Prediction of Neurological Functional Recovery After Carotid Endarterectomy Using Machine Learning and Carotid Computed Tomography Angiography Radiomics.
Authors
Affiliations (9)
Affiliations (9)
- Department of Neurosurgery, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
- Department of Neurosurgery, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
- Radiology Department, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
- Qingdao Central Hospital, University of Health and Rehabilitation Sciences, Qingdao City, 266000, China. Electronic address: [email protected].
- Department of Neurosurgery, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
- Department of Neurosurgery, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
- Department of Neurosurgery, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
- Department of Neurosurgery, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
- Department of Neurosurgery, The Affiliated Hospital of Qingdao University, Qingdao City, 266000, China. Electronic address: [email protected].
Abstract
To develop and validate machine learning models based on carotid computed tomography angiography (CTA) radiomics for predicting neurological functional recovery in patients with symptomatic carotid artery stenosis (SCAS) after carotid endarterectomy (CEA). This retrospective study analyzed 244 SCAS patients undergoing CEA. Based on modified Rankin Scale improvement at 3-month follow-up, patients were classified into good (n=164) and poor (n=80) recovery groups. Radiomic features were extracted from carotid plaque regions on CTA images, with feature selection performed using least absolute shrinkage and selection operator (LASSO) regression. Five machine learning algorithms-logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), light gradient boosting machine (LightGBM), and multilayer perceptron (MLP)-were developed using radiomics, clinical, and combined feature sets. Among all models, the SVM radiomics model demonstrated optimal performance with 90% accuracy, 91% sensitivity, and 88% specificity. The LR and LightGBM radiomics models also showed robust and stable predictive ability, achieving accuracies of 88% and 86% respectively, with balanced sensitivity and specificity. While KNN and MLP radiomics models exhibited relatively lower performance, their combined models incorporating clinical features demonstrated substantial improvement in predictive capability. The integration of radiomic and clinical features consistently enhanced model performance across all algorithms. The SVM, LR, and LightGBM models, particularly those combining radiomic and clinical features, demonstrated good performance in predicting neurological recovery in patients with SCAS following CEA. These models may offer a useful framework for postoperative risk stratification and support future efforts toward individualized perioperative management and rehabilitation planning.