Comparison of whole-lung, physical-dose, and BED3-guided subregional multi-omics for predicting 3-month symptomatic radiation pneumonitis after lung cancer radiotherapy: a dual-center study

Region definition may influence radiation pneumonitis (RP) prediction, particularly when treatment schedules differ. We compared whole-lung, physical-dose subregional, and voxelwise biologically effective dose with α/β = 3 Gy (BED3)-guided strategies. The target outcome was grade ≥ 2 symptomatic RP within 3 months after lung cancer radiotherapy. This retrospective dual-center study included 246 patients between March 2023 and December 2025: 171 in a hospital-defined development cohort and 75 in the center-2 assessment cohort. The outcome was any grade ≥ 2 symptomatic RP recorded within 3 months after radiotherapy. Seven clinical, DVH, radiomic, dosiomic, and combined model specifications were evaluated within each region-design strategy. Development used repeated stratified 5 × 5 nested cross-validation and optimism correction based on 1000 bootstrap samples. All candidate models used the same outer and corresponding inner partitions and bootstrap resamples. Center 2 was excluded from model development and optimism correction and was used for center-specific out-of-sample performance assessment. Forty-three patients developed grade ≥ 2 symptomatic RP (27/171 in development; 16/75 in center 2). Under shared resampling, center-2 AUCs for the complete whole-lung, physical-dose, and BED3 models were 0.676, 0.669, and 0.684, respectively. The differences were small, with overlapping confidence intervals. Across five intervals per representation, the highest point estimates were observed for physical dose 10–20 Gy (AUC 0.745) and BED3 ≥ 70 Gy3 (AUC 0.720). These interval rankings are exploratory and subject to selection optimism. Center-2 Brier scores for the complete models ranged from 0.184 to 0.260, compared with 0.168 for a constant-probability reference based on the center-2 event rate; calibration was inadequate for clinical use. The paired analysis did not identify a clearly superior region design. Complete-model center-2 AUCs differed by at most 0.015, with overlapping confidence intervals. Physical dose 10–20 Gy and BED3 ≥ 70 Gy3 produced the highest interval-level point estimates, but these rankings are exploratory. Recalibration and prospective multicenter validation are required before clinical use. Not applicable.