A multimodal deep learning model for predicting impending rupture in symptomatic abdominal aortic aneurysms using CTA and clinical data

Figures

• 

  Figure 1. Flow chart of the study. AAA, abdominal aortic aneurysm.

• 

  Figure 2. Overall framework of the multimodal model. AAA, abdominal aortic aneurysm; CTA, computed tomography angiography.

• 

  Figure 3. Pearson correlation analysis matrix. (a) Correlation analysis matrix of all clinical features; (b) strongly correlated variables.

• 

  Figure 4. The clinical feature selection based on five machine learning models. (a) The importance ranking of the top six variables based on Random Forest; (b) receiver operating characteristic curves of the clinical feature selection model. RBG, random blood glucose; TC, total cholesterol; DBP, diastolic blood pressure; WBC, white blood cell; AUC, area under the curve; TPR, true positive rate; FPR, false positive rate.

• 

  Figure 5. Performance evaluation of the proposed multimodal model in the development cohort. (a) ROC curve of the multimodal model; (b) confusion matrices of the multimodal model; (c) calibration curve; (d) decision curve analysis; (e) bootstrap confidence intervals. ROC, receiver operating characteristic; AUC, area under the curve.

• 

  Figure 6. Comprehensive performance comparison of the triage models on the development test set. (a) ROC curves for the proposed deep learning, clinical rule-based, and CTA sign models; (b) confusion matrices; (c) calibration curves assessing the agreement between predicted probabilities and observed outcomes; (d) decision curve analysis. ROC, receiver operating characteristic; AUC, area under the curve.

• 

  Figure 7. Performance of the multimodal model on internal temporal validation. (a) ROC curve of the multimodal model; (b) confusion matrices of the multimodal model; (c) calibration curve; (d) decision curve analysis; (e) bootstrap confidence intervals. ROC, receiver operating characteristic; AUC, area under the curve.

• 

  Figure 8. Robustness analysis against domain shifts. The model's performance under temporal drift (development test set) and across four CT scanner vendors (internal temporal test set) was evaluated. (a) ROC curves for temporal subgroups; (b) calibration curves for the temporal subgroups; (c) ROC curves across four CT scanner vendors; (d) calibration curves across vendors. CT, computed tomography; ROC, receiver operating characteristic.

• 

  Figure 9. Grad-CAM visualizations of representative AAA cases from the internal temporal validation cohort. Heatmaps indicate model attention for predicting impending rupture (red = highest, blue = lowest). (a) True positive: The model focuses on the pathological aortic wall exhibiting high-risk CTA signs in a patient with impending rupture. (b) True negative: Minimal attention to the aortic wall in a stable patient correctly classified as low risk. (c) False positive: Attention directed to the peri-aneurysmal area in the absence of definitive high-risk CTA signs, leading to incorrect high-risk classification in a stable patient. (d) False negative: Attention shifts to anatomically irrelevant structures, failing to capture the diseased aortic segment. Grad-CAM, Gradient-Weighted Class Activation Mapping; CTA, computed tomography angiography; AAA, abdominal aortic aneurysm.