Abstract Book

ESTRO 37

S373

using either a set of 17 clinical features (clinical models) or grade (grade models) or both (clinical + grade models). All models included demographics (age, sex, race). Clinical models included information on tumor location, brain or bone invasion, edema, and necrosis. [DRR1] To assess the generalizability of our models, each analysis was performed on 100 subsamples stratified by the outcome (LF or OS) of the full sample, each time obtaining accuracy from the omitted test set. The following algorithms were trained and compared: classification and regression tree (CART), logistic regression, regularized logistic regression (elastic net), support vector machine with linear (SVMLIN) and radial (SVMRAD) kernels, random forest (RF), and gradient boosting (GBM). Grade models were built using logistic regression. Results The best clinical models in each case (prediction of LF and OS) were at least as good as the grade models. Clinical + grade models offered minimal advantage over models based on clinical features alone (Figures 1 and 2[DRR1] ). Models based on grade predicted LF with accuracy of 0.64 on average (AUC = 0.66), compared to the best clinical model with accuracy of 0.69 (AUC = 0.74). In predicting OS, grade and the best clinical model both gave an accuracy of 0.66 (Grade AUC = 0.69, Clinical AUC = 0.71).

gioma as tumor grade. We believe these models may help to inform treatment planning decisions. All code used for this project will be made available online.

Poster: Clinical track: Haematology

PO-0730 Temporal changes in radiation-induced breast cancer in Hodgkin's lymphoma survivors. S. Appel 1 , Y.R. Lawrence 1 , Z. Symon 1 , E. Landau 1 , M. Ben- David 1 1 Chaim Sheba Medical Center, Radiation Oncology, Ramat Gan, Israel Purpose or Objective Radiation-induced breast cancer is a well described long- term side effect of treatment for mediastinal Hodgkin's lymphoma (HD), however there is ongoing debate regarding the shape of the dose-response curve. Over the last two decades radiation fields have shrunken: from large extended fields used forty years ago to the contemporary use of involved field or involved nodes; likewise radiation doses have dropped from 40Gy to 30Gy or even 20 Gy. We hypothesized that the decrease in radiation exposure over recent decades would be associated with a reduced risk of breast cancer (BC). Material and Methods Cases were identified from the SEER-9 NCI database. Inclusion criteria were women aged between 15-50 years at time of HD diagnosis, with the HD diagnosis made between January 1983 to December 2012. Women who had a non-HD first primary were excluded. Follow up was until end November 2014. Additional data extracted included disease stage, the use of external-beam radiation therapy, and the subsequent development of breast cancer. After categorization by radiation use, we compared the observed risk in each group to the expected risk in the general population (relative risk RR was the observed divided by the expected), and also compared the two groups to each other. We then calculated the risk of developing breast cancer across different time periods, based upon year of HD diagnosis: 1990-1995, 1995-1999, 2000-2004, 2005-2009 and 2010-2012. Statistical analysis was performed using Standardized Incidence Ratios tables generated from SEER*stat software v. 8.2.1. Confidence intervals were set at 95%. Results The entire cohort comprised 5,757 HD female cases. Of the 2,247 women who did not receive radiation, BC was subsequently diagnosed in 38, with RR 1.27 (95% CI 0.9- 1.75). External beam radiation treatment (EBRT) was given in 3,510 cases, and subsequent breast cancer diagnosed in 154 women, producing an overall relative risk (RR) compared to the general population of 2.66 (95%CI 2.2-3.1).Differences between groups were statistically significant (Chi square test p=0.009). Absolute excess risk of BC compared to the general population was 21.3 per 10,000 women for EBRT vs. 3.77 for no RT. In the EBRT cases, the RR of BC according to the period of radiation treatment was: during the years 1990-1994 RR 1.7 (95%CI 0.7-3.4); 1995-1999 RR 2.15 (95%CI 1.27- 3.4); 2000-2004 RR 2.5 (95%CI 1.7-3.5); 2005-2009 RR 3.5 (95%CI 2.6-4.5) and 2010-2012 RR 2.6 (95%CI 1.8-3.7). The difference in the RR between the time periods were not statistically significant (Chi square test p=0.7).

Conclusion In conclusion, we demonstrate that preoperative clinical features are at least as good predictors of local failure and overall survival in atypical and anaplastic menin-