Does a machine learning model integrating CTA-derived radiomics features of carotid plaque and PVAT improve the prediction of ipsilateral stroke recurrence compared to clinical factors alone in patients with carotid atherosclerosis?
Integrating CTA-derived radiomics of carotid plaque and perivascular adipose tissue significantly improves the prediction of recurrent ipsilateral stroke compared to clinical and stenosis factors alone.
Objective To develop and validate an interpretable machine learning (ML) model integrating computed tomography angiography (CTA)-derived radiomics features of carotid plaque and perivascular adipose tissue (PVAT) for predicting ipsilateral stroke recurrence in patients with carotid atherosclerosis. Methods In this retrospective study, patients with unilateral carotid atherosclerosis who underwent head and neck CTA between May 2016 and March 2024 were included and followed for recurrent ischemic stroke detected by follow-up MRI. Radiomics features of carotid plaque and PVAT were automatically extracted using a deep learning-based segmentation model and then gathered to constructed a ML model to predict stroke risk. A conventional clinical model based on carotid stenosis degree and clinical factors was also developed. Additionally, a combined model incorporating both radiomics and clinical factors was constructed. The optimal predictive model was chosen among five ML algorithms based on the area under receiver operating characteristics curve (AUC). Model performance was validated through repeated 10-fold cross-validation and tested in an independent testing cohort. Model interpretability was examined using Shapley Additive Explanations (SHAP). Results Of 162 patients (mean age, 69.28 years ± 8.30 SD; 136 83.95% male) were included, of whom 63 (38.9%) experienced ipsilateral stroke recurrence during follow-up (median, 1 years). The combined model using support vector machines achieved the highest AUC of 0.87(95% CI: 0.74–0.97) in the testing set, higher than the radiomics-only model (AUC, 0.80; 0.63–0.94) and the clinical model (AUC, 0.77; 0.58–0.91; all p 0.05). SHAP analysis demonstrated that plaque texture features contributed strongly to recurrence risk, while PVAT-derived features provided complementary inflammatory information. Conclusion An ML model including both radiomics features of carotid plaque and PVAT can improve performance in predicting ipsilateral stroke recurrence risk than clinical factors alone, offering a promising tool for stroke prevent in clinical practice.
Wei et al. (Thu,) studied this question.