Dual-center Development and External Validation of a Venous-phase CT Radiomics Model for Predicting PD-L1 Positivity in Non-small Cell Lung Cancer

Abstract Background: Programmed death-ligand 1 (PD-L1) expression is a clinically important biomarker for immunotherapy stratification in patients with non-small cell lung cancer (NSCLC). However, tissue-based assessment is limited by sampling bias and intratumoral heterogeneity. This study aimed to develop and externally validate a radiomics model based on venous-phase contrast-enhanced computed tomography (CECT) for non-invasive prediction of PD-L1 positivity in NSCLC. Methods: A total of 259 patients with pathologically confirmed NSCLC from two institutions were retrospectively enrolled and divided into a training cohort (n = 135), an internal validation cohort (n = 50), and an independent external validation cohort (n = 74). Radiomics features were extracted from venous-phase CECT. After least absolute shrinkage and selection operator (LASSO) regression, selected features were used to construct XGBoost, random forest, and decision tree models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration analysis, and decision curve analysis. SHapley Additive exPlanations (SHAP) analysis was performed to improve model interpretability. Results: Among 1,688 extracted radiomics features, two features were ultimately selected for model development. The XGBoost model showed favorable validation performance across cohorts, achieving AUCs of 0.838 (95% CI, 0.769–0.901), 0.801 (95% CI, 0.665–0.919), and 0.767 (95% CI, 0.650–0.873) in the training, internal validation, and external validation cohorts, respectively. In the external validation cohort, the XGBoost model showed significantly better discrimination than the decision tree model while maintaining comparable performance to the random forest model. SHAP analysis demonstrated consistent feature contribution patterns across all cohorts. Conclusions: Venous-phase CECT radiomics enabled moderate but stable discrimination for predicting PD-L1 positivity in NSCLC across independent cohorts. The XGBoost model may serve as a non-invasive complementary decision-support tool for identifying patients with a higher likelihood of PD-L1 positivity.