Predicting transient ischemic attack risk in patients with mild carotid stenosis using machine learning and CT radiomics

Figures

• 

  Figure 1. Flowchart for building a predictive model. (A) The carotid CTA image of a patient. (B) The manual VOI segmentation was performed on an image. (C) The radiomic features extracted from each VOI. (D) The feature data were selected. (E) The establishment of prediction models. CTA, computed tomography angiography; VOI, volume of interest; ROC, receiver operating characteristic; AUC, area under curve.

• 

  Figure 2. The procedure of radiomics segmentation. The VOI in each image was segmented with the green arterial lumen and yellow plaques, and a 3D model was synthesized. (A) The edges of atherosclerotic plaques were delineated. (B) Atherosclerotic plaques were segmented. (C) The edges of arteries were delineated. (D) Arteries were segmented. (E) A 3D model was built based on the segmentation of each layer image. (F) Carotid artery CTA corresponding image of the patient. VOI, volume of interest; CTA, computed tomography angiography.

• 

  Figure 3. Feature importance in the random forest model. Radiomics features and combined features show only the top 20 features. Clinical features show all features. (A) The random forest method was used to rank clinical features from high to low importance. (B) The random forest method was used to rank radiomics features from high to low importance. (C) The random forest method was used to rank combined features from high to low importance. BMI, body mass index; UA, uric acid; TG, triglycerides; LDL, low-density lipoprotein; Fib, fibrinogen; HCY, homocysteine.

• 

  Figure 4. Patient selection. The flowchart shows the process of enrolling patients in this study. The original population consisted of 6,314 subjects who had undergone carotid CTA studies, and we first screened the 615 subjects with mild carotid stenosis. According to the inclusion and exclusion criteria, we finally enrolled 179 subjects and included patients with TIA after CTA in the symptomatic group. CTA, computed tomography angiography; TIA, Transient ischemic attack.

• 

  Figure 5. ROC curves of various models. The area under the ROC curves (AUCs) of five different machine learning models constructed from radiomics features, clinical features, and combined features, respectively. (A) ROC curves of five machine learning models built with clinical features. (B) ROC curves of five machine learning models built with radiomics features. (C) ROC curves of five machine learning models built with combined features. ROC, receiver operating characteristic; AUC, area under curve; RF, random forest; XGBoost, eXtreme gradient boosting; LR, logistic regression; SVM, support vector machines; KNN, k-nearest neighbors.

• 

  Figure 6. Comparative analysis of ROC in the random forest model. Among the five models, the random forest model showed the best predictive performance. The random forest model was constructed with radiomics features, clinical features, and combined features, respectively, and the DeLong test was performed. The combined model was significantly better than the clinical model. (A) A DeLong test was used to compare the radiomics model and clinical model. P = 0.813. (B) A DeLong test was used to compare the radiomics model and the combined model. P = 0.098. (C) The DeLong test was used to compare the clinical model and the combined model. P = 0.004. ROC, receiver operating characteristic; RF, random forest.

• 

  Figure 7. Comparative analysis of ROC in the XGBoost model. In the XGBoost model, there were no statistical differences in the models based on clinical features, radiomics features, and combined features. (A) A DeLong test was used to compare the clinical model and the radiomics model. P = 0.598. (B) A DeLong test was used to compare the radiomics model and the combined model. P = 0.676. (C) The DeLong test was used to compare the clinical model and the combined model. P = 0.950. ROC, receiver operating characteristic; RF, random forest.