Interpretable analysis of time-stratified survival prediction for non-small cell lung cancer patients based on FDG PET/CT radiomics

In this study, baseline FDG PET/CT radiomics and clinical parameters were integrated, a time-stratified model framework was developed to explain the dynamic evolution of prognostic factors in non-small cell lung cancer (NSCLC), and the clinical utility of the radiomics score (Rad score) was evaluated to refine a personalized follow-up plan. The goal is to achieve early identification of high-risk patients and enhance clinical trust in model predictions, ultimately improving the management and outcomes of NSCLC patients. A retrospective cohort of 255 treatment-naïve NSCLC patients was randomly partitioned (7:3 ratio) into training ( n = 178) and test sets ( n = 77). Following the extraction of 1888 baseline 18 F-FDG PET/CT radiomic features, least absolute shrinkage and selection operator (LASSO) regression was applied to derive the Rad score. Cox regression was used to analyze clinical and imaging variables to identify prognostic factors for overall survival (OS). Time-stratified prediction models were developed at 12-, 18-, and 24-month intervals using five machine learning (ML) algorithms: light gradient boosting machine (LightGBM), extreme gradient boosting (XGBoost), gradient boosting (GB), support vector machine (SVM), and K-nearest neighbors (KNN). Several evaluation indices, including the area under the receiver operating characteristic curve (AUC), were used to compare the predictive performance of the models. The SHapley Additive exPlanations (SHAP) framework enabled time point-specific global feature importance and local instance explanations to obtain critical determinants. The ICCs of the radiomics features were all greater than 0.75. After eliminating redundant features (with r > 0.9) through correlation screening, 485 features were retained for LASSO-Cox screening. Finally, four independent prognostic radiomic features were obtained, and a radiomic risk score (RS) was constructed. The RS was significantly correlated with OS in both the training set and the test set, and the KM analysis revealed that the high-RS group had a lower survival rate. In terms of time stratification, the AUC values of traditional TNM staging in the validation set at 12, 18, and 24 months were 0.595, 0.647, and 0.557, respectively, indicating generally poor predictive performance. Among the five machine learning models, the LightGBM model performed best at the three time points, with AUC values of 0.61, 0.68, and 0.69 at 12, 18, and 24 months, respectively, and accuracies of 0.71, 0.68, and 0.64, respectively. Therefore, the LightGBM model was selected as the optimal model for interpretation at the three time points. SHAP analysis at the three time points revealed that at 12 months, the clinical stage contributed the most (average |SHAP|= 0.29), followed by the maximum tumor diameter (0.15), RS (0.14), SUVmax (0.06), and age (0.01); at 18 months, the RS contributed the most (0.31), surpassing the clinical stage (0.24), and at 24 months, the RS dominance further increased (0.35), while the weight of the clinical stage decreased to 0.13. The SHAP waterfall plot revealed the individual-level feature contributions. The radiomics-based ML model enables multitimepoint survival prediction in patients with NSCLC, with the Rad score exhibiting dynamically enhanced prognostic power for long-term survival, surpassing conventional clinical staging over time, and providing an explainable imaging decision-support tool for personalized surveillance.