External beam radiation therapy treatment factors prognostic of biochemical failure free survival: A multi-institutional retrospective study for prostate cancer

•Treatment plans with CTV D99 ≥ 100% and PTV V98 ≥ 91% show optimal biochemical failure-free survival.•Treatment planning factors such as dose calculation algorithm are not prognostic.•Treatment delivery factors such as image-guidance type are not prognostic.•Random Survival Forestis an effective tool in identifying prognostic features. The goal of this work is to identify specific treatment planning and delivery features that are prognostic of biochemical failure-free survival (BFFS) for prostate cancer patients treated with external beam radiotherapy (EBRT). This study reviewed patients diagnosed with localized prostate adenocarcinoma between 2005 and 2016, and treated with EBRT on a Varian linear accelerator at one of the four cancer centers in Alberta, Canada. BFFS was calculated using the Kaplan-Meier estimator. Patient demographics, tumor characteristics, and EBRT treatment planning and delivery factors, were collected for each patient. The patient cohort was split into a training dataset with patients from two centers and a validation dataset with patients from another two centers. A random survival forest was used to identify features associated with BFFS. This study included 2827 patients with a median follow-up of 6.4 years. The BFFS for this cohort collectively was 84.9% at 5 years and 69.3% at 10 years. 2519 patients from two centers were used for model training and 308 patients from two different centers were used for model validation. The prognostic features were Gleason score, prostate-specific antigen (PSA) at diagnosis, clinical T stage, CTV D99, pelvic irradiation, IGRT frequency, and PTV V98. Variables including bladder volume, dose calculation algorithm, PTV D99, age at diagnosis, hip prosthesis, number of malignancies, fiducial marker usage, PTV volume, RT modality, PTV HI, rectal volume, hormone treatment, PTV D1cc, equivalent PTV margin, IGRT type, and EQD2_1.5 were unlikely to be prognostic. A random survival forest using only the seven prognostic variables was built. The Harrell’s concordance index for the model was 0.65 for the whole training dataset, 0.62 for out-of-bag samples of the training dataset, and 0.62 for the validation dataset. EBRT features prognostic of BFFS were identified and a random survival forest was developed for predicting prostate cancer patients’ BFFS after EBRT.