Abstract
BACKGROUND AND OBJECTIVES: Moderate hypofractionated radiotherapy (MHRT) is an important treatment modality for lung cancer, offering shorter courses and improved local control, yet it also markedly increases the risk of severe radiation-induced esophagitis (RIE; grade ≥3). Severe RIE compromises quality of life and adherence to therapy and may necessitate interruption of radiotherapy. This study aimed to develop a prediction model based on clinical and dosimetric factors to identify high-risk patients receiving MHRT and to facilitate individualized treatment strategies. METHODS: Lung cancer patients receiving moderate hypofractionated radiotherapy were included, with the endpoint defined as grade ≥3 radiation-induced esophagitis. Baseline characteristics were summarized using non-missing data only. During model development, in each outer bootstrap training set, candidate variables underwent single-rule imputation (median for continuous variables, mode for categorical variables) and standardization, followed by variable selection via elastic-net regression and model building with Firth-penalized logistic regression; the outcome itself was not imputed. Fully nested bootstrap validation (B=1000) was performed to assess internal robustness, with optimism-corrected performance metrics and 95% confidence intervals reported. Discrimination was evaluated using ROC curves and AUC, calibration by calibration plots and the Hosmer-Lemeshow test, and clinical utility through decision curve analysis (DCA). The optimal threshold was determined by the Youden index, with the corresponding confusion matrix presented. Finally, a nomogram was constructed to facilitate clinical visualization and application. RESULTS: A total of 105 patients were included; the incidence of grade ≥3 RIE was 16.2% (17/105). Five predictors entered the final model via elastic-net selection: mean gross tumor volume (mean GTV), V5, D2cc, circumferential 2.6-Gy irradiated length, and circumferential 3.0-Gy irradiated length. The Firth-penalized logistic model showed good apparent performance: AUC=0.771, Brier score = 0.114, calibration slope = 1.16, and calibration intercept = 0.13. After optimism correction by fully nested bootstrap (B=1,000), discrimination decreased to AUC=0.608 (95% CI, 0.464-0.761) with a corresponding Brier score of 0.176 (95% CI, 0.114-0.247). The Hosmer-Lemeshow test yielded χ² = 7.84, p = 0.449, indicating acceptable overall fit. The Youden-index-derived optimal cutoff was 0.130, stratifying patients into high-risk (predicted probability ≥ 0.13) and lower-risk (< 0.13) groups. DCA demonstrated positive net benefit over "treat-all" and "treat-none" strategies across threshold probabilities of 0-0.8. Optimism-corrected calibration parameters were unstable, likely reflecting the limited number of events; these results should be interpreted with caution. CONCLUSION: Using elastic-net feature selection and Firth logistic regression, we developed a model to predict severe (grade ≥3) RIE in lung cancer patients undergoing MHRT. The model exhibited moderate discriminatory ability with generally acceptable calibration, enables risk stratification and identification of high-risk patients, and is presented as a nomogram to support clinical application. It holds promise for guiding individualized radiotherapy decisions and the prevention of treatment-related complications.