Author: Victoria Doss, Tsion Gebre, Rachel B. Ger, Esi A Hagan, Elaina Hales, Russell K Hales, Xun Jia, Heng Li, Dezhi Liu, Todd R. McNutt, Meti Negassa, Anas Obaideen, Tinker Trent, K. Ranh Voong, Cecilia FPM de Sousa 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Johns Hopkins University 🌍
Purpose: As cancer care advances, more patients require re-irradiation, yet evidence-based data is lacking. This study aimed to develop a thoracic re-irradiation database and explore time-dependent recovery dynamics for toxicity modeling.
Methods: Thoracic patients treated at our institution since 2014 were screened and classified per ESTRO/EORTC re-irradiation guidelines. A REDCap database was created, capturing demographics, tumor characteristics, treatments, and outcomes, and integrated with healthcare systems for automated data retrieval. Patient dose was deformably registered to the latest CT scan, and a custom script converted all doses to EQD2. A subcohort of high-risk patients undergoing BID re-irradiation was selected for preliminary analysis evaluating ≥grade 2 esophagitis. First, logistic regression models using the total composite dose were created. We optimized three dose-recovery algorithms—mono-exponential, bi-exponential, and reciprocal-time—via grid search, enhancing discrimination by AUC. Lasso and Elastic Net regularization refined models. Bootstrap techniques were applied to logistic regression models, and k-fold cross-validation was used in Lasso regularization to ensure robustness. The DeLong test compared AUCs of base models with time-adjusted ones. Additional covariates (age, concurrent chemotherapy) were assessed using univariable and multivariable regression and log-likelihood ratio tests.
Results: Of 4,719 screened patients, 371 had type 1 (direct overlap), and 436 had type 2 (cumulative dose concern) re-irradiation. A comprehensive database integrated with clinical and planning systems was successfully established. In the preliminary analysis of 65 BID patients, time-dependent recovery algorithms significantly improved discrimination (AUC from 0.74 to 0.80, p=0.038 for mono-exponential; 0.81, p<0.001 for bi-exponential; 0.80, p=0.038 for reciprocal-time). Adding covariates did not enhance model performance (p=0.267), and their associations were not significant.
Conclusion: This study establishes a robust framework for thoracic re-irradiation data integration and demonstrates the value and feasibility of incorporating time-dependent recovery into toxicity models. These findings highlight the potential for broader applications in toxicity prediction in larger cohorts.