Prediction of radiation pneumonitis after CRT in patients with advanced NSCLC using multi‐region radiomics and attention‐based ensemble learning

Abstract Background Radiation pneumonitis (RP) is a major dose‐limiting toxicity in concurrent chemoradiotherapy (CRT) for stage III non‐small cell lung cancer (NSCLC). Existing models often analyze a single lung region and rely on a single algorithm, limiting accuracy and external validity. Purpose To develop and externally validate an attention‐weighted ensemble model that integrates multi‐region radiomics for individualized prediction of grade ≥2 RP after three‐dimensional conformal radiotherapy (3D‐CRT) or volumetric‐modulated arc therapy (VMAT). Methods We retrospectively analyzed 137 patients with stage III NSCLC from two Japanese centers (training, n = 107 and external validation, n = 30). 40 anatomical and dose‐stratified regions (covering the gross tumor volume [GTV], peritumoral shells, normal lung sub volumes, and dose sub volumes receiving 5–60 Gy) were delineated on the planning CT and dose maps. From each region, 837 radiomic features were extracted from original and wavelet‐filtered images. Region‐wise feature reduction (variance inflation filtering and least absolute shrinkage and selection operator, LASSO) yielded four radiomic scores (Radscore Tumor, _Lung, Dose, Shell). Five base learners (random forest (RF), gradient boosting machine (GBM), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost)) were trained on the four Radscores. Their outputs were combined using an attention‐weighted stacking meta‐learner (SurvBETA: Survival Boosted Ensemble with Tuned Attention) and integrated with clinical covariates into a nomogram. Discrimination, calibration, and risk‐group separation were evaluated using the concordance index (C‐index), calibration plots, and log‐rank tests. Results The SurvBETA + clinical nomogram achieved a C‐index of 0.87 in the training cohort and 0.83 in the external validation cohort, outperforming a clinical‐only model (0.54) and a conventional average‐stacking ensemble (0.65). High‐risk vs. low‐risk groups defined by the Kaplan–Meier curve showed clear separation in cumulative RP incidence (external cohort log‐rank p < 0.01), with visually acceptable calibration. Decision‐curve analysis indicated higher net benefit across clinically relevant thresholds compared with comparators. Conclusions An attention‐weighted ensemble of multi‐region radiomics features, combined with clinical factors, provided accurate and externally validated prediction of symptomatic RP after CRT for stage III NSCLC.