Related Experiment Video
Updated: Sep 11, 2026
Estimating Bilateral Atrial Function by Cardiovascular Magnetic Resonance Feature Tracking in Patients with Paroxysmal Atrial Fibrillation
Published on: July 20, 2022
Integrating preoperative multiregion radiomic features with clinical data to predict atrial fibrillation recurrence
Hui Yan1, Xiao-Le Li2, Jun-Hao Mei3
1Department of Radiology, Jintan Affiliated Hospital of Jiangsu University, Changzhou, China.
Objective:
To develop and validate a prediction framework integrating clinical data, left atrium and pulmonary vein morphology, and radiomic features to identify patients at high risk of atrial fibrillation (AF) recurrence after radiofrequency ablation (RFA).
Methods:
Patients with AF who underwent RFA at three centers between August 2018 and October 2024 were retrospectively screened. Patients from two centers were divided into training and internal validation cohorts, and patients from the third center formed the external validation cohort. Clinical and CTA-derived morphological variables and radiomic features from the left atrium and epicardial adipose tissue were analyzed. Clinical, radiomic, and fusion models were evaluated using five machine-learning algorithms. Training performance was estimated from five-fold out-of-fold predictions. Discrimination, calibration, clinical utility, pairwise AUROC differences, and model contributions were assessed using ROC analysis, calibration plots, decision curve analysis, DeLong tests, and SHAP.
Results:
Of 877 patients, 449 formed the training cohort, 192 the internal validation cohort, and 236 the external validation cohort; recurrence occurred in 23.6%, 30.2%, and 35.2%, respectively. The optimal clinical GBDT, radiomic RF, and fusion RF models achieved AUROCs of 0.739, 0.922, and 0.947 in the training cohort; 0.726, 0.837, and 0.839 in internal validation; and 0.722, 0.837, and 0.848 in external validation. Radiomic and fusion models outperformed the clinical model in both validation cohorts, whereas fusion did not significantly outperform radiomic (DeLong P = 0.95 internally and P = 0.43 externally). Decision curves showed net benefit across relevant threshold ranges, while calibration varied across models and cohorts.
Conclusion:
Radiomic models substantially improved post-RFA recurrence discrimination over the clinical model. The fusion model achieved the numerically highest external AUROC, but did not significantly outperform radiomic alone in either validation cohort. These findings support radiomic-based risk stratification while underscoring the need for prospective calibration and validation.
