides a comprehensive characterization of tissue heterogeneity for subsequent data mining. To ensure the reproducibility and standardization of our radiomics analysis in compliance with the image biomarker standardisation initiative (IBSI) guidelines, a rigorous preprocessing and feature extraction pipeline was implemented using open source platform PyRadiomics (version 3.0.1) (23, 24). All segmented EAT volumes were first subjected to a standardized resampling and discretization process. The image voxels were resampled to an isotropic spacing of 2.0 × 2.0 × 2.0 mm³ using B-spline interpolation. Prior to resampling, intensity normalization was applied. A fixed bin width of 25 was used for gray-level discretization to ensure scale-invariant texture analysis. This method is based on the widely accepted recommendations in the field of radiomics, aiming to ensure the standardization of the gray scale for all cases, thereby reducing the bias caused by differences in scanning parameters (25). A total of 111 radiomics features were extracted from each patient's EAT segmentation to quantify its original imaging phenotype. These features were categorized into three classes: (1) 19 first-order features, summarizing the global intensity distribution; (2) 17 shape features, describing three-dimensional geometric properties; and (3) 75 texture features, capturing spatial intensity variation and heterogeneity. The texture features comprised 24 gray-level co-occurrence matrix (GLCM), 16 gray-level run-length matrix (GLRLM), 16 gray-level size zone matrix (GLSZM), 14 gray-level dependence matrix (GLDM), and 5 neighboring gray-tone difference matrix (NGTDM) features.

The correlation clustering diagram of the radiomic features is shown in Supplementary Figure S1, where it can be observed that most of the radiomic features exhibit correlation and redundancy. This redundancy may introduce multicollinearity, which interferes with model training and reduces its generalization ability. In order to screen out the most discriminative feature subset to construct a robust prediction model, a feature importance assessment method based on random forest (RF) was used for feature selection. The principle of using RF for feature selection mainly lies in its ability to quantify the contribution of each feature to the model's prediction. The core idea is that if a feature is more useful for predicting the target variable, then when this feature is used in the model, the improvement in the model's performance should be more significant. It is precisely by leveraging its ability to robustly and efficiently identify key radiological biomarkers and clinical indicators most relevant to AF recurrence from high-dimensional radiomics and clinical features that the model's interpretability and generalization potential are enhanced. To ensure the robustness of the feature selection process and to prevent data leakage that could optimistically bias model performance estimates, all steps were embedded within a nested cross validation framework. In this study, the algorithm calculated and compared the contribution scores of all candidate features. Finally, according to the model performance, 22 key clinical features and 13 radiomics features with the highest discriminative value were retained. Furthermore, to construct a predictive model that extends beyond a single information domain, this study integrated clinical variables with radiomic features and again employed the RF algorithm to evaluate feature importance and identify key biomarkers within the fused high-dimensional feature space. A total of 34 features were finally selected into the fusion model. This process effectively reduced the data dimensionality and improved the clinical interpretability and generalization performance of the model.

To build robust and interpretable predictive models, and to comprehensively evaluate the performance of different algorithms in predicting AF recurrence, this study systematically selected six representative machine learning algorithms with distinct underlying principles: extreme gradient boosting (XGBoost), RF, Bayesian, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Logistic Regression. The selection of this model was based on rigorous methodological considerations to ensure the robustness of the conclusions by covering different modeling paradigms. XGBoost and RF can effectively capture complex feature interactions. Logistic regression and Bayesian provide interpretable probabilistic output frameworks. KNN is used for prediction based on the local structure of the data. SVM is good at dealing with nonlinear problems. This diversification strategy helps the system to identify the optimal algorithm for this clinical prediction task. Before model training, to eliminate the impact of dimensions among different features and make the model more stable and faster converging during the training process, all data were standardized using the z-score method, transforming the data distribution into a standard normal distribution. All models were trained and evaluated under a 10-fold cross validation framework to ensure unbiased performance evaluation. The model was applied to the processed test set to generate predictions. All reported performance metrics (AUC, accuracy, sensitivity, specificity, F1-score) represent the mean ± standard deviation calculated from the predictions across all 10 outer-loop test folds. This rigorous pipeline guarantees an unbiased estimate of model generalizability by preventing any leakage of information from the test sets into the feature selection or model development stages. To comprehensively assess the predictive performance and clinical applicability of our models, we performed several additional analyses beyond conventional metrics. The clinical net benefit of each model was quantified using decision curve analysis (DCA), which assesses model utility across a range of threshold probabilities. Furthermore, to ensure that the predicted probabilities reliably reflected the actual risk, we further applied sigmoid calibration using CalibratedClassifierCV with 5-fold cross validation. Calibration performance was evaluated visually through calibration curves and quantitatively using the Brier score, a proper scoring rule in which lower values indicate better calibration.

