1130 - Development of a Predictive Model for Rectal Toxicity in Prostate Cancer

K. Shah¹, H. Lee², Z. Soetan³, B. Y. Yeap⁴, M. Moteabbed¹, S. Muise³, J. Cowan³, K. Remillard⁵, B. L. Silvia³, S. C. Kamran³, D. T. Miyamoto⁵, H. Paganetti¹, J. A. Efstathiou⁶, and I. Chamseddine⁵; ¹Massachusetts General Hospital, Harvard Medical School, Boston, MA, ²St Jude Childrens Research Hospital, Memphis, TN, ³Massachusetts General Hospital, Boston, MA, ⁴Department of Medicine, Massachusetts General Hospital, Harvard Medical School, Boston, MA, ⁵Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, ⁶Department of Radiation Oncology, Harvard School of Medicine, Boston, MA

Purpose/Objective(s): This study aims to develop a predictive model for Common Terminology Criteria for Adverse Events (CTCAE) grade 1+ rectal toxicity within 3 months following prostate cancer treatment with photon and proton therapy. The clinical objective is to identify high-risk patients and aid personalized treatment modality selection by integrating dosimetric data with patient-specific features. Materials/

Methods: A multi-institution cohort of 368 prostate cancer patients treated with radiotherapy between 2012 and 2023 across 10 institutions was analyzed. Dosimetric and non-dosimetric variables were collected, and purposeful feature selection was employed to identify clinically meaningful predictors. Following exclusions, the patient cohort consists of 278 patients, which were split into discovery (n = 227) and independent validation (n = 51) datasets. Since the rectum is a hollow organ, the dose plan along the wall was unfolded into a 2D surface and the Dose-area histogram (DAH) was quantified. A convolutional neural network (CNN) model comprising 2 one-dimensional convolution layers was developed to extract dosimetric features from differential DAHs. We benchmarked the model against logistic regression (LR). Model validation on independent data was performed using the area under the receiver operating characteristics curve (AUC).

Results: Purposeful feature selection indicated the importance of rectum length, race, and the presence of hydrogel spacer for predicting rectal toxicity. The CNN model demonstrated predictive power and stability in the discovery datasets, with an AUC of 0.75, with a standard deviation of ± 0.11. The CNN model exceeded the LR model in independent validation, with an AUC of 0.75 versus 0.54. The CNN model was able to capture interactions between different DAH bins which led to better performance compared to a dose-volume based LR model. Testing on photon and proton cohorts separately yielded consistent AUCs of 0.75 and 0.76, indicating robustness across treatment modalities. In the independent validation dataset, the model showed excellent specificity in both modalities in the bottom 25% low-risk groups of 0.83 and 1 and high sensitivity in the photon top 25% high-risk group of 0.83, respectively, indicating the potential to be used for individualized treatment selection.

Conclusion: Our study presents a novel approach to predict rectal toxicity in prostate cancer patients undergoing radiotherapy combining rectal DAHs and patient-specific factors. The model can integrate both proton and photon therapy cohorts. The CNN model offered improved predictive accuracy compared to logistic regression, highlighting the importance of integrating dosimetric and patient-specific features for personalized treatment planning. The models promising performance warrants prospective validation using randomized clinical trials such as PARTIQoL (NCT01617161).