Machine Learning Applied to CT Radiomics Identifies Symptomatic Carotid Plaques

Introduction: This study presents the development and validation of a multimodal framework that incorporates CT-based radiomics, machine learning algorithms, and clinico-radiologic features to enhance the detection accuracy of symptomatic carotid plaques. Clinically applicable nomograms were established for personalized risk stratification and management in asymptomatic patients. Method: This retrospective study analyzed computed tomography angiography (CTA) data from 206 head and neck patients at Ordos Central Hospital (January 2020-June 2024), including 157 symptomatic and 49 asymptomatic individuals. Patients were randomly allocated into training/validation cohorts (7:3 ratio), with an additional prospectively enrolled cohort of 44 carotid plaque cases from Jining First People’s Hospital serving as an independent test set. Following radiomics feature extraction from CTA images, machine learning models were constructed using tree-based algorithms (Random Forest, XGBoost, LightGBM). An integrated multimodal model combining clinical variables, radiological characteristics, and radiomics signatures was developed. Performance was assessed via receiver operating characteristic (ROC) analysis, reporting area under the curve (AUC), sensitivity, and specificity metrics. Results: Multivariable analysis identified high-density lipoprotein cholesterol (HDL-C) levels and a history of diabetes as independent predictors of symptomatic carotid plaques. Radiomics models utilizing machine learning demonstrated moderate to strong diagnostic capability, yielding AUCs of 0.625-0.814 in the validation cohort and 0.743-0.802 in the test cohort. The multimodal integrated framework consistently surpassed standalone models, attaining AUCs of 0.836 in the validation set and 0.845 in the test set, which were significantly higher than both clinical and radiomics predictors alone. Discussion: The CT radiomics-based machine learning model developed in this study demonstrated favorable diagnostic efficacy in discriminating symptomatic carotid plaques. The multimodal framework integrating clinical indicators with radiomic signatures significantly enhanced the model's discriminative power and generalizability. This work confirms the clinical potential of radiomics analysis derived from routine CTA for precision risk assessment in cardiovascular disease, providing a novel supportive tool for individualized risk stratification of carotid atherosclerosis. Conclusion: CT imaging-based radiomics and machine learning models can effectively support clinical decision-making, enhance risk stratification for carotid plaque patients, and facilitate personalized treatment strategies.