Multimodal radiomics prediction of long-term postoperative prognosis in non–small cell lung cancer patients based on CT imaging

To construct a multi-modal radiomics prediction model based on CT imaging and evaluate its value in predicting postoperative prognosis of non-small cell lung cancer (NSCLC) patients. A retrospective analysis was conducted on 298 NSCLC patients who underwent radical surgical treatment, randomly divided into training set (n=208) and internal validation set (n=90) at a 7:3 ratio. PyRadiomics was used to extract 963 CT radiomics features, and a radiomics score (Rad-score) was constructed through variance screening, correlation analysis, and LASSO regression. Cox regression analysis was used to identify independent prognostic factors and construct clinical model, radiomics model, and combined model. Model performance was evaluated using C-index, time-dependent ROC curves, and decision curve analysis. Variance inflation factor (VIF) analysis was performed on the 13 retained radiomics features to assess multicollinearity. All features demonstrated VIF<3.0 (range: 1.12-2.87), well below the conventional threshold of 5.0, confirming that LASSO regularization effectively addressed feature redundancy. Thirteen radiomics parameters were identified for Rad-score development. Multivariable analysis revealed TNM staging (Stage III versus I: HR=2.234, 95%CI: 1.338-3.731, P=0.002), blood vessel infiltration (HR=1.623, 95%CI: 0.992-2.657, P=0.041, approaching statistical significance), and Rad-score (HR=1.854, 95%CI: 1.351-2.543, P<0.001) as independent survival predictors. The integrated framework achieved C-index values of 0.742 (95%CI: 0.688-0.796) in development and 0.728 (95%CI: 0.651-0.805) in testing cohorts, demonstrating significant superiority over individual clinical and radiomics frameworks (P<0.05). Rad-score-based risk categorization revealed 3-year survival rates of 61.2% versus 85.7% (P<0.001) for high-risk and low-risk populations in the development cohort, and 57.5% versus 84.0% (P<0.001) in the testing cohort. The integrated prediction framework combining CT radiomics characteristics with clinical-pathological variables demonstrates effective prognostic capability for NSCLC patients following surgery, enables accurate risk categorization, and supports evidence-based personalized therapeutic planning.