KEY POINTS
- This retrospective dual-center study included 246 patients receiving lung radiotherapy between 2023 and 2025: 171 patients formed the development cohort and 75 an independent center-2 test cohort. Grade ≥2 symptomatic radiation pneumonitis within three months occurred in 27/171 (15.8%) and 16/75 (21.3%), respectively.
- Three spatial strategies were compared: the whole lung excluding GTV; physical-dose subregions of 0–5, 5–10, 10–20, 20–30 and ≥30 Gy; and biologically effective dose subregions of 0–15, 15–30, 30–50, 50–70 and ≥70 Gy₃. Models incorporated clinical variables, DVH parameters, radiomics and dosiomics in multiple combinations.
- On independent testing, the complete radiomics+dosiomics+DVH+clinical models performed almost identically: external AUC was 0.676 for whole lung, 0.669 for physical-dose regions and 0.684 for BED3 regions. The maximum difference between strategies was therefore only 0.015, with overlapping uncertainty.
- Clinical variables alone reached an external AUC of 0.652, compared with 0.534 for DVH alone, 0.637–0.642 for the best radiomics-only approaches and 0.564–0.590 for dosiomics alone. Adding every feature family therefore produced only modest discrimination beyond the clinical baseline.
- Exploratory interval-specific analyses produced higher point estimates in selected regions, including AUC 0.745 for the 10–20 Gy physical-dose region and 0.720 for BED3 ≥70 Gy₃, but these were selected from multiple candidate regions and require independent confirmation.
- Calibration was a major weakness. Center-2 Brier scores were 0.184, 0.260 and 0.255 for the whole-lung, physical-dose and BED3 complete models, respectively, compared with 0.168 for a simple constant-probability reference based on the external event rate.
- The study used repeated nested cross-validation and bootstrap optimism correction, but remained limited by modest event numbers, substantial between-center treatment differences and absent patient-level harmonization of imaging acquisition/reconstruction. The authors therefore do not support immediate clinical risk prediction from these models.
CLINICAL TAKEAWAY
Dividing the lung according to physical or biologically effective dose did not clearly outperform whole-lung modeling for early symptomatic pneumonitis. The work is methodologically useful for future model development, but modest external discrimination and poor calibration mean these approaches are not ready to guide treatment decisions.