Presenter(s)
A. Koron1, S. Rickards2, J. Pham3, M. Casado3, M. Cao4, J. M. Lamb5, L. Valle6, M. L. Steinberg6, A. U. Kishan6, and X. Qi3; 1Radiation Oncology, Los Angeles, CA, 2UCLA, Los Angeles, CA, 3Department of Radiation Oncology, University of California, Los Angeles, Los Angeles, CA, 4Department of Radiation Oncology, University of California San Francisco, San Francisco, CA, 5UCLA, Department of Radiation Oncology, Los Angeles, CA, 6Department of Radiation Oncology, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, CA
Purpose/Objective(s):
To develop and validate predictive models for late genitourinary (GU) toxicity following prostate SBRT, and to assess whether integrating factional delivered dosimetry and early toxicity scores, in addition to planning (SIM) dosimetry, improves predictive performance. Materials/Methods:
Sixty-four prostate SBRT patients enrolled in the NCT04384770 clinical trial were included. All patients received 40 Gy/5 Fx to the PTV (prostate and proximal seminal vesicles with 2 mm margins) on a 0.35 T MR-Linac. GU toxicity was graded using CTCAE v4.03 from treatment initiation through 3-month (early toxicity) and 12-month (late toxicity) follow-up. Potential predictors included the entire DVH curve binned at 0.1 Gy increments, dose statistics (min/mean/max), and structure volumes across multiple time points. Analyzed structures comprised targets (CTV/PTV) and organs at risk (bladder, bladder wall, trigone, urethra, rectum, and rectal wall). Early toxicity scores were additionally incorporated as predictors for late toxicity modeling. Feature dimensionality was reduced using Welch’s t-test and maximum relevance minimum redundancy prefiltering, followed byfraction-wise sparse-group logistic regression (SGLR) to identify the most relevant groups and individual predictors for binary outcome classification. Model performance was evaluated across ten matched random train/test splits (50/14 per split) using Area Under the Curve (AUC). To assess contributions from fractional dose features, LASSO logistic regression models were trained on SIM-only and SIM plus fractional data to avoid confounding by SGLR group-structure effects. Paired split-matched tests with Holm correction assessed performance differences and the impact of early toxicity information.
Results:
In this cohort of patients treated on a MR-Linac, 23.4% (15/64) and 20.3% (13/64) of patients developed grade 2 GU early and late toxicity, respectively. Using SIM-only data, SGLR model achieved an AUC of 0.57 ± 0.13 for predicting late GU toxicity. Incorporation of fractional dosimetry increased the AUC to 0.71 ± 0.11, and further addition of early toxicity scores yielded the highest performance (0.79 ± 0.12). Compared with SIM-only LASSO models without early toxicity features, the combined model incorporating fractional data and SIM data improved by ?AUC = +0.152 (p = 0.029). Further inclusion of early-toxicity features resulted in ?AUC = +0.218 (p = 0.006), with both differences remaining significant after Holm correction.
Conclusion:
Integration of fractional dosimetry enhanced predictive performance for late GU toxicity. Inclusion of early toxicity scores provided a significant additional improvement in model performance. These findings underscore the significance of both high-quality treatment planning and optimal inter-fractional dose delivery, supporting more personalized and adaptive radiotherapy strategies in prostate SBRT.
