Detecting severe coronary artery stenosis in T2DM patients with NAFLD using cardiac fat radiomics-based machine learning

Figures

• 

  Fig. 1. Flowchart depicting patient selection criteria and process in the study. T2DM, type 2 diabetes mellitus; NAFLD, non-alcoholic fatty liver disease; CAD, coronary artery disease; CCTA, coronary computed tomographic angiography; PCI, percutaneous coronary intervention; CABG, coronary artery bypass grafting.

• 

  Fig. 2. Coronary imaging and cardiac fat tissue segmentation for a 57-year-old male patient with a CAD-RADS of 4. (A) Tridimensional reconstruction of the heart. White arrow indicates sites of severe stenosis; (B) Volume-rendered coronary tree. White arrow indicates sites of severe stenosis; (C) CT scan axial slice with PAT and EAT, yellow indicates EAT and red indicates PAT; (D) 3D slicer visualization of semi-automated segmentation of cardiac adipose tissue, yellow for EAT and red for PAT; (E) Coronal CT view illustrating segmented cardiac adipose tissue; (F) Sagittal CT reconstruction with mapped cardiac adipose tissue, yellow for EAT and red for PAT. T2DM, type 2 diabetes mellitus; NAFLD, non-alcoholic fatty liver disease; EAT, epicardial adipose tissue; PAT, paracardial adipose tissue; CT, computed tomography.

• 

  Fig. 3. Radiomics and machine learning workflow for coronary stenosis detection in T2DM with NAFLD, including image acquisition and segmentation, radiomics feature extraction, feature selection, and model development with machine learning. CT, computed tomography; T2DM, type 2 diabetes mellitus; NAFLD, non-alcoholic fatty liver disease; EAT, epicardial adipose tissue; PAT, paracardial adipose tissue; LR, logistic regression; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting.

• 

  Fig. 4. Radiomics feature selection of cardiac adipose tissue identified by LASSO regression with tenfold cross-validation. (A) LASSO coefficient profiles of EAT radiomics features. The red vertical line shows the best λ value that resulted in 5 features with non-zero coefficients; (B) Contribution of five EAT features with non-zero coefficients to radiomics and their respective coefficients. (C) LASSO coefficient profiles of PAT radiomics features. The red vertical line shows the best λ value that resulted in 5 features with non-zero coefficients; (D) Contribution of five PAT features with non-zero coefficients to radiomics and their respective coefficients. LASSO, least absolute shrinkage and selection operator; EAT, epicardial adipose tissue; PAT, paracardial adipose tissue.

• 

  Fig. 5. Detection performance of radiomic, clinical, and combined feature models in the training set. (A) ROC curves of the radiomics model in the training set; (B) Calibration curves of the radiomics model in the training set; (C) DCA curves of the radiomics model in the training set; (D) ROC curves of the clinical model in the training set; (E) Calibration curves of the clinical model in the training set; (F) DCA curves of the clinical model in the training set. (G) ROC curve of the radiomics-clinical model in the training set; (H) Calibration curves of the radiomics-clinical model in the training set; (I) DCA curves of the the radiomics-clinical model in the training set. LR, logistic regression; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting; ROC, receiver operating characteristic curves; AUC, Area under the curve; DCA, decision curve analysis.

• 

  Fig. 6. Performance of radiomics, clinical, and combined radiomics-clinical models in the validation set for disease assessment. (A) ROC curve of the XGBoost-based radiomics model in the validation set; (B) Calibration curves of XGBoost-based radiomics model in the validation set; (C) DCA curves of XGBoost-based radiomics model in the validation set; (D) ROC curve of the SVM-based clinical model in the validation set; (E) Calibration curve of the SVM-based clinical model in the validation set; (F) DCA curves of the SVM-based clinical model in the validation set. (G) ROC curve of the radiomics-clinical model with XGBoost in the validation set; (H) Calibration curve of the radiomics-clinical model with XGBoost in the validation set; (I) DCA curves of the radiomics-clinical model with XGBoost in the validation set. LR, logistic regression; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting; ROC, receiver operating characteristic curves; AUC, Area under the curve; DCA, Decision curve analysis.

• 

  Fig. 7. Radiomics-clinical model interpretation using SHAP. (A) Mean SHAP values of top features in radiomics-clinical models. Bars indicate the mean absolute SHAP values, quantifying the average contribution of the top features to the model’s detection performance. (B) SHAP value beeswarm plot of top features in radiomics-clinical models. Each point represents the SHAP value of a patient feature. Pink has a positive effect on the detection and blue has a negative impact. The horizontal axis measures the SHAP value’s effect on model output, and the vertical axis displays the features ranked by their mean absolute SHAP values.