Radiomics features and clinical factors for predicting restenosis following endovascular therapy in patients with peripheral artery disease

To construct a predictive model for post-endovascular restenosis in patients with peripheral artery disease based on a novel strategy that integrates radiomics features with clinical factors, aiming to enhance predictive performance and risk assessment. A retrospective study was conducted on 110 patients with peripheral artery disease who underwent CT angiography and clinical evaluation. The dataset was randomly divided into training and validation sets in a 7:3 ratio. A combined model was developed using multivariable logistic regression. Subsequently, the model was internally validated and a nomogram was constructed. Additionally, models were trained using SVM, Xgboost, KNN, Adaboost, and CatBoost algorithms. The optimal model was assessed through SHAP analysis to evaluate the importance of each feature in predicting restenosis after endovascular treatment for peripheral artery disease. After standardization, univariate analysis, and LASSO regression with 5-fold cross-validation for dimensionality reduction, five optimal radiomics features and three clinical factors selected via logistic regression were used to construct the model. Among various models, CatBoost demonstrated the highest predictive performance, with the area under the receiver operating characteristic curve (AUC) for the training and validation sets being 0.985 and 0.878, respectively. SHAP interpretability analysis revealed that the four most important global features influencing the CatBoost model output were, in order, Radiomics Score, PACSS, length, and hypertension. The CatBoost model provides superior support for early identification of individuals at higher risk of restenosis after endovascular treatment for peripheral artery disease. It offers a foundation for implementing more targeted prevention and treatment strategies.