Epicardial Adipose Tissue CT Radiomics Improves Acute Coronary Syndrome Prediction Beyond Coronary Artery Calcium Score

Figures

• 

  Figure 1. Study cohort selection and propensity score matching flowchart.

• 

  Figure 2. Epicardial adipose tissue (EAT) segmentation on cardiac CT. The EAT volume is highlighted in purple (left panel). The unsegmented image is shown on the right panel.

• 

  Figure 3. Standardized Mean Differences for Covariates Before and After Propensity Score Matching.

• 

  Figure 4. Post-Propensity Match ROC Curves for Risk Prediction Models on the Matched Training and Test Sets. Receiver operating characteristic curves compare the predictive performance of different ACS risk models in the training set (A) and matched test set (B). The radiomics model significantly outperformed all others, with an AUC of 0.91 ± 0.01 in the training set and 0.89 ± 0.03 in the matched test set (p < 0.01 for all comparisons). The shaded areas represent the 95% confidence interval of model performance across cross-validation folds. The dotted diagonal line indicates the line of no discrimination (AUC = 0.5).

• 

  Figure 5. ROC Curves for Risk Prediction Models on the Test Set. The receiver operating characteristic curves for the complete unmatched test set compare the predictive performance of various risk models for ACS. The radiomics model demonstrated significantly better performance with an AUC of 0.89 ± 0.03 (p < 0.01 for all). The shaded areas represent the 95% confidence interval of model performance across cross-validation folds. The dotted diagonal line indicates the line of no discrimination (AUC = 0.5).

• 

  Figure 6. Calibration curves of the radiomics model. Calibration curves for the radiomics model illustrate the agreement between predicted probabilities and observed outcomes in the training set (A) and test set (B). The dashed line represents perfect calibration, while the solid blue line indicates model predictions. Minor under-prediction at mid-to-high predicted probabilities was observed in the test set. Hosmer–Lemeshow goodness-of-fit tests (p > 0.05) indicated overall acceptable calibration for both datasets.