Risk prediction model for lower extremity artery bailout stenting based on machine learning and CT radiomics

To explore radiomics features derived from preoperative lower extremity arterial CT angiography (CTA) images and develop a machine learning model to predict the risk of bailout stenting in patients with lower extremity artery disease (LEAD) undergoing endovascular treatment. Furthermore, to enhance the model's interpretability using the SHAP (SHapley Additive exPlanations) method. We retrospectively collected clinical and imaging data from 118 lower extremity patients who underwent femoropopliteal arterial intervention at our hospital from January 2020 to December 2025. Patients were divided into a balloon-only group and a stent group based on whether a stent was implanted. We used the LASSO regression algorithm to screen for key radiomics features, and systematically compared the predictive performance of eight mainstream machine learning algorithms, including Random Forest, Support Vector Machine, and XGBoost. Model performance was evaluated using 5-fold nested cross-validation, and the SMOTE oversampling technique was introduced to address the issue of class imbalance. Based on the optimal model, we constructed a fusion model incorporating both clinical features, and employed the SHAP method to interpret the model’s decision-making mechanism. The Random Forest model demonstrated the best overall performance among all algorithms, achieving an AUC of 0.894 (95% CI 0.841–0.939) for predicting bailout stenting. It also yielded the lowest Brier score (0.12) and exhibited the optimal calibration curve. Furthermore, both the SVM (AUC = 0.886) and XGBoost (AUC = 0.882) models showed favorable generalization capabilities. Notably, the combined model incorporating clinical features (lesion type and TASC classification) maintained highly robust predictive performance (AUC = 0.893), confirming the independent value of radiomics features in capturing the micro-heterogeneity of plaques. SHAP analysis further revealed that high-order texture features were the most critical biomarkers driving the model’s predictions. The Random Forest model based on CT imaging radiomics can non-invasively and accurately predict the risk of bailout stenting in lower extremity arterial intervention. However, given the single-center retrospective design and limited sample size, these findings are exploratory and hypothesis-generating. The model requires external validation before it can be considered for real-world decision support.