Integrating preoperative multiregion radiomic features with clinical data to predict atrial fibrillation recurrence after radiofrequency ablation

Figures

• 

  Figure 1. Flow chart of study design.

• 

  Figure 2. Summary of the clinical and radiomic feature-selection workflow. (a,b) Forest plots of univariable and multivariable logistic regression for clinical features. (c) LASSO coefficient path plot. (d) Cross-validation curve for LASSO regression. (e) Coefficients of LASSO-selected radiomic features. (f,g) Clinical and radiomic feature-screening summaries. (h) Final 12 radiomic features ranked by random-forest Gini importance. All feature-selection procedures were performed in the training cohort.

• 

  Figure 3. SHAP-based interpretation of the optimal clinical, radiomic, and fusion models. Beeswarm plots summarize global feature contributions for the clinical (a), radiomic (d), and fusion (g) models. Waterfall plots illustrate representative individual predictions (b,e,h), and heatmaps display feature-contribution patterns across selected patients (c,f,i).

• 

  Figure 4. Discrimination performance of all sub-models. Receiver operating characteristic curves illustrate the predictive performance of the clinical (A–C), radiomic (D–F), and fusion (G–I) models in the training, internal validation, and external validation cohorts, respectively. Shaded areas represent 95% confidence intervals.

• 

  Figure 5. Pairwise comparison of the optimal clinical, radiomic, and fusion models. DeLong tests compare the AUROCs of the optimal clinical, radiomic, and fusion models in the training (a), internal validation (b), and external validation (c) cohorts.

• 

  Figure 6. Pairwise comparison of discrimination among submodels. Heatmaps present DeLong test P values for pairwise AUROC comparisons among the five algorithms within the clinical (A–C), radiomic (D–F), and fusion (G–I) models across the training, internal validation, and external validation cohorts, respectively.

• 

  Figure 7. Clinical utility of all sub-models. Decision curve analysis evaluates net benefit across threshold probabilities for the clinical (A–C), radiomic (D–F), and fusion (G–I) models in the training, internal validation, and external validation cohorts, respectively.

• 

  Figure 8. Calibration performance of all submodels. Calibration plots assess agreement between predicted probabilities and observed AF recurrence in the clinical (A–C), radiomic (D–F), and fusion (G–I) models across the training, internal validation, and external validation cohorts, respectively. The diagonal line represents perfect calibration.