3417 - Time-Dependent Toxicity Modeling and Local Control in BID Thoracic Reirradiation: Implications for Personalized Radiotherapy
Presenter(s)
C. F. P. M. de Sousa1, V. L. Doss1, E. Hales1, T. Trent1, E. Hagan2, T. Gebre2, A. Obaideen2, M. Negassa2, D. Liu1, C. Hu3, A. N. Viswanathan1, H. Li1, K. R. Voong1, X. Jia1, R. K. Hales1, T. R. McNutt1, and R. Ger1; 1Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 2Johns Hopkins University Bloomberg School of Public Health, Baltimore, MD, 3Division of Biostatistics and Bioinformatics, Sidney Kimmel Comprehensive Cancer Center, Johns Hopkins University School of Medicine, Baltimore, MD
Purpose/Objective(s): In re-irradiation (re-RT), the need for local control is counterbalanced by the high risk of adverse effects from cumulative doses. BID schedules may offset challenges in high-risk patients with dose overlap, but there is limited guidance from outcome and toxicity models. We investigated factors associated with local control and assessed the feasibility of a time-dependent recovery model for improved toxicity predictions in re-RT.
Materials/Methods: 65 thoracic BID patients were included (treated 2012-2024). Doses were deformed to the latest CT scan and converted to EQD2 (a/ß = 3) by a custom script. The cumulative incidence of local failure (LF) was estimated and compared across histology and dose via Gray’s test. For predicting =G2 esophagitis, we evaluated conventional direct summation - courses summed together without time consideration - and 3 time-dependent recovery models to discount previous doses: mono-exponential (ME), bi-exponential (BE), and reciprocal time (RC), each optimized through grid search. Logistic regression was used in the primary models, with bootstrapping to enhance robustness. Lasso and Elastic Net with k-fold cross-validation were also evaluated. AUC and ROC curves were calculated and compared via the Venkatraman method. Additional covariates (age, concurrent chemotherapy, smoking) were assessed via univariate and multivariate regression and log-likelihood ratio tests. P< 0.05 was considered significant.
Results: The 24-month cumulative incidence of LF was 42.9%, varying by histology (small cell: 22.2%, adenocarcinoma: 48.4%, squamous cell: 73.6%, other: 39.4%, p=0.05), and non-significantly by dose (= 45 Gy: 47.6%, > 45 Gy: 15.1%, p = 0.17). The median time from initial RT to BID re-RT was 22.2 months (range 7.5–162.7), reflecting treatment interval variability.
When modeling =G2 esophagitis, the direct summation model achieved an AUC of 0.74. Incorporating time-dependent recovery improved discrimination compared to direct summation, which was significant for 2 models (AUC ME: 0.80, p=0.29; BE: 0.81, p = 0.04; RC 0.80, p = 0.04). There was also a greater separation between summed dose distributions for the time-dependent recovery models, eliminating interquartile range overlap between patients with and without =G2 esophagitis, supporting discrimination improvements. Adding covariates did not improve performance (p=0.267).
Conclusion: BID is a promising re-RT strategy, though further investigation into dose escalation for specific histologies could optimize control. Our findings show that models incorporating time-dependent recovery are feasible and support their potential use over direct summation for more accurate toxicity prediction. They will pave the way for developing advanced outcome models for evidence-based re-RT, ultimately reducing toxicity and optimizing tumor dose for personalized and effective re-RT.