M. Moteabbed1, B. Y. Yeap2, Z. Soetan3, K. Shah1, I. Chamseddine4, S. Muise3, J. Cowan3, K. Remillard4, B. L. Silvia3, S. C. Kamran3, D. T. Miyamoto4, A. L. Zietman4, H. Paganetti1, and J. A. Efstathiou5; 1Massachusetts General Hospital, Harvard Medical School, Boston, MA, 2Department of Medicine, Massachusetts General Hospital, Harvard Medical School, Boston, MA, 3Massachusetts General Hospital, Boston, MA, 4Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, 5Department of Radiation Oncology, Harvard School of Medicine, Boston, MA
Purpose/Objective(s):Urinary toxicities are very common among prostate cancer patients undergoing radiation therapy (RT). Acute urinary side effects (onset <3 months post-RT), especially moderate or severe (i.e. grade 2+), could have a significant impact on patient quality of life. We hypothesize and investigate the predictability of grade 2+ acute urinary toxicities through machine learning methods using pre-RT simulation CT-based radiomics features, combined with patient clinical characteristics and dosimetric information. Materials/
Methods: Data from 292 patients (204 cases for training (75%) and testing (25%), and 88 cases for independent validation) participating in the registry portion of a multi-institutional Phase III randomized clinical trial, collected between 2012 and 2023, was prospectively analyzed. Toxicity scores, baseline clinical characteristics, simulation CT images and respective contours, and bladder dose-volume metrics were retrieved from the patient database. Radiomics analysis was performed to extract 107 features (shape, texture and first order categories) for the bladder volume using an open source software. Ten clinical and 5 dose features were also investigated. Feature selection was performed to recognize and remove highly correlated features (Ccorr<0.6) to eliminate redundancy. The outcome classes were defined as cases with none or grade=1 versus grade=2 acute CTCAE-defined urinary toxicities. Supervised ensemble machine learning classification models, i.e. Random Forest, Gradient Boosting, Logistic Regression, KNeighbors and Support Vector Machine, were trained and tested using 5-fold cross validation and compared using receiver operating characteristics (ROC) analysis. Further evaluation of the model performance was carried out using the independent validation cohort. Results: A total of 17 uncorrelated features were selected for this analysis, which yielded a ROC area under the curve (AUC) of 0.75 ± 0.07 in the testing cross validation and 0.7 in the independent validation phase, using an optimized Gradient Boosting classifier. The sensitivity and specificity averaged across all probability thresholds were 0.7. Meanwhile, radiomics texture and shape features (Grey level variance and emphasis, skewness, elongation) were found amongst the top most important factors for predicting risk of higher grade urinary toxicities, along with treatment-specific characteristics (treatment modality and fractionation) as well as patient characteristics (age and Prostate Specific Antigen level). Conclusion: Supervised classification algorithms were found to offer good predictive value to distinguish patients at higher risk of moderate to severe RT-induced acute urinary side effects. Pre-RT imaging and patient specific characteristics could be valuable in determining the potential risk of complications, and hence guide individualized treatment and follow up strategies.
