Predicting radiation esophagitis after lung cancer radiotherapy using dose-volume, radiomic and dosiomic features: an external validation study

Abstract

Acute radiation esophagitis can interrupt thoracic radiotherapy, but it remains unclear whether image- and dose-derived features improve prediction beyond dose-volume histogram (DVH) metrics and retain their performance across institutions. We externally validated seven models based on different combinations of clinical, DVH, radiomic and dosiomic predictors for Grade ≥ 2 acute radiation esophagitis using 226 Fuzhou patients for development and 76 Xiamen patients for external validation. The original esophageal ROI and an ROI expanded isotropically by 3 mm were analysed separately. Elastic-net logistic regression with embedded feature selection was used, and development performance was estimated using nested cross-validation with five outer and three inner folds. In external validation, the clinical model achieved an AUC of 0.650. Clinical plus DVH achieved AUCs of 0.802 and 0.797 for the original and expanded ROIs. The highest observed AUCs were 0.866 and 0.826 for clinical plus DVH and CT radiomics, with AUPRCs of 0.602 and 0.545; however, no pairwise AUC comparison remained significant after Holm correction. EQD2-based sensitivity analysis did not show a consistent advantage over physical dose. DVH variables provided relatively consistent discrimination across the two ROI definitions and remained a stable predictive foundation, while the highest observed external AUCs suggest potential complementary value from CT radiomics. The limited external event count, variable calibration and non-significant Holm-adjusted comparisons require prospective multicentre confirmation before clinical application.