1248 - External Validation and Development of Radiomics Models for Radiation Pneumonitis after Chemoradio-Immunotherapy in Locally Advanced NSCLC

Purpose/Objective(s): Radiation pneumonitis (RP) remains a major treatment-limiting toxicity in locally advanced non–small cell lung cancer (LA-NSCLC) treated with combined chemo-radiotherapy (cCRT) and immunotherapy (IO). Although several pretreatment CT-based radiomics signatures have been reported for RP prediction after cCRT, their generalizability to contemporary cCRT+IO cohorts is unclear. This study externally validated published radiomics signatures and developed new signatures for RP and early-onset RP (RPEarly).

Materials/Methods: A literature review identified pretreatment CT-based radiomics signatures with reported AUC >0.70. Eligible signatures were reconstructed and externally validated in 195 LA-NSCLC patients treated with cCRT (60 Gy) followed by immunotherapy (2017–2022). Grade =2 RP and RPEarly (=3 months post-cCRT) occurred in 25% and 13% of patients, respectively. Radiomics features were extracted using the IBSI-compliant pyCERR framework, and original modeling pipelines were replicated. In parallel, an internal radiomics analysis was performed using features from original, wavelet-, and Gabor-filtered CT images of normal lungs and dose-stratified regions (NL, NL5Gy, NL10Gy, NL20Gy). Preprocessing and nested cross-validation with 20 repeated stratified 80/20 training/validation splits and 5-fold inner tuning were applied. Random forest, ElasticNet, and XGBoost models were evaluated.

Results: For external validation, four published radiomics signatures (2021–2024) were included. Validation in the independent cohort demonstrated substantial performance degradation relative to the original studies. The median (range) AUC decreased from 0.84 (0.72–0.89) to 0.56 (0.49–0.59) for RP and to 0.57 (0.45–0.66) for RPEarly. The largest decline was observed for models derived from normal lungs receiving >30 Gy (?AUC=0.40), whereas dose-map–based models showed greater robustness (?AUC=0.16 for RP, 0.06 for RPEarly). In the internally developed radiomics analysis, a total of 1,641 features were extracted. The median (range) test AUC was 0.54 (0.46–0.62) for RP and 0.51 (0.46–0.61) for RPEarly. Best performance was achieved by random forest models, with maximum AUCs of 0.62±0.08 (RP, NL5Gy) and 0.61±0.13 (RPEarly, NL10Gy). Among the 20 highest-ranked features, 11 were repeatedly selected across ROIs, including eight wavelet-derived and three original features.

Conclusion: Published radiomics RP risk signatures developed in cCRT-treated cohorts showed limited generalizability to patients receiving cCRT+IO. Newly developed models incorporating rigorous validation and dose-stratified regions demonstrated improved performance, supporting treatment-era–specific and dose-informed radiomics for RP risk stratification. Wavelet-derived features appear particularly promising.