Predicting radiation pneumonitis with dose-segmented radiomics in locally advanced non- small cell lung cancer patients undergoing consolidative immunotherapy post- concurrent chemoradiotherapy.

Objective: To develop and validate a machine learning model that integrates dose distribution-based radiomics, clinicopathological parameters, and hematological inflammatory biomarkers for predicting radiation pneumonitis (RP) in locally advanced non-small cell lung cancer (LA-NSCLC) patients receiving immuno-consolidation therapy after concurrent chemoradiation (CCRT). Methods: This retrospective study analyzed 161 locally advanced non-small cell lung cancer (LA-NSCLC) patients divided into training (n=112) and validation (n=49) cohorts. Radiomics features were extracted from planning CT scans across nine 5-Gy dose gradients (0-60 Gy), including the initial positioning CT (before radiotherapy) and a resetting CT (after a cumulative dose of 40-50 Gy), all within regions of interest (ROIs). Longitudinal feature changes were analyzed, followed by LASSO-based feature selection and logistic regression modeling. Machine learning methods evaluated associations between radiomics signatures (RS), clinical features, hematological inflammatory markers, and RP. Model performance was evaluated with AUC metrics and decision curve analysis (DCA). Results: Radiomics signatures across dose ranges (RS1:5 Gy; RS3:10-15 Gy; RS4:15-20 Gy; RS5:20-30 Gy; RS7:40-50 Gy; RS8:50-55 Gy; RS9:55-60 Gy) were developed. RS8 demonstrated the highest validation AUC (0.854). The model based on RS8 combined with tumor location achieved an AUC of 0.918 in the training cohort for predicting RP, whereas the addition of the neutrophil-to-lymphocyte ratio at 4 week (NLR 4w) to this model resulted in a marginally higher AUC of 0.938. Conclusions: The combined model improves RP prediction in LA-NSCLC patients undergoing post-CCRT consolidative immunotherapy, offering a novel approach for personalized patient management.