External Validation of Dose Patterns and Dosimetric Predictors for Radiation-Induced Esophagitis in Non-Small Cell Lung Cancer

Radiation-induced Esophagitis (RE) is a dose-limiting complication associated with chemo-radiation therapy (RT) for Non-Small-Cell Lung Cancer (NSCLC). We aimed to externally validate dose patterns and dosimetric predictors of grade ≥ 2 RE (RE2+) in an independent NSCLC cohort. We analyzed RTOG-0617 trial patient data. Patients received concurrent chemotherapy with or without cetuximab, and 60- versus 74-Gy radiation doses. Voxel-Based Analysis (VBA) assessed spatial dose differences between patients with and without RE2+. The Generalized Linear Model (GLM) included dose maps and each non-dosimetric variables selected by a LASSO regularized GLM of RE2+. A non-parametric permutation test of the maximum threshold-free cluster-enhanced statistic accounting for multiple comparisons was performed, the significance p- maps generated and compared using Dice-over-volume metric (DIV) with previous VBA results. External validity of esophageal predictors was also assessed in terms of discrimination (ROC-AUC) and calibration. Among 495 patients, 326 patients met inclusion criteria for analysis. The RE2+ incidence (44%) was consistent with the development cohort. Clinical factors were not significantly correlated with RE2+. VBA identified the most significant regions of associations between BED and RE2+ in the Upper and Middle thoracic Esophageal (UME) segment, with satisfactory overlap with a prior study (DIV = 0.7). The relative esophageal volume receiving ≥ 55 Gy (V55Gy) and the UME mean dose (Dmean) were both confirmed as good predictors in external validation with comparable discrimination (AUC = 0.7, 95% CI: [0.61, 0.72]). Calibration curves indicated superior risk prediction accuracy for UME-Dmean (Radj2 = 0.80 vs. 0.65). We successfully validated VBA findings on RE in an independent NSCLC cohort, reinforcing the role of the upper-middle esophagus as a critical region for RE2+ risk. The UME-Dmean emerged as a robust predictor across diverse treatment techniques and patient characteristics. These findings provide a foundation for clinical implementation of VBA.