The DSC was used to measure the effect of EAT image segmentation, and PSM was employed to control potential confounding variables. Categorical variables are presented as numbers (n) and percentages (%). Continuous variables are expressed as either mean ± standard deviation or median with interquartile range (IQR), based on their distribution. The 95% confidence interval (CI) was applied to assess statistical uncertainty, and the final quantitative results of the model were evaluated using the AUC, accuracy, sensitivity, specificity, and F1 score to assess the predictive performance of AF recurrence after PVI.

After PSM score matching, 302 patients were finally included in the analysis. As shown in Table 1, the SMD for all covariates after matching were less than 0.1, indicating that baseline characteristics were strictly balanced between the groups. After matching, the average age of the case group was 61.83 years, of which 32.5% were women and 67.5% were men. The mean temperature in this group was 36.43 ℃, and the mean respiratory rate was 18.7 beats per minute. In the body mass index (BMI) subgroup, 49% were overweight (BMI: 23.0–27.5) and 33.8% were obese (BMI > 27.5). In terms of lifestyle habits (smoking, alcohol), medical conditions (hypertension, diabetes), and previous surgical history, the distributions were balanced between the case and control groups.

Despite the high anatomical variability of EAT in CTA images, the SwinUNETR model demonstrated robust and accurate segmentation performance. Quantitative evaluation on the independent validation set demonstrated a DSC of 0.87 ± 0.04 and HD of 2.83 ± 0.2, respectively. Representative visual results of the segmentation are provided in Supplementary Figure S2, confirming the model's effectiveness in delineating EAT boundaries. The training process was monitored through the loss and DSC curves. Optimal performance was achieved at 6,100 epochs, corresponding to the highest DSC on the validation set, with both metrics subsequently stabilizing, indicating convergence. The resulting high-quality segmentations provided the precise ROI necessary for subsequent radiomics feature extraction from the EAT.

Following feature selection, 22 key clinical indicators and 13 radiomics features were retained for subsequent predictive modeling. The selected features, along with their importance scores from RF, are detailed in Supplementary Table S2 and Supplementary Table S3, respectively. The features of the fusion model are shown in Supplementary Table S4. The performance of machine learning algorithms was evaluated across three distinct model types: the clinical model (Table 2), the radiomics model (Table 3), and the integrated clinical-radiomics model (Table 4). Evaluation metrics included accuracy, sensitivity, specificity, F1-score, and the AUC. Figure 3 presents the receiver operating characteristic (ROC) curves for the three models: the clinical model (A), the radiomics model (B), and the fusion model (C).

To translate this discriminative ability into clinical value, DCA analysis was performed, and the DCA curve is shown in Figure 4B. At a clinically relevant decision threshold probability of 0.5, all calibrated models provided a positive net benefit, indicating their utility over the treat-none strategy. The XGBoost-based fusion model yielded the highest net benefit of 0.222, followed by the RF model (Net Benefit = 0.195). This translates to 22 more net beneficial interventions per 100 patients when using the XGBoost model for clinical decision-making at this threshold. The quantified net benefit of the calibrated models at a threshold probability of 0.5 is shown in Supplementary Table S5. The calibration performance of all models was significantly improved after sigmoid calibration. Post-calibration, the predicted probability ranges became more constrained and realistic, with means closely aligning with the overall event rate of 0.5. The calibration curves for the six machine learning models after probability calibration are presented in Figure 4A. The probability statistics and Brier scores of the calibrated models are presented in Supplementary Table S6. The Brier score was lowest for the XGBoost model (0.194), confirming its superior probability accuracy. All calibrated models achieved Brier scores below 0.23, indicating a satisfactory level of calibration for clinical interpretation.